{"id": "1d23ad55a293d9710bb07530b83eb18ff03de7fbb9e034f0fd160323bcaca32c", "sources": ["arxiv"], "title": "Uncertainty-Aware Reward Modeling for Stable RLHF", "abstract": "Reinforcement learning from human feedback (RLHF) aligns large language models by training reward models on preference data and optimizing policies to maximize predicted rewards. However, this pipeline faces two fundamental challenges: (1) reward models cannot signal when their predictions are unreliable, since they usually act as deterministic point estimators; and (2) modern group-based policy optimization can amplify unreliable reward signals, as exemplified by GRPO's uniform treatment of rewards during advantage computation. As policies explore increasingly diverse responses, these two limitations create a critical vulnerability: unreliable reward estimates may be granted disproportionate influence, triggering severe reward hacking. We propose Uncertainty-Aware Reward Modeling (UARM), which equips reward models with calibrated uncertainty via quantile-based conformal prediction and reweights GRPO advantages through heteroscedastic variance decomposition. Experiments across HelpSteer, UltraFeedback, and PKU-SafeRLHF demonstrate that UARM significantly improves reward model calibration, reduces reward hacking, and enhances downstream alignment quality compared to standard GRPO and uncertainty-agnostic baselines.", "authors": ["Licheng Pan", "Haocheng Yang", "Haoxuan Li", "Yichen Sun", "Yunsheng Lu", "Shijian Wang", "Lei Shen", "Yuan Lu", "Zhixuan Chu", "Hao Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": [], "published_date": "2026-06-18", "url": "https://arxiv.org/abs/2606.19818", "pdf_url": "https://arxiv.org/pdf/2606.19818v1", "arxiv_id": "2606.19818", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "536f9f096daad2029c71a41bf0a3d9f383d01be39eeb4abd0fd9f53b086c5a6f", "sources": ["arxiv"], "title": "Pareto Q-Learning with Reward Machines", "abstract": "We present Pareto Q-Learning with Reward Machines (PQLRM), a multi-objective reinforcement learning algorithm for tasks whose reward structure is specified by a set of reward machines (RMs). PQLRM combines Pareto Q-Learning (PQL), which maintains sets of vector-valued Q-estimates to approximate the Pareto front, with enhancements from Q-Learning with Reward Machines (QRM), which exploits the factored automaton structure of the reward signal. This yields a multi-policy algorithm that remains sample-efficient under non-Markovian, RM-encoded rewards. Experimental trials show that PQLRM converges faster than a naive PQL baseline applied to the cross-product MDP and can synthesize Pareto-optimal policies that QRM cannot.", "authors": ["Arnaud Lequen", "Clément Legrand-Lixon", "Léo Saulières"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": [], "published_date": "2026-06-17", "url": "https://arxiv.org/abs/2606.19134", "pdf_url": "https://arxiv.org/pdf/2606.19134v1", "arxiv_id": "2606.19134", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "15bf71c08d26be78e94a18979b27c82d8cebf65daa6462ac70d4d81641d370cf", "sources": ["arxiv"], "title": "Steerable Cultural Preference Optimization of Reward Models", "abstract": "It is essential for large language model (LLM) technology to serve many different cultural sub-communities in a manner that is acceptable to each community. However, research on LLM alignment has so far predominantly focused on predicting a unified response preference of annotators from certain regions. This paper aims to advance the development of alignment models with a more global outlook, that are able to accurately represent the preferences of subcommunities and do not exhibit excessive bias towards any of them. We focus on the development of reward models for this purpose and present a novel reward model training algorithm (SCPO) that can incorporate diverse cultural preferences in a balanced manner. Our method results in performance increases of the minority reward model of up to 7 points over the baseline model across two datasets, PRISM and GlobalOpinionQA, and across 7 countries. SCPO is up to 280% more training data-efficient than full-data finetuning of reward models. In addition, we perform analysis of bias by separately evaluating on the preference of subcommunities and show that excessive bias is mitigated via our weighting method. Our code is available at https://github.com/minsik-ai/Steerable-Cultural-Preference", "authors": ["Minsik Oh", "Advit Deepak", "Sophie Wu", "Douwe Kiela", "Ekaterina Shutova"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": [], "published_date": "2026-06-17", "url": "https://arxiv.org/abs/2606.18606", "pdf_url": "https://arxiv.org/pdf/2606.18606v1", "arxiv_id": "2606.18606", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/minsik-ai/Steerable-Cultural-Preference", "venue": null, "quality_score": 0.65}
{"id": "e53b309a3001159f948afa6b4c085d2dbad13379d42615da4f2af1c957f0bdca", "sources": ["arxiv", "semantic_scholar"], "title": "A Unifying Lens on Reward Uncertainty in RLHF", "abstract": "Reinforcement learning from human feedback (RLHF) is bottlenecked by reward hacking, where the policy exploits errors in a proxy reward model (RM) and produces high RM scores without genuine quality gains. A natural mitigation is pessimism: lowering rewards in regions where the RM is uncertain. However, standard scalar RMs provide no principled notion of uncertainty. We argue that the right object is a distributional reward model $p(r\\mid x,y)$. Under either a Bayesian inference or a KL-distributionally robust optimization (KL-DRO) lens, the KL-regularized RLHF objective admits a closed-form effective reward $\\tilde r(x,y) = \\pmβ\\log\\mathbb{E}_p[e^{\\pm r/β}]$. The pessimistic branch unifies the prior heuristics for RM ensemble aggregation: mean aggregation, worst-case optimization (WCO), and uncertainty-weighted optimization (UWO) all emerge as limits or truncations of this single expression. This also clarifies the implicit assumptions of each existing rule.", "authors": ["Ely Hahami", "Yoel Zimmermann", "Ray Zhou", "Jack Benarroch Jedlicki"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-08", "url": "https://arxiv.org/abs/2606.09073", "pdf_url": "https://arxiv.org/pdf/2606.09073v2", "arxiv_id": "2606.09073", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "2901871b1fafcfd85fc43399f35d27995799b32c96096606d7777709fe9cb5c5", "sources": ["arxiv", "semantic_scholar"], "title": "DynaCF: Mitigating Shortcut Learning in Reward Models via Dynamic Counterfactual Sensitivity", "abstract": "Reward models trained from pairwise preferences often exploit superficial shortcut cues rather than learning true response quality. We propose DynaCF, a dynamic reweighting framework for mitigating shortcut learning in reward model training. Unlike static shortcut heuristics, DynaCF measures shortcut sensitivity online during optimization by applying semantics-preserving counterfactual perturbations and tracking the resulting margin shifts and preference flips under the current model. Samples with higher shortcut sensitivity are dynamically downweighted in the Bradley-Terry objective, encouraging the model to rely less on superficial patterns and more on task-relevant preference signals. Extensive experiments show that DynaCF consistently improves robustness in preference modeling.", "authors": ["Fengyuan Liu", "Yongliang Miao", "Zirui He", "Yanguang Liu", "Fei Sun", "Mengnan Du"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-08", "url": "https://arxiv.org/abs/2606.09043", "pdf_url": "https://arxiv.org/pdf/2606.09043v1", "arxiv_id": "2606.09043", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4c5eda60152941e6f8379b4a6d05fdafc731fd72c6158f78c4bcf5f7820190b9", "sources": ["arxiv", "semantic_scholar"], "title": "When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming", "abstract": "Reinforcement learning from human feedback (RLHF) makes large-scale post-training possible by replacing an underspecified human objective with learned and scalable proxies. The same substitution creates a structured failure surface: optimization can raise the learned reward while external quality falls, degrade both proxy and judge scores, reveal proxy under-alignment, or produce evaluator-specific disagreement. We present an empirical failure-mode study of a compact RLHF pipeline with proximal policy optimization (PPO), direct preference optimization (DPO), uncertainty-penalized PPO (UP-PPO), reward-model uncertainty, approximate policy drift, diversity and repetition diagnostics, and two external LLM judges. Rather than treating reward hacking as a single terminal event, we classify matched transitions between checkpoints using the directions of the learned reward, judge scores, and average judge score. Across 61 checkpoint rows and 1920 row-level transitions, aggressive PPO has the highest localized reward-hacking rate (14.45%; bootstrap 95% CI: 10.16-18.75), while UP-PPO yields lower rates in the same aggressive regime (11.33-10.94%). A pre-transition logistic model predicts future row-level reward hacking with ROC-AUC 0.821, and row-level analysis finds localized reward hacking that checkpoint averages miss in 3 of 12 settings. The central conclusion is methodological: RLHF failures are not only final-model pathologies, but training dynamics that can be classified, localized, and partially anticipated.", "authors": ["Zelalem Abahana"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-02", "url": "https://arxiv.org/abs/2606.03238", "pdf_url": "https://arxiv.org/pdf/2606.03238v1", "arxiv_id": "2606.03238", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ffa7dd593613a170a8aefdf8639b244173ec134c9309d4131f852ba6e5d3818e", "sources": ["arxiv", "semantic_scholar"], "title": "Sparse Mixture-of-Experts Reward Models Learn Interpretable and Specialized Experts for Personalized Preference Modeling", "abstract": "Preference modeling plays a central role in reinforcement learning from human feedback (RLHF), enabling large language models (LLMs) to align with human values. However, most existing approaches assume a universal reward function, neglecting the diversity and heterogeneity of human preferences. To address this limitation without additional annotation costs, recent work has proposed learning multiple preference components from binary data and combining them to model individual preferences. Nevertheless, these components often fail to capture coherent and disentangled patterns, limiting their interpretability and effectiveness for personalization. In this work, we propose a sparse Mixture-of-Experts (MoE) reward model that encourages sparse routing and expert diversity during training on binary preference data. Across controlled and real-world experiments, sparse MoE learns interpretable routing patterns and specialized experts. It also improves test-time personalization, and post-adaptation shifts in expert weights provide a qualitative lens for analyzing how the model adapts to personalized preferences.", "authors": ["Yifan Wang", "Jinyi Mu", "Mayank Jobanputra", "Yu Wang", "Ji-Ung Lee", "Soyoung Oh", "Isabel Valera", "Vera Demberg"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-02", "url": "https://arxiv.org/abs/2606.04284", "pdf_url": "https://arxiv.org/pdf/2606.04284v1", "arxiv_id": "2606.04284", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a8c63d3c04413eb395860ed58cbfd0470d17de452ab7a6ab9b740cdaeab7c469", "sources": ["arxiv", "semantic_scholar"], "title": "EvalStop: Using World Feedback to Detect and Correct Reward Overoptimization in Multi-Tenant RLHF Platforms", "abstract": "Cloud LLM fine-tuning platforms increasingly serve RLHF workloads, where a learned reward model is optimized as a proxy for human quality. As Gao et al. (2023) showed, this proxy diverges from world feedback (downstream eval metrics) under sustained optimization pressure, a phenomenon known as reward overoptimization. Existing platform schedulers ignore this divergence: non-clairvoyant schedulers optimize JCT without any quality signal, SLAQ-style quality-aware schedulers use training loss (a weaker proxy that drops monotonically through hacking), and classical per-job early stopping requires human monitoring and does not free shared GPUs. We propose EvalStop, a composable scheduling primitive that terminates jobs on k consecutive eval-score declines, releases GPUs, preserves the best checkpoint, and delegates to any base scheduler. We frame scheduler-level early stopping as a detection problem and evaluate it in a discrete-event simulator whose RLHF workload mixes reward-hacking and structurally healthy runs, with ground-truth labels hidden from schedulers. On RLHF-heavy workloads (80% RLHF, 64 GPUs), EvalStop achieves precision 98% / recall 99% / FPR 1.5% while improving JCT by 9% and cutting wasted compute by 22% over SRTF-Est (p<0.05). Trivial fixed-progress and loss-plateau competitors either incur 65% FPR on healthy RLHF or miss over half of true hacking cases. Gains compose across every base scheduler tested (9-25% JCT) and detection quality stays stable under eval noise (precision at least 91% at noise std <= 0.05) and hacking base rate (precision at least 89% across 20-80% hacking fractions).", "authors": ["Guilin Zhang", "Chuanyi Sun", "Kai Zhao", "Xu Chu", "Shahryar Sarkani", "John M. Fossaceca"], "categories": ["cs.LG", "cs.AI", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-02", "url": "https://arxiv.org/abs/2606.04145", "pdf_url": "https://arxiv.org/pdf/2606.04145v3", "arxiv_id": "2606.04145", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "06b076868fb0a8d103d7e1af2f5e84f669e7dfc9f50ea242f44e20bda9b83d09", "sources": ["arxiv", "semantic_scholar"], "title": "RDA: Reward Design Agent for Reinforcement Learning", "abstract": "Reinforcement learning has enabled the acquisition of impressive robotic skills, but typically requires hand-crafted reward functions that are slow to design and difficult to align with human intentions. Recent work, such as Eureka, automates reward design by using an LLM to iteratively generate and refine reward code from task descriptions. However, they rely on coarse feedback signals such as success rate, which provide little semantic insight into the learned behavior. As a result, their trained policies achieve the final goal but are frequently poorly aligned with task instructions. We introduce the Reward Design Agent (RDA), a VLM-based agentic framework that injects semantic understanding into reward design. RDA decomposes tasks, visually evaluates trajectories, summarizes failure modes, and iteratively revises reward code to better align with task instructions. Across 12 tabletop manipulation tasks from ManiSkill and 4 whole-body manipulation tasks from HumanoidBench, RDA produces policies substantially more instruction-aligned than those of other baselines, while achieving comparable task success rates. Videos and the generated reward code are available on https://nitinkamra1992.github.io/reward-design-agent.", "authors": ["Hojoon Lee", "Ajay Subramanian", "Ben Abbatematteo", "Vijay Veerabadran", "Pedro Matias", "Karl Ridgeway", "Nitin Kamra"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-06-01", "url": "https://arxiv.org/abs/2606.01672", "pdf_url": "https://arxiv.org/pdf/2606.01672v1", "arxiv_id": "2606.01672", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "40337a1a94e845a0063edbf9bc197dc27f9656b6d1fe319424d7d203a863045f", "sources": ["arxiv", "semantic_scholar"], "title": "From Reward-Free Representations to Preferences: Rethinking Offline Preference-Based Reinforcement Learning", "abstract": "Preference-based reinforcement learning (PbRL) avoids explicit reward engineering by learning from pairwise human preference feedback. Existing offline PbRL methods typically follow a two-stage pipeline, first learning a reward or preference model from labeled preferences and then performing offline RL on unlabeled data. We revisit offline PbRL through the lens of reward-free representation learning (RFRL) from the zero-shot RL literature, and propose a new training framework that first learns latent successor-measure representations from reward-free offline data, followed by contrastive search and fine-tuning using preference data. Through extensive experiments and ablations, we show that our method achieves superior preference efficiency over offline PbRL baselines. This work is the first to connect RFRL with PbRL, highlighting its potential as a feedback-efficient solution. Our code is publicly available at https://github.com/rl-bandits-lab/FB-PbRL.", "authors": ["Jun-Jie Yang", "Chia-Heng Hsu", "Kui-Yuan Chen", "Ping-Chun Hsieh"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-31", "url": "https://arxiv.org/abs/2606.01123", "pdf_url": "https://arxiv.org/pdf/2606.01123v1", "arxiv_id": "2606.01123", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/rl-bandits-lab/FB-PbRL", "venue": null, "quality_score": 0.65}
{"id": "8564bdf28b12160b90415ce403089eae6552e5a6ea6426e36e0391bb816d2151", "sources": ["arxiv", "semantic_scholar"], "title": "The Representation-Rationalizability Tradeoff in Reward Learning", "abstract": "In RLHF, each training example contains a prompt $x$ and two candidate responses $y,y'$, and annotators provide pairwise preferences between these responses. The learning problem is to convert these heterogeneous pairwise judgments into a single scalar reward $r(x,y)$ that measures response quality for each prompt. Classical social choice implies an impossibility because heterogeneous annotator samples can induce pooled preferences with Condorcet cycles, so no scalar reward can evaluate all compared response pairs consistently. A growing literature analyzes RLHF as a social-choice problem, but usually assumes a fixed finite set of alternatives, i.e., a pre-enumerated finite set of candidate responses for each prompt. Modern pipelines instead score responses through a learned representation $φ(x,y)$ before a scalar head, so $φ$ determines which responses are treated as distinguishable alternatives and which comparisons are visible to the reward model. Once this embedding is part of the problem, the impossibility results from social choice theory become a tradeoff. We show that the excess cross-entropy loss of any reward built on $φ$ decomposes exactly into a representational term, which a richer $φ$ shrinks, and an aggregation term, which a richer $φ$ enlarges by exposing more comparisons that no scalar can rank consistently. The same results extend to direct preference optimization (DPO), and jointly training the embedding and the reward cannot guarantee to recover the sweet spot of this tradeoff. Experiments on synthetic data and real preference datasets corroborate our results.", "authors": ["Jing Dong", "Yaoliang Yu", "Pascal Pourpart"], "categories": ["cs.GT", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-29", "url": "https://arxiv.org/abs/2606.00291", "pdf_url": "https://arxiv.org/pdf/2606.00291v1", "arxiv_id": "2606.00291", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "844ca3b2606f222e9163a7702cb4222d89f2a52c181a6567a2b70ab276ed2e20", "sources": ["arxiv", "semantic_scholar"], "title": "In-Context Reward Adaptation for Robust Preference Modeling", "abstract": "Reinforcement Learning from Human Feedback (RLHF) typically relies on static reward models to align Large Language Models with human preferences. However, human values are inherently diverse and heterogeneous, and a single reward model often lacks the robustness required to generalize to unseen preference domains. While existing multi-reward frameworks attempt to address this, they are often restricted to a fixed set of known domains and fail to adapt to unseen human distributions without costly retraining. In this work, we propose In-Context Reward Adaptation, a transformer-based framework designed to model diverse and unseen human preferences on the fly. By leveraging the in-context learning capabilities of transformers, our approach adaptively infers the underlying reward structure from a small set of preference demonstrations. We demonstrate that while a standard transformer architecture is insufficient for this task by characterizing an asymptotic bias to the ground-truth, incorporating human response time as an auxiliary input signal enables the model to successfully adapt to preferences from previously unseen domains. Our findings show that this approach provides a more robust foundation for preference modeling, allowing for the representation of heterogeneous rewards and preference distribution shift, and offering a scalable path toward more flexible human-AI alignment.", "authors": ["Zhenyu Sun", "Zheng Xu", "Ermin Wei"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.30323", "pdf_url": "https://arxiv.org/pdf/2605.30323v1", "arxiv_id": "2605.30323", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "20fc28ab632906daf136f4c08b94f85d79a031ed75b149c41abdc4659973afe4", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Learning from Best-of-$N$ Preference Data: Targets, Tradeoffs, and Design Principles", "abstract": "Best-of-$N$ sampling is widely used to construct pairwise preference data: $N$ candidates are drawn from a base distribution, and the best is paired with a rejected response. Despite its widespread use, what Bradley--Terry (BT) reward learning extracts from such data, and how to choose $N$ and the base distribution, remain unclear. We specialize a recent analysis of preference data via its induced conditional distribution to Best-of-$N$. For independent-reference variants, we derive closed-form reward targets as explicit functions of $N$ and the base distribution, and show that they preserve the latent reward ranking. For the practical Best-vs-Random and Best-vs-Worst variants, chosen and rejected responses are coupled through the same candidate set, so exact BT representability generally fails; nevertheless, bounded-class minimizers approach the reference targets as $N$ grows. Although margin and connectivity are known to govern sample efficiency in pairwise preference learning, Best-of-$N$ couples them through $N$ in opposing directions: larger $N$ widens pairwise margins but reduces connectivity. This trade-off yields two design principles: use larger $N$ when preference labels are the bottleneck, smaller $N$ when generation is the bottleneck; and shape the base distribution to place mass between the responses whose comparison matters most at test time. Experiments on synthetic and real preference data support the predicted dependence on sample size and base-distribution shape.", "authors": ["Rattana Pukdee", "Maria-Florina Balcan", "Pradeep Ravikumar"], "categories": ["stat.ML", "cs.AI", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.30619", "pdf_url": "https://arxiv.org/pdf/2605.30619v1", "arxiv_id": "2605.30619", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "6e1f3104957ece7fbac9192701555b413b33bc2fb003f0935dfe2c6205ece5e7", "sources": ["arxiv", "semantic_scholar"], "title": "Rubric-Guided Process Reward for Stepwise Model Routing", "abstract": "Stepwise model routing improves the efficiency of Large Reasoning Models (LRMs) by assigning each reasoning step to a suitable model. Recent methods formulate routing as a sequential decision process and train the router with reinforcement learning. However, although they model routing as a process, they still supervise the router with outcome rewards. Such rewards only reflect final answer correctness and fail to evaluate intermediate routing decisions, which can weaken performance and generalization. To address this gap, we propose RoRo, a rubric-guided process reward framework for stepwise model routing. RoRo first collects diverse routing trajectories and constructs preference pairs based on outcome, cost, and process quality. It then trains a Rubricor to generate a query-specific evaluation rubric and a Judge to score routing trajectories under this rubric through alternating optimization. The resulting process rewards are combined with outcome rewards to optimize the routing policy via GRPO. Experiments on five reasoning benchmarks under both same-family and cross-family settings show that RoRo consistently outperforms strong baselines and achieves better accuracy and cost trade-offs.", "authors": ["Shenghao Ye", "Yu Guo", "Zhengheng Li", "Shuangwu Chen", "Jian Yang"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-28", "url": "https://arxiv.org/abs/2605.29310", "pdf_url": "https://arxiv.org/pdf/2605.29310v1", "arxiv_id": "2605.29310", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "e8ccfdb70c48392eaa25d06ca579bf1115efef646d6885136194f0af89b5add6", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Pairwise Preferences: Listwise Reward-Aware Alignment for Diffusion Models", "abstract": "Preference optimization has emerged as an efficient alternative to online reinforcement learning from human feedback (RLHF) for aligning text-to-image diffusion models. However, existing methods largely reduce supervision to binary pairwise comparisons. This pairwise reduction is limiting when training data naturally contains multiple candidate images for the same prompt, and when continuous reward scores can provide richer information than a single winner-loser label. To address these limitations, we propose Diffusion LAIR, a reward-aware listwise preference optimization method for diffusion models. For each prompt, LAIR converts reward scores across a group of candidate images into centered advantage weights, then optimizes an advantage-weighted regression objective on the implicit reward, defined as the denoising-loss improvement of the current model over a fixed reference model, with a quadratic penalty that regularizes the magnitude of the implicit reward. The resulting objective uses all candidates simultaneously rather than selecting pairs, and remains conservative by explicitly controlling the magnitude of the implicit reward. The LAIR objective admits a bounded closed-form optimum in implicit-reward space, clarifying how the regularization strength controls the magnitude of the preference update. Experiments show that Diffusion LAIR outperforms strong preference optimization baselines on SD1.5 and SDXL across text-to-image generation, compositional generation, and image editing benchmarks.", "authors": ["Austin Wang", "Jiaqi Han", "Stefano Ermon", "Yisong Yue"], "categories": ["cs.LG", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-26", "url": "https://arxiv.org/abs/2605.26491", "pdf_url": "https://arxiv.org/pdf/2605.26491v1", "arxiv_id": "2605.26491", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "7d765b8ec398948d46ce164bc9fde1a1ae580462a14d12d4fedbc96026cde05e", "sources": ["arxiv", "semantic_scholar"], "title": "Focal Reward: Balanced Reinforcement Learning under Rubric-Based Rewards", "abstract": "The open-ended generation in LLMs usually requires multi-dimensional rubrics to adequately assess quality and guide the improvement of reinforcement learning. However, a critical dilemma inherent in this training paradigm is the imbalanced reward polarization along different rubric dimensions. Under this bottleneck, even if LLMs achieve relatively high rewards after training, they may still exhibit severe deficiencies in certain dimensions, leading to a direct deterioration in user experience. To address this problem, we propose Focal Reward, a novel objective to automatically balance the training of reinforcement learning under rubric-based rewards. Specifically, we first leverage an inverse reward projection mechanism to estimate the saturation degree of each criterion in the rubric, which forms the basis to calibrate the reward direction. Then, the final objective is designed with an automatically reweighting coefficient for each criterion to achieve the fine-grained balancing. Extensive experiments across three model scales and six benchmarks demonstrate that our Focal Reward method outperforms the strongest static aggregation baseline in all 18 model-benchmark comparisons. Rollout, mechanism, and ablation analyses further show that these gains arise from online, saturation-aware reallocation toward rubrics that still have room for improvement.", "authors": ["Yu Huang", "Zihua Zhao", "Zhaoxin Huan", "Wanli Gu", "Feng Hong", "Xinmu Ge", "Lin Yuan", "Weichang Wu", "Qiang Hu", "Xiaolu Zhang", "Jun Zhou", "Jiangchao Yao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-26", "url": "https://arxiv.org/abs/2605.26579", "pdf_url": "https://arxiv.org/pdf/2605.26579v1", "arxiv_id": "2605.26579", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "5ddbc515851e6d77372c247310dfb30c3093b4284e663b00c667ba4e393eeb06", "sources": ["arxiv", "semantic_scholar"], "title": "How Neural Reward Models Learn Features for Policy Optimization: A Single-Index Analysis", "abstract": "Reward modeling is not only a prediction problem: in KL-regularized policy optimization, the learned reward is exponentiated to define the deployed policy, so downstream value depends on errors in reward-tilted regions. We study this feedback in a Gaussian single-index model with $r^*(x) = σ^*(\\langle θ^*, x\\rangle)$ and $x \\sim N(0, I_d)$. We analyze a two-stage neural reward model that first learns the hidden direction $θ^*$ from reward-weighted samples and then fits the readout layer by weighted ridge regression. Exponential reward weighting changes the Hermite signal available to the first layer; for any feature-learning temperature $β_1$ above a dimension-free $O(1)$ threshold, a constant fraction of neurons recover the hidden direction, with weak-recovery complexity governed by the generative exponent. After feature recovery, we derive tilted-policy value-gap bounds for an idealized label-weighted fit with weights $e^{y/β_2}$ and a more practical surrogate-weighted fit with weights $e^{r_{a_0}(x)/β_2}$. Keeping the $β_2$-dependence explicit yields an admissible set of deployment temperatures, balancing the gain from lowering $β_2$ against the learning cost amplified by exponential weighting; in the surrogate-weighted case, proxy-dependent factors shrink this admissible set.", "authors": ["Rei Higuchi", "Ryotaro Kawata", "Akifumi Wachi", "Shokichi Takakura", "Kohei Miyaguchi", "Taiji Suzuki"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2026-05-23", "url": "https://arxiv.org/abs/2605.24749", "pdf_url": "https://arxiv.org/pdf/2605.24749v1", "arxiv_id": "2605.24749", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "ab81af1d9787cef06f5c03c73c7bd8eb042707b2870328d761aa8d02e024b6e6", "sources": ["arxiv", "semantic_scholar"], "title": "PREFINE: Preference-Based Implicit Reward and Cost Fine-Tuning for Safety Alignment", "abstract": "We address the problem of making a pre-trained reinforcement learning (RL) policy safety-aware by incorporating cost constraints without retraining it from scratch. While costs could be numerically encoded, we assume a more general setting is when costs are provided as preferences. Given a reward-optimized policy and a small dataset of preferred (low-cost) and dispreferred (high-cost) trajectories, our goal is to fine-tune the policy to generate low-cost behaviors while retaining high rewards. Unlike standard RLHF in language models, where preferences are defined over responses to the same prompt, our setting involves trajectory-level preferences in continuous control environments. We introduce PREFINE: Preference-based Implicit Reward and Cost Fine-Tuning for Safety Alignment which is a preference-based fine-tuning method that adapts Direct Preference Optimization (DPO), which is now widely used for LLM fine-tuning, to the sequential decision making setting. PREFINE constructs policy-sampled counterfactual trajectories to establish meaningful preference contrasts and jointly optimizes for reward retention and safety alignment. Empirically, PREFINE reduces constraint violations and catastrophic failures by over 60% while maintaining original reward behavior. PREFINE produces policies that achieve low-cost, high-reward performance with significantly improved data and computational efficiency compared to full offline RL or imitation learning, bridging preference alignment and safe policy adaptation in continuous domains.", "authors": ["Richa Verma", "Bavish Kulur", "Sanjay Chawla", "Balaraman Ravindran"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-20", "url": "https://arxiv.org/abs/2605.21225", "pdf_url": "https://arxiv.org/pdf/2605.21225v1", "arxiv_id": "2605.21225", "doi": "10.65109/sdrb4374", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "c74e533a4ffd3bbd16f4460f50dab708a3da199f34d89d8c929269da45cf0869", "sources": ["arxiv", "semantic_scholar"], "title": "Process Rewards with Learned Reliability", "abstract": "Process Reward Models (PRMs) provide step-level feedback for reasoning, but current PRMs usually output only a single reward score for each step. Downstream methods must therefore treat imperfect step-level reward predictions as reliable decision signals, with no indication of when these predictions should be trusted. We propose BetaPRM, a distributional PRM that predicts both a step-level success probability and the reliability of that prediction. Given step-success supervision from Monte Carlo continuations, BetaPRM learns a Beta belief that explains the observed number of successful continuations through a Beta-Binomial likelihood, rather than regressing to the finite-sample success ratio as a point target. This learned reliability signal indicates when a step reward should be trusted, enabling downstream applications to distinguish reliable rewards from uncertain ones. As one application, we introduce Adaptive Computation Allocation (ACA) for PRM-guided Best-of-N reasoning. ACA uses the learned reliability signal to stop when a high-reward solution is reliable and to spend additional computation on uncertain candidate prefixes. Experiments across four backbones and four reasoning benchmarks show that BetaPRM improves PRM-guided Best-of-N selection while preserving standard step-level error detection. Built on this signal, ACA improves the accuracy--token tradeoff over fixed-budget Best-of-16, reducing token usage by up to 33.57% while improving final-answer accuracy.", "authors": ["Jinyuan Li", "Langlin Huang", "Chengsong Huang", "Shaoyang Xu", "Donghong Cai", "Yuyi Yang", "Wenxuan Zhang", "Jiaxin Huang"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-15", "url": "https://arxiv.org/abs/2605.15529", "pdf_url": "https://arxiv.org/pdf/2605.15529v1", "arxiv_id": "2605.15529", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "a724fc1423e57ea3892080898e88cd92924d6946340a2a62e1194d460e6a6b9a", "sources": ["arxiv", "semantic_scholar"], "title": "Variance-aware Reward Modeling with Anchor Guidance", "abstract": "Standard Bradley--Terry (BT) reward models are limited when human preferences are pluralistic. Although soft preference labels preserve disagreement information, BT can only express it by shrinking reward margins. Gaussian reward models provide an alternative by jointly predicting a reward mean and a reward variance, but suffer from a fundamental non-identifiability from pairwise preferences alone. We propose Anchor-guided Variance-aware Reward Modeling, a framework that resolves this non-identifiability by augmenting preference data with two coarse response-level anchor labels. Building on this, we prove that two anchors are sufficient for identification, develop a joint training objective and establish a non-asymptotic convergence rate for both the estimated reward mean and variance functions. Across simulation studies and four real-world diverging-preference datasets, our method consistently improves reward modeling performance and downstream RLHF, including PPO training and best-of-$N$ selection.", "authors": ["Shuxing Fang", "Ruijian Han", "Liangyu Zhang", "Fan Zhou"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2026-05-12", "url": "https://arxiv.org/abs/2605.11865", "pdf_url": "https://arxiv.org/pdf/2605.11865v1", "arxiv_id": "2605.11865", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "18e939e0a31d45dbe0a09aabd7f95f26cbc0f600567ba040b98259bb6f97c109", "sources": ["arxiv", "semantic_scholar"], "title": "Unsupervised Process Reward Models", "abstract": "Process Reward Models (PRMs) are a powerful mechanism for steering large language model reasoning by providing fine-grained, step-level supervision. However, this effectiveness comes at a significant cost: PRMs require expert annotations for every reasoning step, making them costly and difficult to scale. Here, we propose a method for training unsupervised PRMs (uPRM) that requires no human supervision, neither at the level of step-by-step annotations nor through ground-truth verification of final answers. The key idea behind our approach is to define a scoring function, derived from LLM next-token probabilities, that jointly assesses candidate positions of first erroneous steps across a batch of reasoning trajectories. We demonstrate the effectiveness of uPRM across diverse scenarios: (i) uPRM achieves up to 15% absolute accuracy improvements over the LLM-as-a-Judge in identifying first erroneous steps on the ProcessBench dataset; (ii) as a verifier for test-time scaling, uPRM performs comparably to supervised PRMs and outperforms the majority voting baseline by up to 6.9%, and (iii) when used as a reward signal in reinforcement learning, uPRM enables more robust policy optimization throughout training compared to a supervised PRM trained using ground-truth labels. Overall, our results open a path toward scalable reward modeling for complex reasoning tasks.", "authors": ["Artyom Gadetsky", "Maxim Kodryan", "Siba Smarak Panigrahi", "Hang Guo", "Maria Brbic"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-11", "url": "https://arxiv.org/abs/2605.10158", "pdf_url": "https://arxiv.org/pdf/2605.10158v1", "arxiv_id": "2605.10158", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "065489ca5e818afb74c9b2672e38d9cf7e575bb2e927e5945881d4a9f3916acb", "sources": ["arxiv", "semantic_scholar"], "title": "Spurious Correlation Learning in Preference Optimization: Mechanisms, Consequences, and Mitigation via Tie Training", "abstract": "Preference learning methods like Direct Preference Optimization (DPO) are known to induce reliance on spurious correlations, leading to sycophancy and length bias in today's language models and potentially severe goal misgeneralization in future systems. In this work, we provide a unified theoretical analysis of this phenomenon, characterizing the mechanisms of spurious learning, its consequences on deployment, and a provable mitigation strategy. Focusing on log-linear policies, we show that standard preference-learning objectives induce reliance on spurious features at the population level through two channels: mean spurious bias and causal-spurious correlation leakage. We then show that this reliance creates an irreducible vulnerability to distribution shift: more data from the same training distribution fails to reduce the model's dependence on spurious features. To address this, we propose tie training, a data augmentation strategy using ties (equal-utility preference pairs) to introduce data-driven regularization. We demonstrate that this approach selectively reduces spurious learning without degrading causal learning. Finally, we validate our theory on log-linear models and provide empirical evidence that both the spurious learning mechanisms and the benefits of tie training persist for neural networks and large language models.", "authors": ["Christian Moya", "Alex Semendinger", "Guang Lin", "Elliott Thornley"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-11", "url": "https://arxiv.org/abs/2605.11134", "pdf_url": "https://arxiv.org/pdf/2605.11134v2", "arxiv_id": "2605.11134", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Proceedings of the 43rd International Conference on Machine Learning, 2026, Seoul, South Korea", "quality_score": 0.55}
{"id": "b73983bba8a2b09656b1a0b2df466bdce3688205c4af0f8b4a1ac0f58b959a5e", "sources": ["arxiv", "semantic_scholar"], "title": "$ξ$-DPO: Direct Preference Optimization via Ratio Reward Margin", "abstract": "Reference-free preference optimization has emerged as an efficient alternative to reinforcement learning from human feedback, with Simple Preference Optimization(SimPO) demonstrating strong performance by eliminating the explicit reference model through a simple objective. However, the joint tuning of the hyperparameters $β$ and $γ$ in SimPO remains a central challenge. We argue that this difficulty arises because the margin formulation in SimPO is not easily interpretable across datasets with different reward gap structures. To better understand this issue, we conduct a comprehensive analysis of SimPO and find that $β$ implicitly controls sample filtering, while the effect of $γ$ depends on the reward gap structure of the dataset. Motivated by these observations, we propose $ξ$-DPO: Direct preference optimization via ratio reward margin. We first reformulate the preference objective through an equivalent transformation, changing the optimization target from maximizing the likelihood of reward gaps to minimizing the distance between reward gaps and optimal margins. Then, we redefine the reward in a ratio form between the chosen and rejected, which effectively cancels the effect of $β$ and yields a bounded and interpretable margin. This margin is called the ratio reward margin and is denoted by $ξ$. Unlike the margin $γ$ in SimPO, $ξ$ explicitly represents the desired relative separation between chosen and rejected responses and can be determined from the initial reward gap distribution, avoiding repeated trial-and-error tuning. ....", "authors": ["Zhengyuan Fan", "Zhonghua Wu", "Yuxuan Du", "Qun Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-09", "url": "https://arxiv.org/abs/2605.10981", "pdf_url": "https://arxiv.org/pdf/2605.10981v1", "arxiv_id": "2605.10981", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "4033de934a2c65fc0a94e7348ed3813ecfd866e349786c733922fe1989038e90", "sources": ["arxiv", "semantic_scholar"], "title": "Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders", "abstract": "Preference learning in large language models relies on reward models as proxies for human judgment. However, these models frequently exhibit preference instability, producing contradictory preference assignments in response to subtle, meaning-preserving input variations. We analyze this instability at the representation level under three semantic-preserving perturbation types: paraphrasing, pattern injection, and backdoor triggers. We attribute this instability to over-reliance on predictive yet brittle features, which we term unstable features, and isolate them via Sparse Autoencoders (SAEs) in a sparse latent space where benign and perturbed inputs activate distinctly separable patterns. Building on this separability, we propose two SAE-based instability mitigation strategies: SAE Feature Steering, which identifies and suppresses anomalously activated features at inference, and SAE Residual Correction, which learns adaptive adjustments over SAE features to restore correct preferences. Our methods substantially reduce incorrect preference assignments on harmlessness and hallucination benchmarks while preserving benign performance and general utility on other tasks, without retraining the reward model. Our code and data are available in \\url{https://github.com/shunchang-liu/pisa}.", "authors": ["Shunchang Liu", "Xin Chen", "Belen Martin Urcelay", "Francesco Croce"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.16339", "pdf_url": "https://arxiv.org/pdf/2605.16339v1", "arxiv_id": "2605.16339", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/shunchang-liu/pisa}", "venue": null, "quality_score": 0.65}
{"id": "aeb37bbf2614e2a8cda00646ba389d0c1c23824cd3b7bade15b0ab05fd22afd6", "sources": ["arxiv", "semantic_scholar"], "title": "Optimal Transport for LLM Reward Modeling from Noisy Preference", "abstract": "Reward models are fundamental to Reinforcement Learning from Human Feedback (RLHF), yet real-world datasets are inevitably corrupted by noisy preference. Conventional training objectives tend to overfit these errors, while existing denoising approaches often rely on homogeneous noise assumptions that fail to capture the complexity of linguistic preferences. To handle these challenges, we propose SelectiveRM, a framework grounded in optimal transport. We first devise a Joint Consistency Discrepancy to align the distribution of model predictions with preference data. Furthermore, to address the limitation of strict mass conservation which compels the model to fit outliers, we incorporate a Mass Relaxation mechanism via partial transport. This enables the autonomous exclusion of samples with noisy preference that contradict semantic consistency. Theoretically, we demonstrate that SelectiveRM optimizes a tighter upper bound on the unobserved clean risk. Extensive experiments validate that our approach significantly outperforms state-of-the-art baselines across diverse benchmarks.", "authors": ["Licheng Pan", "Haochen Yang", "Haoxuan Li", "Yunsheng Lu", "Yongqi Tong", "Yinuo Wang", "Shijian Wang", "Zhixuan Chu", "Lei Shen", "Yuan Lu", "Hao Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-07", "url": "https://arxiv.org/abs/2605.06036", "pdf_url": "https://arxiv.org/pdf/2605.06036v1", "arxiv_id": "2605.06036", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "dddc48c0563e8fc55c53d1e2008fbd2ade4997d8d23f6d404bde57e11b365b82", "sources": ["arxiv", "semantic_scholar"], "title": "Misaligned by Reward: Socially Undesirable Preferences in LLMs", "abstract": "Reward models are a key component of large language model alignment, serving as proxies for human preferences during training. However, existing evaluations focus primarily on broad instruction-following benchmarks, providing limited insight into whether these models capture socially desirable preferences. As a result, important failures in social alignment can remain hidden. We extend reward-model benchmarking to four socially consequential domains: bias, safety, morality, and ethical reasoning. We introduce a framework that converts social evaluation datasets into pairwise preference data, leveraging gold labels where available and directional bias indicators otherwise. This enables us to test whether reward models prefer socially undesirable responses, and whether their preferences produce systematically biased distributions over selected outputs. Across five publicly available reward models and two instruction-tuned models used as reward proxies, we find substantial variation across domains, with no single model performing best overall. The models fall well short of strong social intelligence: they often prefer socially undesirable options, and their preferences produce systematically biased distributions. Moreover, stronger bias avoidance can reduce sensitivity to context, revealing a key alignment trade-off between avoiding biased outcomes and preserving contextual faithfulness. These findings show that standard reward benchmarks are insufficient for assessing social alignment and highlight the need for evaluations that directly measure the social preferences encoded in reward models.", "authors": ["Gayane Ghazaryan", "Esra Dönmez"], "categories": ["cs.CL", "cs.AI", "cs.CY"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-06", "url": "https://arxiv.org/abs/2605.05003", "pdf_url": "https://arxiv.org/pdf/2605.05003v1", "arxiv_id": "2605.05003", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "dd9ac75204f67cadf4b7605710e0e54ed8127f921bd72c7fbdbc8dd96e5c7adc", "sources": ["arxiv", "semantic_scholar"], "title": "RMGAP: Benchmarking the Generalization of Reward Models across Diverse Preferences", "abstract": "Reinforcement Learning from Human Feedback has become the standard paradigm for language model alignment, where reward models directly determine alignment effectiveness. In this work, we focus on how to evaluate the generalizability of reward models. By \"generalizability\", we mean the ability of RMs to correctly rank responses to align with diverse user preferences. However, existing reward model benchmarks are typically designed around a universal preference, failing to assess this generalization. To address this critical gap, we introduce RMGAP, a benchmark comprising 1,097 instances across Chat, Writing, Reasoning, and Safety domains. Since different users exhibit diverse preferences for the same task, we first generate four distinct responses with different linguistic profiles for each collected prompt. However, the original prompt set lacks the specificity to convey different preferences. We therefore construct tailored prompts by contrasting these candidates and designing scenarios in which one response becomes the uniquely appropriate choice. Moreover, we observe that users often express the same preference using different phrasings, and thus extend each prompt with two paraphrased variants. Our evaluation of 24 state-of-the-art RMs reveals their substantial limitations: even the best RM achieves only 49.27% Best-of-N accuracy, highlighting considerable room for improvement in reward model generalization. Related data and code are available at https://github.com/nanzhi84/RMGAP.", "authors": ["Yangyang Zhou", "Yi-Chen Li"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-03", "url": "https://arxiv.org/abs/2605.01831", "pdf_url": "https://arxiv.org/pdf/2605.01831v1", "arxiv_id": "2605.01831", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/nanzhi84/RMGAP", "venue": null, "quality_score": 0.65}
{"id": "a175076b1e4da2c8f0bfdc8bf45d60fcdd7a9353faa5081acef1c7f2da9f0146", "sources": ["arxiv", "semantic_scholar"], "title": "PrefMoE: Robust Preference Modeling with Mixture-of-Experts Reward Learning", "abstract": "Preference-based reinforcement learning offers a scalable alternative to manual reward engineering by learning reward structures from comparative feedback. However, large-scale preference datasets, whether collected from crowdsourced annotators or generated by synthetic teachers, often contain heterogeneous and partially conflicting supervision, including disagreement across annotators and inconsistency within annotators. Existing reward learning methods typically fit a single reward model to such data, forcing it to average incompatible signals and thereby limiting robustness. To solve this, we propose PrefMoE, a mixture-of-experts reward learning framework for robust preference modeling. PrefMoE learns multiple specialized reward experts and uses trajectory-level soft routing to combine them adaptively, enabling the model to capture diverse latent preference patterns under noisy and heterogeneous preference supervision. A load-balancing regularizer further stabilizes training by preventing expert collapse. Across locomotion benchmarks from D4RL and manipulation tasks from MetaWorld, PrefMoE improves preference prediction robustness and leads to more reliable downstream policy learning than strong single-model baselines.", "authors": ["Ziqin Yuan", "Ruiqi Wang", "Dezhong Zhao", "Baijian Yang", "Byung-Cheol Min"], "categories": ["cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-05-01", "url": "https://arxiv.org/abs/2605.00384", "pdf_url": "https://arxiv.org/pdf/2605.00384v1", "arxiv_id": "2605.00384", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.35}
{"id": "bacd73bd7af44419d88f197e663d70e8b86fdf93f365435948ea2bb4e927c076", "sources": ["arxiv", "semantic_scholar"], "title": "Uncertainty-Aware Reward Discounting for Mitigating Reward Hacking", "abstract": "Reinforcement learning (RL) systems typically optimize scalar reward functions that assume precise and reliable evaluation of outcomes. However, real-world objectives--especially those derived from human preferences--are often uncertain, context-dependent, and internally inconsistent. This mismatch can lead to alignment failures such as reward hacking, over-optimization, and overconfident behavior. We introduce a dual-source uncertainty-aware reward framework that explicitly models both epistemic uncertainty in value estimation and uncertainty in human preferences. Model uncertainty is captured via ensemble disagreement over value predictions, while preference uncertainty is derived from variability in reward annotations. We combine these signals through a confidence-adjusted Reliability Filter that adaptively modulates action selection, encouraging a balance between exploitation and caution. Empirical results across multiple discrete grid configurations (6x6, 8x8, 10x10) and high-dimensional continuous control environments (Hopper-v4, Walker2d-v4) demonstrate that our approach yields more stable training dynamics and reduces exploitative behaviors under reward ambiguity, achieving a 93.7% reduction in reward-hacking behavior as measured by trap visitation frequency. We demonstrate statistical significance of these improvements and robustness under up to 30% supervisory noise, albeit with a trade-off in peak observed reward compared to unconstrained baselines. By treating uncertainty as a first-class component of the reward signal, this work offers a principled approach toward more reliable and aligned reinforcement learning systems.", "authors": ["Disha Singha"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-29", "url": "https://arxiv.org/abs/2604.26360", "pdf_url": "https://arxiv.org/pdf/2604.26360v1", "arxiv_id": "2604.26360", "doi": "10.48550/arXiv.2604.26360", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.55}
{"id": "9c20d095227a499cdf65ee8b3ff8e2716a9a92d2d07e4bb37d4efef9b2305b6b", "sources": ["arxiv", "semantic_scholar"], "title": "reward-lens: A Mechanistic Interpretability Library for Reward Models", "abstract": "Every RLHF-trained language model is shaped by a reward model, yet the mechanistic interpretability toolkit -- logit lens, direct logit attribution, activation patching, sparse autoencoders -- was built for generative LLMs whose primitives all project onto a vocabulary unembedding. Reward models replace that with a scalar regression head, breaking each tool. We present reward-lens, an open-source library that ports this toolkit to reward models, organised around one observation: the reward head's weight vector $w_r$ is the natural axis for every interpretability question. The library provides a Reward Lens, component attribution, three-mode activation patching, a reward-hacking probe suite, TopK SAE feature attribution, cross-model comparison, and five theory-grounded extensions (distortion index, divergence-aware patching, misalignment cascade detection, reward-term conflict analysis, concept-vector analysis). A ten-method adapter protocol covers Llama, Mistral, Gemma-2, and ArmoRM multi-objective heads, with a generic adapter for any HuggingFace sequence classification model. We validate on two production reward models across ~695 RewardBench pairs. The central empirical finding is negative: linear attribution does not predict causal patching effects (mean Spearman $ρ= -0.256$ on Skywork, $-0.027$ on ArmoRM). The framework treats this disagreement as a property to expose, not a bug -- motivating a design that keeps observational and causal views first-class and directly comparable.", "authors": ["Mohammed Suhail B Nadaf"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-28", "url": "https://arxiv.org/abs/2604.26130", "pdf_url": "https://arxiv.org/pdf/2604.26130v1", "arxiv_id": "2604.26130", "doi": "10.48550/arXiv.2604.26130", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/suhailnadaf509/reward-lens", "venue": "arXiv.org", "quality_score": 0.85}
{"id": "377546b10133d0877254ba3ade1fc5ec3e16068b9e9d8543a42d76bf986005f3", "sources": ["arxiv", "semantic_scholar"], "title": "Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis", "abstract": "Process Reward Models (PRMs) have achieved remarkable success in augmenting the reasoning capabilities of Large Language Models (LLMs) within static domains such as mathematics. However, their potential in dynamic data analysis tasks remains underexplored. In this work, we first present a empirical study revealing that general-domain PRMs struggle to supervise data analysis agents. Specifically, they fail to detect silent errors, logical flaws that yield incorrect results without triggering interpreter exceptions, and erroneously penalize exploratory actions, mistaking necessary trial-and-error exploration for grounding failures. To bridge this gap, we introduce DataPRM, a novel environment-aware generative process reward model that (1) can serve as an active verifier, autonomously interacting with the environment to probe intermediate execution states and uncover silent errors, and (2) employs a reflection-aware ternary reward strategy that distinguishes between correctable grounding errors and irrecoverable mistakes. We design a scalable pipeline to construct over 8K high-quality training instances for DataPRM via diversity-driven trajectory generation and knowledge-augmented step-level annotation. Experimental results demonstrate that DataPRM improves downstream policy LLMs by 7.21% on ScienceAgentBench and 11.28% on DABStep using Best-of-N inference. Notably, with only 4B parameters, DataPRM outperforms strong baselines, and exhibits robust generalizability across diverse Test-Time Scaling strategies. Furthermore, integrating DataPRM into Reinforcement Learning yields substantial gains over outcome-reward baselines, achieving 78.73% on DABench and 64.84% on TableBench, validating the effectiveness of process reward supervision. Code is available at https://github.com/zjunlp/DataMind.", "authors": ["Zhisong Qiu", "Shuofei Qiao", "Kewei Xu", "Yuqi Zhu", "Lun Du", "Ningyu Zhang", "Huajun Chen"], "categories": ["cs.CL", "cs.AI", "cs.CE", "cs.LG", "cs.MA"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-27", "url": "https://arxiv.org/abs/2604.24198", "pdf_url": "https://arxiv.org/pdf/2604.24198v1", "arxiv_id": "2604.24198", "doi": "10.48550/arXiv.2604.24198", "citation_count": 3, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/zjunlp/DataMind", "venue": "arXiv.org", "quality_score": 0.85}
{"id": "927c290a876589751e7f70d291e3b0fa85230a853ad3f9e6a605682bb7df10af", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Models Are Secretly Value Functions: Temporally Coherent Reward Modeling", "abstract": "Reward models in RLHF are trained to score only the final token of a response - a choice that discards rich signal from every intermediate position and produces models whose token-level outputs are noise. We argue this is a missed opportunity: a well-trained reward model's output at any token should represent the conditional expectation of the final reward given the response so far. We introduce Temporally Coherent Reward Modeling (TCRM), which induces this property via two regularization terms on top of the standard Bradley-Terry loss, with minimizers provably equal to conditional expectations. The regularizers correspond to Monte Carlo and TD value-learning objectives, establishing a direct connection to RL value functions. TCRM requires zero changes to architecture, data, or inference, yet unlocks three capabilities from one principle: interpretable token-level reward trajectories (middle-token pairwise accuracy improved from 50% to 88.9%, final-token accuracy preserved); state-of-the-art PRM performance on ProcessBench (44.9% average F1) among models trained only on outcome data; and unified reward/value modeling in PPO, reducing peak GPU memory by 27% and step time by 19% with matching LLM quality.", "authors": ["Alex Nikulkov"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-24", "url": "https://arxiv.org/abs/2604.22981", "pdf_url": "https://arxiv.org/pdf/2604.22981v1", "arxiv_id": "2604.22981", "doi": "10.48550/arXiv.2604.22981", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.55}
{"id": "0b523a07c46024bc53dcbeca2dc03d867eb2b77f0a0c8f9b84c0dac1b8d1575b", "sources": ["arxiv", "semantic_scholar"], "title": "K-Score: Kalman Filter as a Principled Alternative to Reward Normalization in Reinforcement Learning", "abstract": "We propose a simple yet effective alternative to reward normalization in policy gradient reinforcement learning by integrating a 1D Kalman filter for online reward estimation. Instead of relying on fixed heuristics, our method recursively estimates the latent reward mean, smoothing high-variance returns and adapting to non-stationary environments. This approach incurs minimal overhead and requires no modification to existing policy architectures. Experiments on \\textit{LunarLander} and \\textit{CartPole} demonstrate that Kalman-filtered rewards significantly accelerate convergence and reduce training variance compared to standard normalization techniques. Code is available at https://github.com/Sumxiaa/Kalman_Normalization.", "authors": ["Zixuan Xia", "Quanxi Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-24", "url": "https://arxiv.org/abs/2604.23056", "pdf_url": "https://arxiv.org/pdf/2604.23056v1", "arxiv_id": "2604.23056", "doi": "10.48550/arXiv.2604.23056", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Sumxiaa/Kalman_Normalization", "venue": "arXiv.org", "quality_score": 0.85}
{"id": "e750fa7fc69af813791eadc3212ebeb6415c72866aaf6181ca116f1976264c68", "sources": ["arxiv", "semantic_scholar"], "title": "Mitigating Multimodal Hallucination via Phase-wise Self-reward", "abstract": "Large Vision-Language Models (LVLMs) still struggle with vision hallucination, where generated responses are inconsistent with the visual input. Existing methods either rely on large-scale annotated data for fine-tuning, which incurs massive computational overhead, or employ static post-hoc strategies that overlook the dynamic nature of hallucination emergence. To address these, we introduce a new self-rewarding framework, enabling dynamic hallucination mitigation at inference time without external supervision. On the empirical side, we reveal that visual hallucination exhibits phase-wise dynamic patterns, peaking at the onset of each semantic phase. Drawing on these insights, we propose \\textbf{PSRD} (\\textbf{Phase-wise \\textbf{S}elf-\\textbf{R}eward \\textbf{D}ecoding) for online hallucination correction guided by phase-wise self-reward signals. To reduce the cost of repeated self-evaluation during decoding, we distill the hallucination guidance signal from LVLMs into a lightweight reward model. The reward model subsequently provides on-the-fly guidance for targeted intervention during the decoding process, enabling precise hallucination suppression. The proposed PSRD significantly reduces the hallucination rate of LLaVA-1.5-7B by 50.0% and consistently outperforms existing post-hoc methods across five hallucination evaluation benchmarks for four LVLMs. Further analysis confirms that PSRD effectively mitigates hallucination propagation and achieves a highly controllable trade-off between strong performance and inference efficiency.", "authors": ["Yu Zhang", "Chuyang Sun", "Kehai Chen", "Xuefeng Bai", "Yang Xiang", "Min Zhang"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-20", "url": "https://arxiv.org/abs/2604.17982", "pdf_url": "https://arxiv.org/pdf/2604.17982v1", "arxiv_id": "2604.17982", "doi": "10.48550/arXiv.2604.17982", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.55}
{"id": "488df9c33414a110b0eabcf74d54efec3f5de9688c65b631c60318667c9a5027", "sources": ["arxiv", "semantic_scholar"], "title": "PARM: Pipeline-Adapted Reward Model", "abstract": "Reward models (RMs) are central to aligning large language models (LLMs) with human preferences, powering RLHF and advanced decoding strategies. While most prior work focuses on single-step generation, real-world applications increasingly adopt multi-stage LLM pipelines, where effective reward guidance remains underexplored. We investigate this through code generation for combinatorial optimization, constructing a pipeline that integrates reward models into both formulation and solution stages. We identify a critical challenge: inconsistency between reward model predictions and actual pipeline execution outcomes. To address this, we propose the Pipeline-Adapted Reward Model (PARM), which leverages pipeline-specific data and direct preference optimization to align rewards with downstream feedback. We instantiate PARM as a two-stage pipeline (formulation -> code generation) and evaluate it on four public optimization benchmarks, measuring execution rate and solving accuracy against baselines and sampling methods. A supplementary cross-domain experiment on GSM8K assesses transferability. Results demonstrate that PARM consistently improves pipeline output quality and stability, providing new insights into reward modeling for multi-stage LLM reasoning.", "authors": ["Xingyu Fan", "Wei Shao", "Jiacheng Liu", "Linqi Song", "Pheng Ann Heng"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-20", "url": "https://arxiv.org/abs/2604.18327", "pdf_url": "https://arxiv.org/pdf/2604.18327v1", "arxiv_id": "2604.18327", "doi": "10.1109/jstsp.2026.3690098", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Journal on Selected Topics in Signal Processing", "quality_score": 0.55}
{"id": "5c49c09021c7d5ab0e7d77041e89345ccd7a3b431a2493251e6ab48b9d2a151f", "sources": ["arxiv", "semantic_scholar"], "title": "C2: Scalable Rubric-Augmented Reward Modeling from Binary Preferences", "abstract": "Rubric-augmented verification guides reward models with explicit evaluation criteria, yielding more reliable judgments than single-model verification. However, most existing methods require costly rubric annotations, limiting scalability. Moreover, we find that rubric generation is vulnerable to a failure of cooperation; low-quality rubrics actively mislead reward models rather than help. Inspired by the principle of cooperative communication, we propose Cooperative yet Critical reward modeling (C2), a framework that significantly improves reward model judgments by having the reward model critically collaborate with a rubric generator trained solely from binary preferences. In C2, we synthesize helpful and misleading rubric pairs by measuring how each rubric shifts the reward model toward or away from the correct preference. Using these contrastive pairs, we train a cooperative rubric generator to propose helpful rubrics, and a critical verifier to assess rubric validity before making its judgment, following only rubrics it deems helpful at inference time. C2 outperforms reasoning reward models trained on the same binary preferences, with gains of up to 6.5 points on RM-Bench and 6.0 points length-controlled win rate on AlpacaEval 2.0. Without external rubric annotations, C2 enables an 8B reward model to match performance achieved with rubrics from a 4$\\times$ larger model. Overall, our work demonstrates that eliciting deliberate cooperation in rubric-augmented verification makes reward models more trustworthy in a scalable way.", "authors": ["Akira Kawabata", "Saku Sugawara"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-15", "url": "https://arxiv.org/abs/2604.13618", "pdf_url": "https://arxiv.org/pdf/2604.13618v1", "arxiv_id": "2604.13618", "doi": "10.48550/arXiv.2604.13618", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5443}
{"id": "e6e74114ac20ae26049b171efcb6f3003d69530623b6d38241c7d727af695027", "sources": ["arxiv", "semantic_scholar"], "title": "Enhancing Reinforcement Learning for Radiology Report Generation with Evidence-aware Rewards and Self-correcting Preference Learning", "abstract": "Recent reinforcement learning (RL) approaches have advanced radiology report generation (RRG), yet two core limitations persist: (1) report-level rewards offer limited evidence-grounded guidance for clinical faithfulness; and (2) current methods lack an explicit self-improving mechanism to align with clinical preference. We introduce clinically aligned Evidence-aware Self-Correcting Reinforcement Learning (ESC-RL), comprising two key components. First, a Group-wise Evidence-aware Alignment Reward (GEAR) delivers group-wise, evidence-aware feedback. GEAR reinforces consistent grounding for true positives, recovers missed findings for false negatives, and suppresses unsupported content for false positives. Second, a Self-correcting Preference Learning (SPL) strategy automatically constructs a reliable, disease-aware preference dataset from multiple noisy observations and leverages an LLM to synthesize refined reports without human supervision. ESC-RL promotes clinically faithful, disease-aligned reward and supports continual self-improvement during training. Extensive experiments on two public chest X-ray datasets demonstrate consistent gains and state-of-the-art performance.", "authors": ["Qin Zhou", "Guoyan Liang", "Qianyi Yang", "Jingyuan Chen", "Sai Wu", "Chang Yao", "Zhe Wang"], "categories": ["cs.LG", "stat.ME"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-15", "url": "https://arxiv.org/abs/2604.13598", "pdf_url": "https://arxiv.org/pdf/2604.13598v1", "arxiv_id": "2604.13598", "doi": "10.48550/arXiv.2604.13598", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5443}
{"id": "70528fd8625a16e32b22bbb1e4e3f68891e0c39f400175acae7b0582a8805852", "sources": ["arxiv", "semantic_scholar"], "title": "DDO-RM: Distribution-Level Policy Improvement after Reward Learning", "abstract": "Recent theory suggests that reward-model-first methods can be more sample-efficient than direct policy fitting when the reward function is statistically simpler than the induced policy. We propose DDO-RM, a finite-candidate decision-optimization method that converts reward scores into an explicit target distribution. Unlike PPO-based RLHF or DPO, DDO-RM performs a KL-regularized mirror-descent update to project the policy toward a reward-improved distribution over a candidate set. Preliminary experiments on Pythia-410M show that DDO-RM outperforms DPO in pair accuracy (0.52 to 0.56) and mean margin (0.13 to 0.53). Our framework provides a principled connection between reward learning and mirror-descent policy improvement.", "authors": ["Tiantian Zhang", "Jierui Zuo", "Michael Chen", "Wenping Wang"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-13", "url": "https://arxiv.org/abs/2604.11119", "pdf_url": "https://arxiv.org/pdf/2604.11119v2", "arxiv_id": "2604.11119", "doi": "10.48550/arXiv.2604.11119", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.542}
{"id": "16b03ee8ce722b99ee0a46924915bfb54c790ff7296d08b3fecdbe2159c7dea8", "sources": ["arxiv", "semantic_scholar"], "title": "Mitigating Reward Hacking in RLHF via Advantage Sign Robustness", "abstract": "Reward models (RMs) used in reinforcement learning from human feedback (RLHF) are vulnerable to reward hacking: as the policy maximizes a learned proxy reward, true quality plateaus or degrades. We make the assumption that reward hacking is often caused by flipped advantage signs: instead of reducing the likelihood of a bad response, a flipped sign causes the update to increase it. By considering an adversarial perturbation in the RM parameter space, we can derive a certified sign-preservation radius, which is the smallest perturbation that can flip the advantage sign during policy optimization. Based on this formulation, we propose Sign-Certified Policy Optimization (SignCert-PO), down-weighting non-robust completions in the policy gradient update. Unlike prior approaches that require multiple RMs or access to the RM training data, SignCert-PO is lightweight and operates purely at the policy optimization stage using only the RM parameters and on-policy completions. On TL;DR summarization and AlpacaFarm benchmarks, SignCert-PO consistently achieves a better win rate than baselines and reduces reward hacking.", "authors": ["Shinnosuke Ono", "Johannes Ackermann", "Soichiro Nishimori", "Takashi Ishida", "Masashi Sugiyama"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-03", "url": "https://arxiv.org/abs/2604.02986", "pdf_url": "https://arxiv.org/pdf/2604.02986v1", "arxiv_id": "2604.02986", "doi": "10.48550/arXiv.2604.02986", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5305}
{"id": "3531a9a3b952f22acbd55e51d7eb154c37648c0905d56eea4ed2a553fd4755c0", "sources": ["arxiv", "semantic_scholar"], "title": "PAC-Bayesian Reward-Certified Outcome Weighted Learning", "abstract": "Estimating optimal individualized treatment rules (ITRs) via outcome weighted learning (OWL) often relies on observed rewards that are noisy or optimistic proxies for the true latent utility. Ignoring this reward uncertainty leads to the selection of policies with inflated apparent performance, yet existing OWL frameworks lack the finite-sample guarantees required to systematically embed such uncertainty into the learning objective. To address this issue, we propose PAC-Bayesian Reward-Certified Outcome Weighted Learning (PROWL). Given a one-sided uncertainty certificate, PROWL constructs a conservative reward and a strictly policy-dependent lower bound on the true expected value. Theoretically, we prove an exact certified reduction that transforms robust policy learning into a unified, split-free cost-sensitive classification task. This formulation enables the derivation of a nonasymptotic PAC-Bayes lower bound for randomized ITRs, where we establish that the optimal posterior maximizing this bound is exactly characterized by a general Bayes update. To overcome the learning-rate selection problem inherent in generalized Bayesian inference, we introduce a fully automated, bounds-based calibration procedure, coupled with a Fisher-consistent certified hinge surrogate for efficient optimization. Our experiments demonstrate that PROWL achieves improvements in estimating robust, high-value treatment regimes under severe reward uncertainty compared to standard methods for ITR estimation.", "authors": ["Yuya Ishikawa", "Shu Tamano"], "categories": ["cs.LG", "stat.ME", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-04-02", "url": "https://arxiv.org/abs/2604.01946", "pdf_url": "https://arxiv.org/pdf/2604.01946v1", "arxiv_id": "2604.01946", "doi": "10.48550/arXiv.2604.01946", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5294}
{"id": "2c74fde5a11989ebf10a786fa84414f456c829735004804bedbb79fbd43a5c6e", "sources": ["arxiv", "semantic_scholar"], "title": "Preference learning in shades of gray: Interpretable and bias-aware reward modeling for human preferences", "abstract": "Learning human preferences in language models remains fundamentally challenging, as reward modeling relies on subtle, subjective comparisons or shades of gray rather than clear-cut labels. This study investigates the limits of current approaches and proposes a feature-augmented framework to better capture the multidimensional nature of human judgment. Using the Anthropic HHRLHF dataset, we evaluate ten diverse large language models LLMs under a standard pairwise preference setting, where baseline performance remains below 0.74 ROC AUC, highlighting the difficulty of the task. To address this, we enrich textual representations with interpretable signals: response length, refusal indicators, toxicity scores and prompt response semantic similarity, enabling models to explicitly capture key aspects of helpfulness, safety and relevance. The proposed hybrid approach yields consistent improvements across all models, achieving up to 0.84 ROC AUC and significantly higher pairwise accuracy, with DeBERTav3Large demonstrating the best performance. Beyond accuracy, we integrate SHAP and LIME to provide fine-grained interpretability, revealing that model decisions depend on contextualized safety and supportive framing rather than isolated keywords. We further analyze bias amplification, showing that while individual features have weak marginal effects, their interactions influence preference learning.", "authors": ["Simona-Vasilica Oprea", "Adela Bâra"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-04-01", "url": "https://arxiv.org/abs/2604.01312", "pdf_url": "https://arxiv.org/pdf/2604.01312v1", "arxiv_id": "2604.01312", "doi": "10.48550/arXiv.2604.01312", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5282}
{"id": "d55bdbfe95a58da4928d8d5dd6786c6ca701f36332954ef6a0076eb23830074e", "sources": ["arxiv", "semantic_scholar"], "title": "ImplicitRM: Unbiased Reward Modeling from Implicit Preference Data for LLM alignment", "abstract": "Reward modeling represents a long-standing challenge in reinforcement learning from human feedback (RLHF) for aligning language models. Current reward modeling is heavily contingent upon experimental feedback data with high collection costs. In this work, we study \\textit{implicit reward modeling} -- learning reward models from implicit human feedback (e.g., clicks and copies) -- as a cost-effective alternative. We identify two fundamental challenges in implicit reward modeling: (1) Implicit preference data lacks definitive negative samples, which makes standard positive-negative classification methods inapplicable; (2) Implicit preference data suffers from user preference bias, where different responses have different propensities to elicit user feedback actions, which exacerbates the difficulty of distinguishing definitive negative samples. To address these challenges, we propose ImplicitRM, which aims to learn unbiased reward models from implicit preference data. ImplicitRM stratifies training samples into four latent groups via a stratification model. Building on this, it derives a learning objective through likelihood maximization, which we prove is theoretically unbiased, effectively resolving both challenges. Experiments demonstrate that ImplicitRM learns accurate reward models across implicit preference datasets. Code is available on our project website.", "authors": ["Hao Wang", "Haocheng Yang", "Licheng Pan", "Lei Shen", "Xiaoxi Li", "Yinuo Wang", "Zhichao Chen", "Yuan Lu", "Haoxuan Li", "Zhouchen Lin"], "categories": ["cs.CL", "cs.AI", "stat.AP"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-03-24", "url": "https://arxiv.org/abs/2603.23184", "pdf_url": "https://arxiv.org/pdf/2603.23184v1", "arxiv_id": "2603.23184", "doi": "10.48550/arXiv.2603.23184", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.8022}
{"id": "6f6a1eaef745956cf726a27aac38d02bc26068dd2fb7012368029ea1781768eb", "sources": ["arxiv", "semantic_scholar"], "title": "Privacy-Preserving Reinforcement Learning from Human Feedback via Decoupled Reward Modeling", "abstract": "Preference-based fine-tuning has become an important component in training large language models, and the data used at this stage may contain sensitive user information. A central question is how to design a differentially private pipeline that is well suited to the distinct structure of reinforcement learning from human feedback. We propose a privacy-preserving framework that imposes differential privacy only on reward learning and derives the final policy from the resulting private reward model. Theoretically, we study the suboptimality gap and show that privacy contributes an additional additive term beyond the usual non-private statistical error. We also establish a minimax lower bound and show that the dominant term changes with sample size and privacy level, which in turn characterizes regimes in which the upper bound is rate-optimal up to logarithmic factors. Empirically, synthetic experiments confirm the scaling predicted by the theory, and experiments on the Anthropic HH-RLHF dataset using the Gemma-2B-IT model show stronger private alignment performance than existing differentially private baseline methods across privacy budgets.", "authors": ["Young Hyun Cho", "Will Wei Sun"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-03-23", "url": "https://arxiv.org/abs/2603.22563", "pdf_url": "https://arxiv.org/pdf/2603.22563v1", "arxiv_id": "2603.22563", "doi": "10.48550/arXiv.2603.22563", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5179}
{"id": "04ba9ecb076ea88060dd8beac0e4df9f84ceda7cedf0cff73be6d3d24c05c7d0", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Sharpness-Aware Fine-Tuning for Diffusion Models", "abstract": "Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models with human preferences, inspiring the development of reward-centric diffusion reinforcement learning (RDRL) to achieve similar alignment and controllability. While diffusion models can generate high-quality outputs, RDRL remains susceptible to reward hacking, where the reward score increases without corresponding improvements in perceptual quality. We demonstrate that this vulnerability arises from the non-robustness of reward model gradients, particularly when the reward landscape with respect to the input image is sharp. To mitigate this issue, we introduce methods that exploit gradients from a robustified reward model without requiring its retraining. Specifically, we employ gradients from a flattened reward model, obtained through parameter perturbations of the diffusion model and perturbations of its generated samples. Empirically, each method independently alleviates reward hacking and improves robustness, while their joint use amplifies these benefits. Our resulting framework, RSA-FT (Reward Sharpness-Aware Fine-Tuning), is simple, broadly compatible, and consistently enhances the reliability of RDRL.", "authors": ["Kwanyoung Kim", "Byeongsu Sim"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-22", "url": "https://arxiv.org/abs/2603.21175", "pdf_url": "https://arxiv.org/pdf/2603.21175v1", "arxiv_id": "2603.21175", "doi": "10.48550/arXiv.2603.21175", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5168}
{"id": "b7755cda26a08d6a0a98515a4029778a35d0c5354bd55321b4cd2f6d9c77f4d1", "sources": ["arxiv", "semantic_scholar"], "title": "CausalRM: Causal-Theoretic Reward Modeling for RLHF from Observational User Feedbacks", "abstract": "Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models, current reward modeling heavily relies on experimental feedback data collected from human annotators under controlled and costly conditions. In this work, we introduce observational reward modeling -- learning reward models with observational user feedback (e.g., clicks, copies, and upvotes) -- as a scalable and cost-effective alternative. We identify two fundamental challenges in this setting: (1) observational feedback is noisy due to annotation errors, which deviates it from true user preference; (2) observational feedback is biased by user preference, where users preferentially provide feedback on responses they feel strongly about, which creats a distribution shift between training and inference data. To address these challenges, we propose CausalRM, a causal-theoretic reward modeling framework that aims to learn unbiased reward models from observational feedback. To tackle challenge (1), CausalRM introduces a noise-aware surrogate loss term that is provably equivalent to the primal loss under noise-free conditions by explicitly modeling the annotation error generation process. To tackle challenge (2), CausalRM uses propensity scores -- the probability of a user providing feedback for a given response -- to reweight training samples, yielding a loss function that eliminates user preference bias. Extensive experiments across diverse LLM backbones and benchmark datasets validate that CausalRM effectively learns accurate reward signals from noisy and biased observational feedback and delivers substantial performance improvements on downstream RLHF tasks -- including a 49.2% gain on WildGuardMix and a 32.7% improvement on HarmBench. Code is available on our project website.", "authors": ["Hao Wang", "Licheng Pan", "Zhichao Chen", "Chunyuan Zheng", "Zhixuan Chu", "Xiaoxi Li", "Yuan Lu", "Xinggao Liu", "Haoxuan Li", "Zhouchen Lin"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-03-19", "url": "https://arxiv.org/abs/2603.18736", "pdf_url": "https://arxiv.org/pdf/2603.18736v1", "arxiv_id": "2603.18736", "doi": "10.48550/arXiv.2603.18736", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7933}
{"id": "81a02f3ea5e263070c212c956e961ab23ba89b52a58069ec9d8bb2b7f2027cae", "sources": ["arxiv", "semantic_scholar"], "title": "Robust Post-Training for Generative Recommenders: Why Exponential Reward-Weighted SFT Outperforms RLHF", "abstract": "Aligning generative recommender systems to user preferences via post-training is critical for closing the gap between next-item prediction and actual recommendation quality. Existing post-training methods are ill-suited for production-scale systems: RLHF methods reward hack due to noisy user feedback and unreliable reward models, offline RL alternatives require propensity scores that are unavailable, and online interaction is infeasible. We identify exponential reward-weighted SFT with weights $w = \\exp(r/λ)$ as uniquely suited to this setting, and provide the theoretical and empirical foundations that explain why. By optimizing directly on observed rewards without querying a learned reward model, the method is immune to reward hacking, requires no propensity scores, and is fully offline. We prove the first policy improvement guarantees for this setting under noisy rewards, showing that the gap scales only logarithmically with catalog size and remains informative even for large item catalogs. Crucially, we show that temperature $λ$ explicitly and quantifiably controls the robustness-improvement tradeoff, providing practitioners with a single interpretable regularization hyperparameter with theoretical grounding. Experiments on three open-source and one proprietary dataset against four baselines confirm that exponential reward weighting is simple, scalable, and consistently outperforms RLHF-based alternatives.", "authors": ["Keertana Chidambaram", "Sanath Kumar Krishnamurthy", "Qiuling Xu", "Ko-Jen Hsiao", "Moumita Bhattacharya"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-10", "url": "https://arxiv.org/abs/2603.10279", "pdf_url": "https://arxiv.org/pdf/2603.10279v1", "arxiv_id": "2603.10279", "doi": "10.48550/arXiv.2603.10279", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7774}
{"id": "4219af8f68a24e0e4bf67a43c1224bcd7f0313b8860afb03c7427c02a32a27fa", "sources": ["arxiv", "semantic_scholar"], "title": "Causally Robust Reward Learning from Reason-Augmented Preference Feedback", "abstract": "Preference-based reward learning is widely used for shaping agent behavior to match a user's preference, yet its sparse binary feedback makes it especially vulnerable to causal confusion. The learned reward often latches onto spurious features that merely co-occur with preferred trajectories during training, collapsing when those correlations disappear or reverse at test time. We introduce ReCouPLe, a lightweight framework that uses natural language rationales to provide the missing causal signal. Each rationale is treated as a guiding projection axis in an embedding space, training the model to score trajectories based on features aligned with that axis while de-emphasizing context that is unrelated to the stated reason. Because the same rationales (e.g., \"avoids collisions\", \"completes the task faster\") can appear across multiple tasks, ReCouPLe naturally reuses the same causal direction whenever tasks share semantics, and transfers preference knowledge to novel tasks without extra data or language-model fine-tuning. Our learned reward model can ground preferences on the articulated reason, aligning better with user intent and generalizing beyond spurious features. ReCouPLe outperforms baselines by up to 1.5x in reward accuracy under distribution shifts, and 2x in downstream policy performance in novel tasks. We have released our code at https://github.com/mj-hwang/ReCouPLe", "authors": ["Minjune Hwang", "Yigit Korkmaz", "Daniel Seita", "Erdem Bıyık"], "categories": ["cs.AI", "cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-05", "url": "https://arxiv.org/abs/2603.04861", "pdf_url": "https://arxiv.org/pdf/2603.04861v1", "arxiv_id": "2603.04861", "doi": "10.48550/arXiv.2603.04861", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/mj-hwang/ReCouPLe", "venue": "arXiv.org", "quality_score": 0.7685}
{"id": "cce5dccf585671458d0f4fd545b84948f2cf4e2cf04b345d1f7b4ff00cb5013f", "sources": ["arxiv", "semantic_scholar"], "title": "VRM: Teaching Reward Models to Understand Authentic Human Preferences", "abstract": "Large Language Models (LLMs) have achieved remarkable success across diverse natural language tasks, yet the reward models employed for aligning LLMs often encounter challenges of reward hacking, where the approaches predominantly rely on directly mapping prompt-response pairs to scalar scores, which may inadvertently capture spurious correlations rather than authentic human preferences. In contrast, human evaluation employs a sophisticated process that initially weighs the relative importance of multiple high-dimensional objectives according to the prompt context, subsequently evaluating response quality through low-dimensional semantic features such as logical coherence and contextual appropriateness. Motivated by this consideration, we propose VRM, i.e., Variational Reward Modeling, a novel framework that explicitly models the evaluation process of human preference judgments by incorporating both high-dimensional objective weights and low-dimensional semantic features as latent variables, which are inferred through variational inference techniques. Additionally, we provide a theoretical analysis showing that VRM can achieve a tighter generalization error bound compared to the traditional reward model. Extensive experiments on benchmark datasets demonstrate that VRM significantly outperforms existing methods in capturing authentic human preferences.", "authors": ["Biao Liu", "Ning Xu", "Junming Yang", "Hao Xu", "Xin Geng"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-05", "url": "https://arxiv.org/abs/2603.04974", "pdf_url": "https://arxiv.org/pdf/2603.04974v1", "arxiv_id": "2603.04974", "doi": "10.48550/arXiv.2603.04974", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4973}
{"id": "3064227dd486b1b97e8bb41ab984bd7db9db8840a0d2c571e27e0835be210f96", "sources": ["arxiv", "semantic_scholar"], "title": "Reward-Conditioned Reinforcement Learning", "abstract": "Single-task RL agents are typically trained under a fixed reward function, which limits their robustness to reward misspecification and their ability to adapt to changing preferences. We introduce Reward-Conditioned Reinforcement Learning (RCRL), an off-policy method that conditions agents on reward parameterizations while collecting experience under a single nominal objective. By recomputing counterfactual rewards from shared replay data, RCRL exposes the agent to multiple reward objectives without additional environment interaction, connecting single-task RL with ideas from multi-objective and multi-task learning. Across single-task, multi-task, and vision-based benchmarks, RCRL improves sample efficiency under the nominal reward parameterization, enables efficient adaptation to new parameterizations, and supports zero-shot behavioral adjustment at deployment. Our results show that RCRL provides a scalable mechanism for learning robust, steerable policies without sacrificing the simplicity of single-task training.", "authors": ["Michal Nauman", "Marek Cygan", "Pieter Abbeel"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-05", "url": "https://arxiv.org/abs/2603.05066", "pdf_url": "https://arxiv.org/pdf/2603.05066v3", "arxiv_id": "2603.05066", "doi": "10.48550/arXiv.2603.05066", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4973}
{"id": "505aff0dcfb0d2d18e2f487adb8f45a371efe9008837e2b55dbb6b690656042e", "sources": ["arxiv", "semantic_scholar"], "title": "Alternating Reinforcement Learning with Contextual Rubric Rewards: Beyond the Scalarization Strategy", "abstract": "Reinforcement Learning with Rubric Rewards (RLRR) is a framework that extends conventional reinforcement learning from human feedback (RLHF) and verifiable rewards (RLVR) by replacing scalar preference signals with structured, multi-dimensional, contextual rubric-based evaluations. However, existing approaches in RLRR are limited to linearly compressing vector rewards into a scalar reward with a fixed weightings, which is sensitive to artificial score design and fails to capture correlations among reward dimensions. To overcome the limitations of reward aggregation, this work proposes Alternating Reinforcement Learning with Rubric Rewards (ARL-RR), a framework that eliminates the need for a fixed scalarization by optimizing one semantic rubric meta-class at a time. Theoretically, we show that reward aggregation induces a variance contraction effect, which helps explain the performance gains. We further introduce a lightweight, search-based adaptation procedure that selects the next meta-class dynamically based on task performance, enabling the policy to emphasize critical objectives and thereby improve the model performance. Empirically, our experiments on the HealthBench dataset with experts annotations demonstrate that ARL-RR uniformly outperforms scalarized methods in both model performance and training efficiency across different model scales (1.7B, 4B, 8B, and 14B).", "authors": ["Guangchen Lan", "Lian Xiong", "Xin Zhou", "Hejie Cui", "Yuwei Zhang", "Mao Li", "Zhenyu Shi", "Besnik Fetahu", "Lihong Li", "Xian Li"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-03-04", "url": "https://arxiv.org/abs/2603.15646", "pdf_url": "https://arxiv.org/pdf/2603.15646v2", "arxiv_id": "2603.15646", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3157}
{"id": "d6a0a3b1cce8eeccf718a8b38a227fd78986f08c14b14ed109873e77cf55dbea", "sources": ["arxiv", "semantic_scholar"], "title": "Generalisation of RLHF under Reward Shift and Clipped KL Regularisation", "abstract": "Alignment and adaptation in large language models heavily rely on reinforcement learning from human feedback (RLHF); yet, theoretical understanding of its generalisability remains premature, especially when the learned reward could shift, and the KL control is estimated and clipped. To address this issue, we develop generalisation theory for RLHF that explicitly accounts for (1) \\emph{reward shift}: reward models are trained on preference data from earlier or mixed behaviour policies while RLHF optimises the current policy on its own rollouts; and (2) \\emph{clipped KL regularisation}: the KL regulariser is estimated from sampled log-probability ratios and then clipped for stabilisation, resulting in an error to RLHF. We present generalisation bounds for RLHF, suggesting that the generalisation error stems from a sampling error from prompts and rollouts, a reward shift error, and a KL clipping error. We also discuss special cases of (1) initialising RLHF parameters with a uniform prior over a finite space, and (2) training RLHF by stochastic gradient descent, as an Ornstein-Uhlenbeck process. The theory yields practical implications in (1) optimal KL clipping threshold, and (2) budget allocation in prompts, rollouts, and preference data.", "authors": ["Kenton Tang", "Yuzhu Chen", "Fengxiang He"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-02-25", "url": "https://arxiv.org/abs/2602.21765", "pdf_url": "https://arxiv.org/pdf/2602.21765v1", "arxiv_id": "2602.21765", "doi": "10.48550/arXiv.2602.21765", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4881}
{"id": "e37d453a9265820365d533c683bd4c97821ea0ed6eb7c87d3b748451d57929a7", "sources": ["arxiv", "semantic_scholar"], "title": "The Art of Efficient Reasoning: Data, Reward, and Optimization", "abstract": "Large Language Models (LLMs) consistently benefit from scaled Chain-of-Thought (CoT) reasoning, but also suffer from heavy computational overhead. To address this issue, efficient reasoning aims to incentivize short yet accurate thinking trajectories, typically through reward shaping with Reinforcement Learning (RL). In this paper, we systematically investigate the mechanics of efficient reasoning for LLMs. For comprehensive evaluation, we advocate for more fine-grained metrics, including length distribution conditioned on correctness and performance across a wide spectrum of token budgets ranging from 2k to 32k. First, we reveal that the training process follows a two-stage paradigm: length adaptation and reasoning refinement. Through extensive experiments (about 0.2 million GPU hours) in a unified protocol, we deconstruct training prompts and rollouts, reward shaping, and optimization strategies. A central finding is to maintain a sufficient density of positive reward signals and avoid the short-is-correct trap. Moreover, the learned length bias generalizes across domains and difficulty levels. We distill these findings into valuable insights and practical guidelines, and validate them across the Qwen3 models ranging from 0.6B to 30B, demonstrating the robustness and generalization. Weights are available at https://wutaiqiang.github.io/project/Art", "authors": ["Taiqiang Wu", "Zenan Xu", "Bo Zhou", "Ngai Wong"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-24", "url": "https://arxiv.org/abs/2602.20945", "pdf_url": "https://arxiv.org/pdf/2602.20945v3", "arxiv_id": "2602.20945", "doi": "10.48550/arXiv.2602.20945", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.487}
{"id": "33167b1d802e267f69577dc16b98db9db2e288d66559296187d9cbbcf0a686ed", "sources": ["arxiv", "semantic_scholar"], "title": "IR$^3$: Contrastive Inverse Reinforcement Learning for Interpretable Detection and Mitigation of Reward Hacking", "abstract": "Reinforcement Learning from Human Feedback (RLHF) enables powerful LLM alignment but can introduce reward hacking - models exploit spurious correlations in proxy rewards without genuine alignment. Compounding this, the objectives internalized during RLHF remain opaque, making hacking behaviors difficult to detect or correct. We introduce IR3 (Interpretable Reward Reconstruction and Rectification), a framework that reverse-engineers, interprets, and surgically repairs the implicit objectives driving RLHF-tuned models. We propose Contrastive Inverse Reinforcement Learning (C-IRL), which reconstructs the implicit reward function by contrasting paired responses from post-alignment and baseline policies to explain behavioral shifts during RLHF. We then decompose the reconstructed reward via sparse autoencoders into interpretable features, enabling identification of hacking signatures through contribution analysis. Finally, we propose mitigation strategies - clean reward optimization, adversarial shaping, constrained optimization, and feature-guided distillation - that target problematic features while preserving beneficial alignment. Experiments across multiple reward model configurations show that IR3 achieves 0.89 correlation with ground-truth rewards, identifies hacking features with over 90% precision, and significantly reduces hacking behaviors while maintaining capabilities within 3% of the original model.", "authors": ["Mohammad Beigi", "Ming Jin", "Junshan Zhang", "Jiaxin Zhang", "Qifan Wang", "Lifu Huang"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-23", "url": "https://arxiv.org/abs/2602.19416", "pdf_url": "https://arxiv.org/pdf/2602.19416v1", "arxiv_id": "2602.19416", "doi": "10.48550/arXiv.2602.19416", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4858}
{"id": "a7f61a901a3e4b2026e4f73c945d6927b84a80ea4fb05588971d257cbfaf9fe9", "sources": ["arxiv", "semantic_scholar"], "title": "Gradient Regularization Prevents Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards", "abstract": "Reinforcement Learning from Human Feedback (RLHF) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works address this by limiting the policy update with a Kullback-Leibler (KL) penalty towards a reference model. We propose a different framing: Train the LM in a way that biases policy updates towards regions in which the reward is more accurate. First, we derive a theoretical connection between the accuracy of a reward model and the flatness of an optimum at convergence. Gradient regularization (GR) can then be used to bias training to flatter regions and thereby maintain reward model accuracy. We confirm these results by showing that the gradient norm and reward accuracy are empirically correlated in RLHF. We then show that Reference Resets of the KL penalty implicitly use GR to find flatter regions with higher reward accuracy. We further improve on this by proposing to use explicit GR with an efficient finite-difference estimate. Empirically, GR performs better than a KL penalty across a diverse set of RL experiments with LMs. GR achieves a higher GPT-judged win-rate in RLHF, avoids overly focusing on the format in rule-based math rewards, and prevents hacking the judge in LLM-as-a-Judge math tasks.", "authors": ["Johannes Ackermann", "Michael Noukhovitch", "Takashi Ishida", "Masashi Sugiyama"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-20", "url": "https://arxiv.org/abs/2602.18037", "pdf_url": "https://arxiv.org/pdf/2602.18037v1", "arxiv_id": "2602.18037", "doi": "10.48550/arXiv.2602.18037", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4824}
{"id": "67dcf1b604136a621eb4fa56229acea84ff4ea6f4fbc42cabaaa8e1ec55a6d5a", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Under Attack: Analyzing the Robustness and Hackability of Process Reward Models", "abstract": "Process Reward Models (PRMs) are rapidly becoming the backbone of LLM reasoning pipelines, yet we demonstrate that state-of-the-art PRMs are systematically exploitable under adversarial optimization pressure. To address this, we introduce a three-tiered diagnostic framework that applies increasing adversarial pressure to quantify these vulnerabilities. Static perturbation analysis uncovers a fluency-logic dissociation: high invariance to surface-level style changes reward changes $<$0.1, yet inconsistent detection of logically-corrupted reasoning, with different models failing on different attack types. Adversarial optimization demonstrates that gradient-based attacks inflate rewards on invalid trajectories, with reward landscapes exhibiting wide, exploitable peaks. RL-induced reward hacking exposes the critical failure mode: policies trained on AIME problems achieve near-perfect PRM rewards ($>$0.9), while ground-truth accuracy remains low (below 4%), with 43% of reward gains attributable to stylistic shortcuts. These findings reveal that current PRMs function as fluency detectors rather than reasoning verifiers, creating systematic blind spots that undermine their use as training signals. We release PRM-BiasBench and a diagnostic toolkit to enable robustness evaluation before deployment. The code and dataset are available at https://github.com/SqueezeAILab/reward-under-attack.", "authors": ["Rishabh Tiwari", "Aditya Tomar", "Udbhav Bamba", "Monishwaran Maheswaran", "Heng Yang", "Michael W. Mahoney", "Kurt Keutzer", "Amir Gholami"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-20", "url": "https://arxiv.org/abs/2603.06621", "pdf_url": "https://arxiv.org/pdf/2603.06621v1", "arxiv_id": "2603.06621", "doi": "10.48550/arXiv.2603.06621", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/SqueezeAILab/reward-under-attack", "venue": "arXiv.org", "quality_score": 0.7455}
{"id": "86b41c50c6a849d49e1748be05dc3ba199cbcfde880e33334135121ede283f37", "sources": ["arxiv", "semantic_scholar"], "title": "MARS: Margin and Semantic-Aware Data Augmentation for Reward Modeling", "abstract": "Reward modeling is central to alignment pipelines such as RLHF, RLAIF, and PPO-based policy optimization, yet its reliability is constrained by limited and heterogeneous human preference data that are expensive to collect at scale. While synthetic augmentation can expand preference supervision, existing methods often augment uniformly or at the representation level, without targeting examples where the reward model is uncertain or prone to mis-ranking. In this paper, we introduce MARS (Margin and Semantic-Aware Data Augmentation for Reward Modeling), an adaptive augmentation framework that prioritizes low-margin preference pairs and uses semantic distance as a second layer for refinement to enhance the contrast between the chosen and rejected responses. Across multiple preference datasets, reward-model backbones, downstream alignment settings, and benchmarks including RewardBench and AlpacaEval, MARS improves both reward-model quality and alignment performance over existing baselines. Our results show that reward-model augmentation is most effective when guided by both model margins and semantic structure.", "authors": ["Payel Bhattacharjee", "Osvaldo Simeone", "Ravi Tandon"], "categories": ["cs.LG", "cs.AI", "cs.IT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2026-02-19", "url": "https://arxiv.org/abs/2602.17658", "pdf_url": "https://arxiv.org/pdf/2602.17658v2", "arxiv_id": "2602.17658", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3062}
{"id": "93dd90acd3e8e82783f0433ea7131f25dc0584e13f079e198ace2776a57f6b6d", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Binary Preferences: A Principled Framework for Reward Modeling with Ordinal Feedback", "abstract": "Reward modeling is crucial for aligning large language models with human preferences, yet current approaches lack a principled mathematical framework for leveraging ordinal preference data. When human annotators provide graded preferences on a Likert scale (e.g., significantly better, better, slightly better, negligibly better), existing methods typically apply ad-hoc heuristics, such as margin terms or scaling factors, to loss functions derived from binary preference models like Bradley-Terry. These approaches lack an underlying mathematical model for how ordinal preference data is generated. We present a theoretically grounded framework that formulates reward modeling with Likert scale preferences as a discrete ordinal regression problem. We derive two loss functions from this formulation: a negative log-likelihood loss and an all-threshold loss, both of which learn threshold parameters that naturally capture the ordinal structure of preferences. Unlike existing heuristic methods that manually specify fixed margins or scaling weights, our approach learns these parameters directly from data within a coherent probabilistic framework. Experimental results on multiple benchmarks demonstrate that our ordinal regression approach consistently achieves competitive or superior performance compared to existing heuristic methods across diverse evaluation categories including chat, reasoning, and safety tasks. Our work provides the first principled mathematical framework for incorporating Likert scale preferences into reward model training, moving beyond ad-hoc modifications of binary preference models to enable more effective utilization of fine-grained human feedback.", "authors": ["Amirhossein Afsharrad", "Ruida Zhou", "Luca Viano", "Sanjay Lall", "Mohammad Ghavamzadeh"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-13", "url": "https://arxiv.org/abs/2603.02232", "pdf_url": "https://arxiv.org/pdf/2603.02232v1", "arxiv_id": "2603.02232", "doi": "10.48550/arXiv.2603.02232", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4744}
{"id": "6fca7a4418a93f8d902956a2a6d9644a190a1df0a5054e883259700a30996269", "sources": ["arxiv", "semantic_scholar"], "title": "Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling", "abstract": "Reward models learned from human preferences are central to aligning large language models (LLMs) via reinforcement learning from human feedback, yet they are often vulnerable to reward hacking due to noisy annotations and systematic biases such as response length or style. We propose Bayesian Non-Negative Reward Model (BNRM), a principled reward modeling framework that integrates non-negative factor analysis into Bradley-Terry (BT) preference model. BNRM represents rewards through a sparse, non-negative latent factor generative process that operates at two complementary levels: instance-specific latent variables induce disentangled reward representations, while sparsity over global latent factors acts as an implicit debiasing mechanism that suppresses spurious correlations. Together, this disentanglement-then-debiasing structure enables robust uncertainty-aware reward learning. To scale BNRM to modern LLMs, we develop an amortized variational inference network conditioned on deep model representations, allowing efficient end-to-end training. Extensive empirical results demonstrate that BNRM substantially mitigates reward over-optimization, improves robustness under distribution shifts, and yields more interpretable reward decompositions than strong baselines.", "authors": ["Zhibin Duan", "Guowei Rong", "Zhuo Li", "Bo Chen", "Mingyuan Zhou", "Dandan Guo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-11", "url": "https://arxiv.org/abs/2602.10623", "pdf_url": "https://arxiv.org/pdf/2602.10623v2", "arxiv_id": "2602.10623", "doi": "10.48550/arXiv.2602.10623", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/GuoweiRong/Bayesian-Non-negative-Reward-Model", "venue": "arXiv.org", "quality_score": 0.7296}
{"id": "c7a50c90ecb99ef8c56cee427a936ace5443babf80db584a72b4b6dbe7531a36", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Modeling for Reinforcement Learning-Based LLM Reasoning: Design, Challenges, and Evaluation", "abstract": "Large Language Models (LLMs) demonstrate transformative potential, yet their reasoning remains inconsistent and unreliable. Reinforcement learning (RL)-based fine-tuning is a key mechanism for improvement, but its effectiveness is fundamentally governed by reward design. Despite its importance, the relationship between reward modeling and core LLM challenges--such as evaluation bias, hallucination, distribution shift, and efficient learning--remains poorly understood. This work argues that reward modeling is not merely an implementation detail but a central architect of reasoning alignment, shaping what models learn, how they generalize, and whether their outputs can be trusted. We introduce Reasoning-Aligned Reinforcement Learning (RARL), a unifying framework that systematizes diverse reward paradigms for multi-step reasoning. Within this framework, we present a taxonomy of reward mechanisms, analyze reward hacking as a pervasive failure mode, and examine how reward signals unify challenges ranging from inference-time scaling to hallucination mitigation. We further critically evaluate existing benchmarks, highlighting vulnerabilities such as data contamination and reward misalignment, and outline directions for more robust evaluation. By integrating fragmented research threads and clarifying the interplay between reward design and fundamental reasoning capabilities, this work provides a foundational roadmap for building reasoning models that are robust, verifiable, and trustworthy.", "authors": ["Pei-Chi Pan", "Yingbin Liang", "Sen Lin"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-10", "url": "https://arxiv.org/abs/2602.09305", "pdf_url": "https://arxiv.org/pdf/2602.09305v1", "arxiv_id": "2602.09305", "doi": "10.48550/arXiv.2602.09305", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4709}
{"id": "daf167b4a8054a79bd16fae3e332d9df12e9ca647dbb265613b596351fe274cf", "sources": ["arxiv", "semantic_scholar"], "title": "Bayesian Preference Learning for Test-Time Steerable Reward Models", "abstract": "Reward models are central to aligning language models with human preferences via reinforcement learning (RL). As RL is increasingly applied to settings such as verifiable rewards and multi-objective alignment, RMs are expected to encode more complex and multifaceted preference distributions. However, classifier RMs remain static once trained, limiting their adaptability at test time. We propose Variational In-Context Reward Modeling (ICRM), a novel Bayesian reward modeling objective that enables test-time steerability via in-context preference demonstrations. ICRM casts reward modeling as amortized variational inference over a latent preference probability under the Bradley-Terry model using a conjugate Beta prior. We show that ICRM adapts to unseen preference distributions at test time for both single and multi-objective settings. With more demonstrations, ICRM improves RM-Bench accuracy from 60.5 to 70.8, achieves lower calibration error than a generative judge on moral dilemma preferences, and expands the attainable Pareto frontier under conflicting preferences. We further study the practical applicability of ICRM for RL training, showing that it can effectively encode verifiable rewards by outperforming a conventional RM in math reasoning. Finally, we provide theoretical guarantees that the variational objective admits a global interior optimum with finite confidence, and we analyze how KL regularization mitigates reward over-optimization.", "authors": ["Jiwoo Hong", "Shao Tang", "Zhipeng Wang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.08819", "pdf_url": "https://arxiv.org/pdf/2602.08819v2", "arxiv_id": "2602.08819", "doi": "10.48550/arXiv.2602.08819", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4698}
{"id": "45d3d9274615e97b8050cbbb5f3a3ecc738b9f6380f2615e802b5c330e39acd2", "sources": ["arxiv", "semantic_scholar"], "title": "Learning in Context, Guided by Choice: A Reward-Free Paradigm for Reinforcement Learning with Transformers", "abstract": "In-context reinforcement learning (ICRL) leverages the in-context learning capabilities of transformer models (TMs) to efficiently generalize to unseen sequential decision-making tasks without parameter updates. However, existing ICRL methods rely on explicit reward signals during pretraining, which limits their applicability when rewards are ambiguous, hard to specify, or costly to obtain. To overcome this limitation, we propose a new learning paradigm, In-Context Preference-based Reinforcement Learning (ICPRL), in which both pretraining and deployment rely solely on preference feedback, eliminating the need for reward supervision. We study two variants that differ in the granularity of feedback: Immediate Preference-based RL (I-PRL) with per-step preferences, and Trajectory Preference-based RL (T-PRL) with trajectory-level comparisons. We first show that supervised pretraining, a standard approach in ICRL, remains effective under preference-only context datasets, demonstrating the feasibility of in-context reinforcement learning using only preference signals. To further improve data efficiency, we introduce alternative preference-native frameworks for I-PRL and T-PRL that directly optimize TM policies from preference data without requiring reward signals nor optimal action labels.Experiments on dueling bandits, navigation, and continuous control tasks demonstrate that ICPRL enables strong in-context generalization to unseen tasks, achieving performance comparable to ICRL methods trained with full reward supervision.", "authors": ["Juncheng Dong", "Bowen He", "Moyang Guo", "Ethan X. Fang", "Zhuoran Yang", "Vahid Tarokh"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-09", "url": "https://arxiv.org/abs/2602.08244", "pdf_url": "https://arxiv.org/pdf/2602.08244v1", "arxiv_id": "2602.08244", "doi": "10.48550/arXiv.2602.08244", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4698}
{"id": "578200056e911dd35d7dc86ec61613a7dc5e8f9d6b2bfc057f1304511c40431c", "sources": ["arxiv", "semantic_scholar"], "title": "Adversarial Reward Auditing for Active Detection and Mitigation of Reward Hacking", "abstract": "Reinforcement Learning from Human Feedback (RLHF) remains vulnerable to reward hacking, where models exploit spurious correlations in learned reward models to achieve high scores while violating human intent. Existing mitigations rely on static defenses that cannot adapt to novel exploitation strategies. We propose Adversarial Reward Auditing (ARA), a framework that reconceptualizes reward hacking as a dynamic, competitive game. ARA operates in two stages: first, a Hacker policy discovers reward model vulnerabilities while an Auditor learns to detect exploitation from latent representations; second, Auditor-Guided RLHF (AG-RLHF) gates reward signals to penalize detected hacking, transforming reward hacking from an unobservable failure into a measurable, controllable signal. Experiments across three hacking scenarios demonstrate that ARA achieves the best alignment-utility tradeoff among all baselines: reducing sycophancy to near-SFT levels while improving helpfulness, decreasing verbosity while achieving the highest ROUGE-L, and suppressing code gaming while improving Pass@1. Beyond single-domain evaluation, we show that reward hacking, detection, and mitigation all generalize across domains -- a Hacker trained on code gaming exhibits increased sycophancy despite no reward for this behavior, and an Auditor trained on one domain effectively suppresses exploitation in others, enabling efficient multi-domain defense with a single model.", "authors": ["Mohammad Beigi", "Ming Jin", "Junshan Zhang", "Qifan Wang", "Lifu Huang"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-02-02", "url": "https://arxiv.org/abs/2602.01750", "pdf_url": "https://arxiv.org/pdf/2602.01750v1", "arxiv_id": "2602.01750", "doi": "10.48550/arXiv.2602.01750", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4618}
{"id": "54e69b3bd365374b891861a6aae36a8a89cf165afbef9db558a17c9ac4c17840", "sources": ["arxiv", "semantic_scholar"], "title": "Omni-RRM: Advancing Omni Reward Modeling via Automatic Rubric-Grounded Preference Synthesis", "abstract": "Multimodal large language models (MLLMs) have shown remarkable capabilities, yet their performance is often capped by the coarse nature of existing alignment techniques. A critical bottleneck remains the lack of effective reward models (RMs): existing RMs are predominantly vision-centric, return opaque scalar scores, and rely on costly human annotations. We introduce \\textbf{Omni-RRM}, the first open-source rubric-grounded reward model that produces structured, multi-dimension preference judgments with dimension-wise justifications across \\textbf{text, image, video, and audio}. At the core of our approach is \\textbf{Omni-Preference}, a large-scale dataset built via a fully automated pipeline: we synthesize candidate response pairs by contrasting models of different capabilities, and use strong teacher models to \\emph{reconcile and filter} preferences while providing a modality-aware \\emph{rubric-grounded rationale} for each pair. This eliminates the need for human-labeled training preferences. Omni-RRM is trained in two stages: supervised fine-tuning to learn the rubric-grounded outputs, followed by reinforcement learning (GRPO) to sharpen discrimination on difficult, low-contrast pairs. Comprehensive evaluations show that Omni-RRM achieves state-of-the-art accuracy on video (80.2\\% on ShareGPT-V) and audio (66.8\\% on Audio-HH-RLHF) benchmarks, and substantially outperforms existing open-source RMs on image tasks, with a 17.7\\% absolute gain over its base model on overall accuracy. Omni-RRM also improves downstream performance via Best-of-$N$ selection and transfers to text-only preference benchmarks. Our data, code, and models are available at https://anonymous.4open.science/r/Omni-RRM-CC08.", "authors": ["Zicheng Kong", "Dehua Ma", "Zhenbo Xu", "Alven Yang", "Yiwei Ru", "Haoran Wang", "Zixuan Zhou", "Fuqing Bie", "Liuyu Xiang", "Huijia Wu", "Jian Zhao", "Zhaofeng He"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-31", "url": "https://arxiv.org/abs/2602.00846", "pdf_url": "https://arxiv.org/pdf/2602.00846v1", "arxiv_id": "2602.00846", "doi": "10.48550/arXiv.2602.00846", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7101}
{"id": "8dd40a8027df67e156a73e8a83985867627de1e594d465f241e7d247246fe38f", "sources": ["arxiv", "semantic_scholar"], "title": "Factored Causal Representation Learning for Robust Reward Modeling in RLHF", "abstract": "A reliable reward model is essential for aligning large language models with human preferences through reinforcement learning from human feedback. However, standard reward models are susceptible to spurious features that are not causally related to human labels. This can lead to reward hacking, where high predicted reward does not translate into better behavior. In this work, we address this problem from a causal perspective by proposing a factored representation learning framework that decomposes the model's contextual embedding into (1) causal factors that are sufficient for reward prediction and (2) non-causal factors that capture reward-irrelevant attributes such as length or sycophantic bias. The reward head is then constrained to depend only on the causal component. In addition, we introduce an adversarial head trained to predict reward from the non-causal factors, while applying gradient reversal to discourage them from encoding reward-relevant information. Experiments on both mathematical and dialogue tasks demonstrate that our method learns more robust reward models and consistently improves downstream RLHF performance over state-of-the-art baselines. Analyses on length and sycophantic bias further validate the effectiveness of our method in mitigating reward hacking behaviors.", "authors": ["Yupei Yang", "Lin Yang", "Wanxi Deng", "Lin Qu", "Fan Feng", "Biwei Huang", "Shikui Tu", "Lei Xu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.21350", "pdf_url": "https://arxiv.org/pdf/2601.21350v2", "arxiv_id": "2601.21350", "doi": "10.48550/arXiv.2601.21350", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4572}
{"id": "64de0b1b3f061e1b15a4c4de727cdecdc3a3bb115fa5043446a6e2114765ab6d", "sources": ["arxiv", "semantic_scholar"], "title": "FunPRM: Function-as-Step Process Reward Model with Meta Reward Correction for Code Generation", "abstract": "Code generation is a core application of large language models (LLMs), yet LLMs still frequently fail on complex programming tasks. Given its success in mathematical reasoning, test-time scaling approaches such as Process Reward Model (PRM)-based Best-of-N selection offer a promising way to improve performance. However, existing PRMs remain ineffective for code generation due to the lack of meaningful step decomposition in code and the noise of Monte Carlo-estimated partial-solution correctness scores (rewards). To address these challenges, we propose FunPRM. FunPRM prompts LLMs to encourage modular code generation organized into functions, with functions treated as PRM reasoning steps. Furthermore, FunPRM introduces a novel meta-learning-based reward correction mechanism that leverages clean final-solution rewards obtained via a unit-test-based evaluation system to purify noisy partial-solution rewards. Experiments on LiveCodeBench and BigCodeBench demonstrate that FunPRM consistently outperforms existing test-time scaling methods across five base LLMs, notably achieving state-of-the-art performance on LiveCodeBench when combined with O4-mini. Furthermore, FunPRM produces code that is more readable and reusable for developers.", "authors": ["Ruiyi Zhang", "Peijia Qin", "Qi Cao", "Eric Xue", "Pengtao Xie"], "categories": ["cs.LG", "cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-29", "url": "https://arxiv.org/abs/2601.22249", "pdf_url": "https://arxiv.org/pdf/2601.22249v1", "arxiv_id": "2601.22249", "doi": "10.48550/arXiv.2601.22249", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4572}
{"id": "86b88a864495f1cf6c00ef5d5912c6063e89e3e4ba743ce24e0af6b64c2b66a1", "sources": ["arxiv", "semantic_scholar"], "title": "The Trajectory Alignment Coefficient in Two Acts: From Reward Tuning to Reward Learning", "abstract": "The success of reinforcement learning (RL) is fundamentally tied to having a reward function that accurately reflects the task objective. Yet, designing reward functions is notoriously time-consuming and prone to misspecification. To address this issue, our first goal is to understand how to support RL practitioners in specifying appropriate weights for a reward function. We leverage the Trajectory Alignment Coefficient (TAC), a metric that evaluates how closely a reward function's induced preferences match those of a domain expert. To evaluate whether TAC provides effective support in practice, we conducted a human-subject study in which RL practitioners tuned reward weights for Lunar Lander. We found that providing TAC during reward tuning led participants to produce more performant reward functions and report lower cognitive workload relative to standard tuning without TAC. However, the study also underscored that manual reward design, even with TAC, remains labor-intensive. This limitation motivated our second goal: to learn a reward model that maximizes TAC directly. Specifically, we propose Soft-TAC, a differentiable approximation of TAC that can be used as a loss function to train reward models from human preference data. Validated in the racing simulator Gran Turismo 7, reward models trained using Soft-TAC successfully captured preference-specific objectives, resulting in policies with qualitatively more distinct behaviors than models trained with standard Cross-Entropy loss. This work demonstrates that TAC can serve as both a practical tool for guiding reward tuning and a reward learning objective in complex domains.", "authors": ["Calarina Muslimani", "Yunshu Du", "Kenta Kawamoto", "Kaushik Subramanian", "Peter Stone", "Peter Wurman"], "categories": ["cs.LG", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-23", "url": "https://arxiv.org/abs/2601.16906", "pdf_url": "https://arxiv.org/pdf/2601.16906v1", "arxiv_id": "2601.16906", "doi": "10.48550/arXiv.2601.16906", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4503}
{"id": "177fc0edf41a8ecd68b1a5bf29e9d64dc2a7ef5f9d1d97bbc4279cb23880d5a8", "sources": ["arxiv", "semantic_scholar"], "title": "Rewarding Creativity: A Human-Aligned Generative Reward Model for Reinforcement Learning in Storytelling", "abstract": "While Large Language Models (LLMs) can generate fluent text, producing high-quality creative stories remains challenging. Reinforcement Learning (RL) offers a promising solution but faces two critical obstacles: designing reliable reward signals for subjective storytelling quality and mitigating training instability. This paper introduces the Reinforcement Learning for Creative Storytelling (RLCS) framework to systematically address both challenges. First, we develop a Generative Reward Model (GenRM) that provides multi-dimensional analysis and explicit reasoning about story preferences, trained through supervised fine-tuning on demonstrations with reasoning chains distilled from strong teacher models, followed by GRPO-based refinement on expanded preference data. Second, we introduce an entropy-based reward shaping strategy that dynamically prioritizes learning on confident errors and uncertain correct predictions, preventing overfitting on already-mastered patterns. Experiments demonstrate that GenRM achieves 68\\% alignment with human creativity judgments, and RLCS significantly outperforms strong baselines including Gemini-2.5-Pro in overall story quality. This work provides a practical pipeline for applying RL to creative domains, effectively navigating the dual challenges of reward modeling and training stability.", "authors": ["Zhaoyan Li", "Hang Lei", "Yujia Wang", "Lanbo Liu", "Hao Liu", "Liang Yu"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-12", "url": "https://arxiv.org/abs/2601.07149", "pdf_url": "https://arxiv.org/pdf/2601.07149v1", "arxiv_id": "2601.07149", "doi": "10.48550/arXiv.2601.07149", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4377}
{"id": "50619689f08b499da3b28fd371878d0a3c7a5cd8dfe8e0e16bf7aa0333a457a1", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Modeling from Natural Language Human Feedback", "abstract": "Reinforcement Learning with Verifiable reward (RLVR) on preference data has become the mainstream approach for training Generative Reward Models (GRMs). Typically in pairwise rewarding tasks, GRMs generate reasoning chains ending with critiques and preference labels, and RLVR then relies on the correctness of the preference labels as the training reward. However, in this paper, we demonstrate that such binary classification tasks make GRMs susceptible to guessing correct outcomes without sound critiques. Consequently, these spurious successes introduce substantial noise into the reward signal, thereby impairing the effectiveness of reinforcement learning. To address this issue, we propose Reward Modeling from Natural Language Human Feedback (RM-NLHF), which leverages natural language feedback to obtain process reward signals, thereby mitigating the problem of limited solution space inherent in binary tasks. Specifically, we compute the similarity between GRM-generated and human critiques as the training reward, which provides more accurate reward signals than outcome-only supervision. Additionally, considering that human critiques are difficult to scale up, we introduce Meta Reward Model (MetaRM) which learns to predict process reward from datasets with human critiques and then generalizes to data without human critiques. Experiments on multiple benchmarks demonstrate that our method consistently outperforms state-of-the-art GRMs trained with outcome-only reward, confirming the superiority of integrating natural language over binary human feedback as supervision.", "authors": ["Zongqi Wang", "Rui Wang", "Yuchuan Wu", "Yiyao Yu", "Pinyi Zhang", "Shaoning Sun", "Yujiu Yang", "Yongbin Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-12", "url": "https://arxiv.org/abs/2601.07349", "pdf_url": "https://arxiv.org/pdf/2601.07349v3", "arxiv_id": "2601.07349", "doi": "10.48550/arXiv.2601.07349", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4377}
{"id": "1290db498e9302ce8c7d5c4b2740fac1b2288f4e52e1aff0eab040d0138b4c74", "sources": ["arxiv", "semantic_scholar"], "title": "IRPM: Intergroup Relative Preference Modeling for Pointwise Generative Reward Models", "abstract": "Generative Reward Models (GRMs) have demonstrated strong performance in reward modeling, due to their interpretability and potential for refinement through reinforcement learning (RL). However, widely used pairwise GRMs create a computational bottleneck in reinforcement learning from human feedback (RLHF), when calibrating or aggregating preference signals over n candidates, often incurring O(n^2) pairwise judgments. To address this issue, we propose Intergroup Relative Preference Modeling (IRPM), an RL-based method that extends the Bradley--Terry preference-learning paradigm via intergroup comparisons to train pointwise GRMs from pairwise preference data. IRPM derives pointwise reward for each response by contrasting groups of chosen vs. rejected samples, enabling pointwise scores comparable across candidate sets and O(n) reward evaluation for a variable number of candidates during RL training, while preserving interpretability and scalability. Experiments show that IRPM achieves state-of-the-art performance among pointwise GRMs on RM-Bench, JudgeBench and RewardBench, and approaches the performance of leading pairwise GRMs. In addition, IRPM achieves substantial gains in post-training evaluations, demonstrating its effectiveness.", "authors": ["Haonan Song", "Qingchen Xie", "Huan Zhu", "Feng Xiao", "Luxi Xing", "Liu Kang", "Fuzhen Li", "Zhiyong Zheng", "Feng Jiang", "Ziheng Li", "Kun Yan", "Qingyi Si", "Yanghua Xiao", "Hongcheng Guo", "Fan Yang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2026-01-02", "url": "https://arxiv.org/abs/2601.00677", "pdf_url": "https://arxiv.org/pdf/2601.00677v2", "arxiv_id": "2601.00677", "doi": "10.48550/arXiv.2601.00677", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4263}
{"id": "9bfa7f674c61a2e0075bfd4810ee32eaf0f8e44d20ffd9eef23b94a23fa5da06", "sources": ["arxiv", "semantic_scholar"], "title": "Eliminating Inductive Bias in Reward Models with Information-Theoretic Guidance", "abstract": "Reward models (RMs) are essential in reinforcement learning from human feedback (RLHF) to align large language models (LLMs) with human values. However, RM training data is commonly recognized as low-quality, containing inductive biases that can easily lead to overfitting and reward hacking. For example, more detailed and comprehensive responses are usually human-preferred but with more words, leading response length to become one of the inevitable inductive biases. A limited number of prior RM debiasing approaches either target a single specific type of bias or model the problem with only simple linear correlations, \\textit{e.g.}, Pearson coefficients. To mitigate more complex and diverse inductive biases in reward modeling, we introduce a novel information-theoretic debiasing method called \\textbf{D}ebiasing via \\textbf{I}nformation optimization for \\textbf{R}M (DIR). Inspired by the information bottleneck (IB), we maximize the mutual information (MI) between RM scores and human preference pairs, while minimizing the MI between RM outputs and biased attributes of preference inputs. With theoretical justification from information theory, DIR can handle more sophisticated types of biases with non-linear correlations, broadly extending the real-world application scenarios for RM debiasing methods. In experiments, we verify the effectiveness of DIR with three types of inductive biases: \\textit{response length}, \\textit{sycophancy}, and \\textit{format}. We discover that DIR not only effectively mitigates target inductive biases but also enhances RLHF performance across diverse benchmarks, yielding better generalization abilities. The code and training recipes are available at https://github.com/Qwen-Applications/DIR.", "authors": ["Zhuo Li", "Pengyu Cheng", "Zhechao Yu", "Feifei Tong", "Anningzhe Gao", "Tsung-Hui Chang", "Xiang Wan", "Erchao Zhao", "Xiaoxi Jiang", "Guanjun Jiang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-29", "url": "https://arxiv.org/abs/2512.23461", "pdf_url": "https://arxiv.org/pdf/2512.23461v2", "arxiv_id": "2512.23461", "doi": "10.48550/arXiv.2512.23461", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Qwen-Applications/DIR", "venue": "arXiv.org", "quality_score": 0.6517}
{"id": "a78990a4fcec3e83dd61b2e41c4845911b97d230532b8bcdd541eb4807047f8e", "sources": ["arxiv", "semantic_scholar"], "title": "Revisiting the Learning Objectives of Vision-Language Reward Models", "abstract": "Learning generalizable reward functions is a core challenge in embodied intelligence. Recent work leverages contrastive vision language models (VLMs) to obtain dense, domain-agnostic rewards without human supervision. These methods adapt VLMs into reward models through increasingly complex learning objectives, yet meaningful comparison remains difficult due to differences in training data, architectures, and evaluation settings. In this work, we isolate the impact of the learning objective by evaluating recent VLM-based reward models under a unified framework with identical backbones, finetuning data, and evaluation environments. Using Meta-World tasks, we assess modeling accuracy by measuring consistency with ground truth reward and correlation with expert progress. Remarkably, we show that a simple triplet loss outperforms state-of-the-art methods, suggesting that much of the improvements in recent approaches could be attributed to differences in data and architectures.", "authors": ["Simon Roy", "Samuel Barbeau", "Giovanni Beltrame", "Christian Desrosiers", "Nicolas Thome"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-20", "url": "https://arxiv.org/abs/2512.20675", "pdf_url": "https://arxiv.org/pdf/2512.20675v1", "arxiv_id": "2512.20675", "doi": "10.48550/arXiv.2512.20675", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4114}
{"id": "6e8eb89a2dd7b852bd82376eddf7ddd1e7652a05dc26b991a1d46d100d8301bb", "sources": ["arxiv", "semantic_scholar"], "title": "A First-Order Logic-Based Alternative to Reward Models in RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) plays a crucial role in aligning large language models (LLMs) with human values and preferences. However, the quality and stability of the trained reward model largely determine the final alignment performance. Existing approaches such as Proximal Policy Optimization (PPO) rely heavily on reward models to guide LLMs toward human-aligned behaviors. In this work, we propose a logic-similarity-based reward mechanism as an alternative to conventional reward modeling. Instead of relying on heuristic reward estimation, our method leverages formal logical consistency to steer model alignment with human preferences. Since real-world questions can be interpreted from multiple perspectives, to ensure that logic-based reinforcement learning does not cause model collapse, we introduce S-GRPO, a supervised variant of the GRPO framework. S-GRPO incorporates an additional supervised component and jointly optimizes the generation term, KL-divergence regularization, and label-based objective during training. Experimental results demonstrate that S-GRPO consistently outperforms standard supervised fine-tuning (SFT) in both performance and robustness. Furthermore, it extends existing preference-learning frameworks such as GRPO and DPO, offering a more flexible and task-adaptive approach to alignment training. Our code is available at https://github.com/ChunjinJiang/sgrpo.", "authors": ["Chunjin Jian", "Xinhua Zhu"], "categories": ["cs.LG", "cs.LO"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-16", "url": "https://arxiv.org/abs/2512.14100", "pdf_url": "https://arxiv.org/pdf/2512.14100v1", "arxiv_id": "2512.14100", "doi": "10.1109/ICNLP69856.2026.11527781", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ChunjinJiang/sgrpo", "venue": "ICON", "quality_score": 0.6286}
{"id": "c9762f1bf7f7c666904cea7812e1215ef6a2b6c83da8894b728a244f92448324", "sources": ["arxiv", "semantic_scholar"], "title": "Model-Based Reinforcement Learning in Discrete-Action Non-Markovian Reward Decision Processes", "abstract": "Many practical decision-making problems involve tasks whose success depends on the entire system history, rather than on achieving a state with desired properties. Markovian Reinforcement Learning (RL) approaches are not suitable for such tasks, while RL with non-Markovian reward decision processes (NMRDPs) enables agents to tackle temporal-dependency tasks. This approach has long been known to lack formal guarantees on both (near-)optimality and sample efficiency. We contribute to solving both issues with QR-MAX, a novel model-based algorithm for discrete NMRDPs that factorizes Markovian transition learning from non-Markovian reward handling via reward machines. To the best of our knowledge, this is the first model-based RL algorithm for discrete-action NMRDPs that exploits this factorization to obtain PAC convergence to $\\varepsilon$-optimal policies with polynomial sample complexity. We then extend QR-MAX to continuous state spaces with Bucket-QR-MAX, a SimHash-based discretiser that preserves the same factorized structure and achieves fast and stable learning without manual gridding or function approximation. We experimentally compare our method with modern state-of-the-art model-based RL approaches on environments of increasing complexity, showing a significant improvement in sample efficiency and increased robustness in finding optimal policies.", "authors": ["Alessandro Trapasso", "Luca Iocchi", "Fabio Patrizi"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-16", "url": "https://arxiv.org/abs/2512.14617", "pdf_url": "https://arxiv.org/pdf/2512.14617v2", "arxiv_id": "2512.14617", "doi": "10.48550/arXiv.2512.14617", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4068}
{"id": "77f31f9101a90c6148d76fbc78f0bf80dbe539c87f8f7181ac1edab5ebf51e20", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Objective Reward and Preference Optimization: Theory and Algorithms", "abstract": "This thesis develops theoretical frameworks and algorithms that advance constrained reinforcement learning (RL) across control, preference learning, and alignment of large language models. The first contribution addresses constrained Markov Decision Processes (CMDPs) under the average-cost criterion through the Average-Constrained Policy Optimization (ACPO) algorithm. ACPO integrates sensitivity analysis with trust-region updates to ensure stable constraint handling, achieving state-of-the-art empirical performance with theoretical guarantees. Constrained RL is then extended to finite-horizon settings via e-COP, the first policy optimization method for episodic CMDPs. Built on an episodic policy difference lemma, e-COP offers provable performance, simplicity, and scalability in safety-critical environments. The thesis then investigates reinforcement learning from human preferences. warmPref-PS introduces a posterior sampling strategy for linear bandits that integrates offline preference data from heterogeneous raters into online learning. Explicit modeling of rater competence yields substantial regret reduction and more efficient data collection for RLHF. The PSPL algorithm further advances preference-based RL by jointly sampling reward models and transition dynamics from pairwise trajectory comparisons, providing Bayesian simple-regret guarantees and robust empirical identification of optimal policies. The final contribution applies these methods to large-scale model alignment. A multi-objective constrained optimization view yields MOPO, an iterative algorithm with closed-form updates that scales to multi-billion-parameter language models and remains robust across alignment settings. Collectively, the thesis unifies constrained RL across average-cost, episodic, and preference-driven paradigms, delivering theoretical advances and practical tools for safe and aligned decision-making.", "authors": ["Akhil Agnihotri"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-11", "url": "https://arxiv.org/abs/2512.10601", "pdf_url": "https://arxiv.org/pdf/2512.10601v1", "arxiv_id": "2512.10601", "doi": "10.48550/arXiv.2512.10601", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.401}
{"id": "f7b39439ec680bd3ea44113d2831368bb08976abd3ecdd2cc2c38d5f43ff43fd", "sources": ["arxiv", "semantic_scholar"], "title": "Parent-Guided Semantic Reward Model (PGSRM): Embedding-Based Reward Functions for Reinforcement Learning of Transformer Language Models", "abstract": "We introduce the Parent-Guided Semantic Reward Model (PGSRM), a lightweight reward framework for reinforcement learning (RL) of transformer language models. PGSRM replaces binary correctness signals, human preference data, and trained reward models with a simple signal: cosine similarity between a parent model's reference output embedding and a child model's generated output for the same input. This yields a dense, semantically meaningful reward with no human annotation or additional model training. We apply PGSRM on five language tasks and find that it produces smoother reward improvement and more stable PPO dynamics than a binary reward baseline, suggesting that embedding-based semantic rewards are a practical alternative to RLHF-style reward modeling for parent-guided alignment in smaller transformer models.", "authors": ["Alexandr Plashchinsky"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-07", "url": "https://arxiv.org/abs/2512.06920", "pdf_url": "https://arxiv.org/pdf/2512.06920v1", "arxiv_id": "2512.06920", "doi": "10.48550/arXiv.2512.06920", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3965}
{"id": "6de827cd2b0b9894f239791ed7558a923b8fe3d6aa2ac586df7df0df2f064c7f", "sources": ["arxiv", "semantic_scholar"], "title": "Average-reward reinforcement learning in semi-Markov decision processes via relative value iteration", "abstract": "This paper applies the authors' recent results on asynchronous stochastic approximation (SA) in the Borkar-Meyn framework to reinforcement learning in average-reward semi-Markov decision processes (SMDPs). We establish the convergence of an asynchronous SA analogue of Schweitzer's classical relative value iteration algorithm, RVI Q-learning, for finite-space, weakly communicating SMDPs. In particular, we show that the algorithm converges almost surely to a compact, connected subset of solutions to the average-reward optimality equation, with convergence to a unique, sample path-dependent solution under additional stepsize and asynchrony conditions. Moreover, to make full use of the SA framework, we introduce new monotonicity conditions for estimating the optimal reward rate in RVI Q-learning. These conditions substantially expand the previously considered algorithmic framework and are addressed through novel arguments in the stability and convergence analysis of RVI Q-learning.", "authors": ["Huizhen Yu", "Yi Wan", "Richard S. Sutton"], "categories": ["cs.LG", "math.OC"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-12-05", "url": "https://arxiv.org/abs/2512.06218", "pdf_url": "https://arxiv.org/pdf/2512.06218v1", "arxiv_id": "2512.06218", "doi": "10.48550/arXiv.2512.06218", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3942}
{"id": "24b58c066e5553cd845f8261e95c6ba99a36a3f26d2111eaae2a1221b02b36e1", "sources": ["arxiv", "semantic_scholar"], "title": "Uncertainty Quantification for Large Language Model Reward Learning under Heterogeneous Human Feedback", "abstract": "We study estimation and statistical inference for reward models used in aligning large language models (LLMs). A key component of LLM alignment is reinforcement learning from human feedback (RLHF), where humans compare pairs of model-generated answers and their preferences are used to train a reward model. However, human feedback is inherently heterogeneous, creating significant challenges for reliable reward learning. To address this, we adopt a heterogeneous preference framework that jointly models the latent reward of answers and human rationality. This leads to a challenging biconvex optimization problem, which we solve via an alternating gradient descent algorithm. We establish theoretical guarantees for the resulting estimator, including its convergence and asymptotic distribution. These results enable the construction of confidence intervals for reward estimates. Leveraging these uncertainty quantification results, we conduct valid statistical comparisons between rewards and incorporate uncertainty into the best-of-$N$ (BoN) policy framework. Extensive simulations demonstrate the effectiveness of our method, and applications to real LLM data highlight the practical value of accounting for uncertainty in reward modeling for LLM alignment.", "authors": ["Pangpang Liu", "Junwei Lu", "Will Wei Sun"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-12-02", "url": "https://arxiv.org/abs/2512.03208", "pdf_url": "https://arxiv.org/pdf/2512.03208v1", "arxiv_id": "2512.03208", "doi": "10.48550/arXiv.2512.03208", "citation_count": 7, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3907}
{"id": "4da791dd8bbc581d9144c98519d29b224f412a983f2ad2b7325fd859fd851971", "sources": ["arxiv", "semantic_scholar"], "title": "SPARK: Stepwise Process-Aware Rewards for Reference-Free Reinforcement Learning", "abstract": "Process reward models (PRMs) that provide dense, step-level feedback have shown promise for reinforcement learning, yet their adoption remains limited by the need for expensive step-level annotations or ground truth references. We propose SPARK: a three-stage framework where in the first stage a generator model produces diverse solutions and a verifier model evaluates them using parallel scaling (self-consistency) and sequential scaling (meta-critique). In the second stage, we use these verification outputs as synthetic training data to fine-tune generative process reward models, which subsequently serve as reward signals during training. We show that aggregating multiple independent verifications at the step level produces training data for process reward models that surpass ground-truth outcome supervision, achieving 67.5 F1 on ProcessBench (a benchmark for identifying erroneous steps in mathematical reasoning) compared to 66.4 for reference-guided training and 61.9 for GPT-4o. In the final stage, we apply our generative PRM with chain-of-thought verification (PRM-CoT) as the reward model in RL experiments on mathematical reasoning, and introduce format constraints to prevent reward hacking. Using Qwen2.5-Math-7B, we achieve 47.4% average accuracy across six mathematical reasoning benchmarks, outperforming ground-truth-based RLVR (43.9%). Our work enables reference-free RL training that exceeds ground-truth methods, opening new possibilities for domains lacking verifiable answers or accessible ground truth.", "authors": ["Salman Rahman", "Sruthi Gorantla", "Arpit Gupta", "Swastik Roy", "Nanyun Peng", "Yang Liu"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-12-02", "url": "https://arxiv.org/abs/2512.03244", "pdf_url": "https://arxiv.org/pdf/2512.03244v1", "arxiv_id": "2512.03244", "doi": "10.48550/arXiv.2512.03244", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3907}
{"id": "c476fad52f45bc2b5654fd1cbae91a0178ee36102baab796bd6572298b8de484", "sources": ["arxiv", "semantic_scholar"], "title": "UMM-RM: An Upcycle-and-Merge MoE Reward Model for Mitigating Reward Hacking", "abstract": "Reward models (RMs) are a critical component of reinforcement learning from human feedback (RLHF). However, conventional dense RMs are susceptible to exploitation by policy models through biases or spurious correlations, resulting in reward hacking: RM scores increase during training while alignment with human preferences deteriorates, a problem that is further exacerbated under distribution shift.To address this issue, we propose UMM-RM (Upcycle-and-Merge MoE Reward Model). UMM-RM first upscales the feed-forward layers of a dense backbone into a mixture-of-experts (MoE) reward model with shared experts. The shared experts are always activated to capture instruction-agnostic preference signals, while the remaining experts model fine-grained preferences across instructions or task regimes. After training, the experts are consolidated into a single dense RM via learnable merging weights.This design retains the robustness and exploitation resistance provided by expert diversity while avoiding the inference overhead of MoE architectures or explicit ensembles. Experiments across multiple base models and preference datasets show that, compared with standard dense RMs, UMM-RM improves accuracy on preference data, reduces reward hacking during PPO training, and yields more stable preference alignment.", "authors": ["Lingling Fu", "Yongfu Xue"], "categories": ["cs.LG", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-30", "url": "https://arxiv.org/abs/2512.00724", "pdf_url": "https://arxiv.org/pdf/2512.00724v2", "arxiv_id": "2512.00724", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2472}
{"id": "7c9dea143b8b749e2c4db110e4996ad16cbad1dfe607511c212853bb407d4c37", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Engineering for Spatial Epidemic Simulations: A Reinforcement Learning Platform for Individual Behavioral Learning", "abstract": "We present ContagionRL, a Gymnasium-compatible reinforcement learning platform specifically designed for systematic reward engineering in spatial epidemic simulations. Unlike traditional agent-based models that rely on fixed behavioral rules, our platform enables rigorous evaluation of how reward function design affects learned survival strategies across diverse epidemic scenarios. ContagionRL integrates a spatial SIRS+D epidemiological model with configurable environmental parameters, allowing researchers to stress-test reward functions under varying conditions including limited observability, different movement patterns, and heterogeneous population dynamics. We evaluate five distinct reward designs, ranging from sparse survival bonuses to a novel potential field approach, across multiple RL algorithms (PPO, SAC, A2C). Through systematic ablation studies, we identify that directional guidance and explicit adherence incentives are critical components for robust policy learning. Our comprehensive evaluation across varying infection rates, grid sizes, visibility constraints, and movement patterns reveals that reward function choice dramatically impacts agent behavior and survival outcomes. Agents trained with our potential field reward consistently achieve superior performance, learning maximal adherence to non-pharmaceutical interventions while developing sophisticated spatial avoidance strategies. The platform's modular design enables systematic exploration of reward-behavior relationships, addressing a knowledge gap in models of this type where reward engineering has received limited attention. ContagionRL is an effective platform for studying adaptive behavioral responses in epidemic contexts and highlight the importance of reward design, information structure, and environmental predictability in learning. Our code is publicly available at https://github.com/redradman/ContagionRL", "authors": ["Radman Rakhshandehroo", "Daniel Coombs"], "categories": ["cs.LG", "cs.AI", "q-bio.PE"], "fields_of_study": ["Computer Science", "Biology"], "published_date": "2025-11-22", "url": "https://arxiv.org/abs/2511.18000", "pdf_url": "https://arxiv.org/pdf/2511.18000v2", "arxiv_id": "2511.18000", "doi": "10.48550/arXiv.2511.18000", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/redradman/ContagionRL", "venue": "Transactions on Machine Learning Research, 2026", "quality_score": 0.5861}
{"id": "0bee6d3826c3610569d98ba75a51489bd84e27e9fbf5cf4ba64475f99818c312", "sources": ["arxiv", "semantic_scholar"], "title": "The Good, The Bad, and The Hybrid: A Reward Structure Showdown in Reasoning Models Training", "abstract": "Reward design is central to reinforcement learning from human feedback (RLHF) and alignment research. In this work, we propose a unified framework to study hard, continuous, and hybrid reward structures for fine-tuning large language models (LLMs) on mathematical reasoning tasks. Using Qwen3-4B with LoRA fine-tuning on the GSM8K dataset, we formalize and empirically evaluate reward formulations that incorporate correctness, perplexity, reasoning quality, and consistency. We introduce an adaptive hybrid reward scheduler that transitions between discrete and continuous signals, balancing exploration and stability. Our results show that hybrid reward structures improve convergence speed and training stability over purely hard or continuous approaches, offering insights for alignment via adaptive reward modeling.", "authors": ["Subramanyam Sahoo"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-17", "url": "https://arxiv.org/abs/2511.13016", "pdf_url": "https://arxiv.org/pdf/2511.13016v1", "arxiv_id": "2511.13016", "doi": "10.48550/arXiv.2511.13016", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3735}
{"id": "17fb9d747f0085583fe1ed2fa5b050460c94d272e89c861702b1164e991bab9b", "sources": ["arxiv", "semantic_scholar"], "title": "Probing Preference Representations: A Multi-Dimensional Evaluation and Analysis Method for Reward Models", "abstract": "Previous methods evaluate reward models by testing them on a fixed pairwise ranking test set, but they typically do not provide performance information on each preference dimension. In this work, we address the evaluation challenge of reward models by probing preference representations. To confirm the effectiveness of this evaluation method, we construct a Multi-dimensional Reward Model Benchmark (MRMBench), a collection of six probing tasks for different preference dimensions. We design it to favor and encourage reward models that better capture preferences across different dimensions. Furthermore, we introduce an analysis method, inference-time probing, which identifies the dimensions used during the reward prediction and enhances its interpretability. Through extensive experiments, we find that MRMBench strongly correlates with the alignment performance of large language models (LLMs), making it a reliable reference for developing advanced reward models. Our analysis of MRMBench evaluation results reveals that reward models often struggle to capture preferences across multiple dimensions, highlighting the potential of multi-objective optimization in reward modeling. Additionally, our findings show that the proposed inference-time probing method offers a reliable metric for assessing the confidence of reward predictions, which ultimately improves the alignment of LLMs.", "authors": ["Chenglong Wang", "Yifu Huo", "Yang Gan", "Yongyu Mu", "Qiaozhi He", "Murun Yang", "Bei Li", "Chunliang Zhang", "Tongran Liu", "Anxiang Ma", "Zhengtao Yu", "Jingbo Zhu", "Tong Xiao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-16", "url": "https://arxiv.org/abs/2511.12464", "pdf_url": "https://arxiv.org/pdf/2511.12464v1", "arxiv_id": "2511.12464", "doi": "10.48550/arXiv.2511.12464", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3724}
{"id": "af7324416da21e6c5342ffa6ca1b534996e046adee852042e4ec580a1fbc8374", "sources": ["arxiv", "semantic_scholar"], "title": "PROF: An LLM-based Reward Code Preference Optimization Framework for Offline Imitation Learning", "abstract": "Offline imitation learning (offline IL) enables training effective policies without requiring explicit reward annotations. Recent approaches attempt to estimate rewards for unlabeled datasets using a small set of expert demonstrations. However, these methods often assume that the similarity between a trajectory and an expert demonstration is positively correlated with the reward, which oversimplifies the underlying reward structure. We propose PROF, a novel framework that leverages large language models (LLMs) to generate and improve executable reward function codes from natural language descriptions and a single expert trajectory. We propose Reward Preference Ranking (RPR), a novel reward function quality assessment and ranking strategy without requiring environment interactions or RL training. RPR calculates the dominance scores of the reward functions, where higher scores indicate better alignment with expert preferences. By alternating between RPR and text-based gradient optimization, PROF fully automates the selection and refinement of optimal reward functions for downstream policy learning. Empirical results on D4RL demonstrate that PROF surpasses or matches recent strong baselines across numerous datasets and domains, highlighting the effectiveness of our approach.", "authors": ["Shengjie Sun", "Jiafei Lyu", "Runze Liu", "Mengbei Yan", "Bo Liu", "Deheng Ye", "Xiu Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-14", "url": "https://arxiv.org/abs/2511.13765", "pdf_url": "https://arxiv.org/pdf/2511.13765v1", "arxiv_id": "2511.13765", "doi": "10.48550/arXiv.2511.13765", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3701}
{"id": "561c5dfc9d4c1ea42581f6dc60d07341cc6a5dca6f5f6ada4ad3fab581832120", "sources": ["arxiv", "semantic_scholar"], "title": "PIRA: Preference-Oriented Instruction-Tuned Reward Models with Dual Aggregation", "abstract": "Reward models are pivotal for aligning Large Language Models (LLMs) with human preferences. Existing approaches face two key limitations: Discriminative reward models require large-scale annotated data, as they cannot exploit the preference instruction-following capability of LLMs available to generative reward models. Moreover, reward models are particularly prone to reward overoptimization, where LLMs exploit weaknesses in the reward function instead of improving true alignment. We introduce \\textbf{PIRA}, a training paradigm that integrates three complementary strategies to address these challenges: (1) reformulating question-answer pairs into preference-task instructions to explicitly leverage LLMs' preference instruction-following capability, (2) averaging the rewards aggregated from diverse preference-task instructions for each sample, which mitigates task-specific bias and enhances robustness across evaluation perspectives, and (3) averaging outputs from the value head under different dropout rates to stabilize reward estimation. Experiments on public datasets show that PIRA improves performance considerably, enhances generalization, and effectively mitigates reward overoptimization.", "authors": ["Yongfu Xue"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-11-14", "url": "https://arxiv.org/abs/2511.20668", "pdf_url": "https://arxiv.org/pdf/2511.20668v2", "arxiv_id": "2511.20668", "doi": "10.48550/arXiv.2511.20668", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference of the European Chapter of the Association for Computational Linguistics", "quality_score": 0.3701}
{"id": "8f189317b9004a615962e67e509af7b4cde256222f8c864355e37070f2533d8d", "sources": ["arxiv", "semantic_scholar"], "title": "Debiasing Reward Models by Representation Learning with Guarantees", "abstract": "Recent alignment techniques, such as reinforcement learning from human feedback, have been widely adopted to align large language models with human preferences by learning and leveraging reward models. In practice, these models often exploit spurious correlations, involving, e.g., response length, discrimination, sycophancy, and conceptual bias, which is a problem that has received increasing attention. In this work, we propose a principled framework that mitigates these biases in reward models while preserving the underlying factors that reflect intended preferences. We first provide a formulation of the data-generating process, assuming that the observed data (e.g., text) is generated from both spurious and non-spurious latent variables. We show that, interestingly, these non-spurious latent variables can be theoretically identified from data, regardless of whether a surrogate for the spurious latent variables is available. This further inspires a practical method that uses variational inference to recover these variables and leverages them to train reward models. Experiments on synthetic and real-world datasets demonstrate that our method effectively mitigates spurious correlation issues and yields more robust reward models.", "authors": ["Ignavier Ng", "Patrick Blöbaum", "Siddharth Bhandari", "Kun Zhang", "Shiva Kasiviswanathan"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-10-27", "url": "https://arxiv.org/abs/2510.23751", "pdf_url": "https://arxiv.org/pdf/2510.23751v1", "arxiv_id": "2510.23751", "doi": "10.48550/arXiv.2510.23751", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "4c9fe65e0624d37e8f46f3e953777ae447e302a5e1206d86df0e5a1eb4f51104", "sources": ["arxiv", "semantic_scholar"], "title": "Rectifying Shortcut Behaviors in Preference-based Reward Learning", "abstract": "In reinforcement learning from human feedback, preference-based reward models play a central role in aligning large language models to human-aligned behavior. However, recent studies show that these models are prone to reward hacking and often fail to generalize well due to over-optimization. They achieve high reward scores by exploiting shortcuts, that is, exploiting spurious features (e.g., response verbosity, agreeable tone, or sycophancy) that correlate with human preference labels in the training data rather than genuinely reflecting the intended objectives. In this paper, instead of probing these issues one at a time, we take a broader view of the reward hacking problem as shortcut behaviors and introduce a principled yet flexible approach to mitigate shortcut behaviors in preference-based reward learning. Inspired by the invariant theory in the kernel perspective, we propose Preference-based Reward Invariance for Shortcut Mitigation (PRISM), which learns group-invariant kernels with feature maps in a closed-form learning objective. Experimental results in several benchmarks show that our method consistently improves the accuracy of the reward model on diverse out-of-distribution tasks and reduces the dependency on shortcuts in downstream policy models, establishing a robust framework for preference-based alignment.", "authors": ["Wenqian Ye", "Guangtao Zheng", "Aidong Zhang"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-21", "url": "https://arxiv.org/abs/2510.19050", "pdf_url": "https://arxiv.org/pdf/2510.19050v1", "arxiv_id": "2510.19050", "doi": "10.48550/arXiv.2510.19050", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3426}
{"id": "b6934406d685908450255a2122583c3770a56586d7640b93eacb6670c89bd5c0", "sources": ["arxiv", "semantic_scholar"], "title": "Auto-Rubric: Learning From Implicit Weights to Explicit Rubrics for Reward Modeling", "abstract": "Conventional reward modeling relies on gradient descent over neural weights, creating opaque, data-hungry \"black boxes.\" We propose a paradigm shift from implicit to explicit reward parameterization, recasting optimization from continuous weight spaces to the discrete space of natural language rubrics. We introduce a training-free framework based on iterative rubric learning: it locally induces discriminative criteria via verification-driven refinement, and globally compresses the candidate criteria pool into a compact core set by maximizing an information-theoretic coding rate objective. We organize the compressed core set into a hierarchical rubric structure -- high-level evaluation dimensions supported by concrete verification checks -- serving as an interpretable, portable reward function. Empirically, our approach challenges prevailing data scaling assumptions: using only 70 preference pairs, our rubric-guided judges outperform fully trained reward models on diverse benchmarks. For instance, Qwen3-8B equipped with our learned rubrics achieves 80.91% on RewardBench2, surpassing the specialized Skywork-Reward-V2-Qwen3-8B (78.20%). These results demonstrate that alignment signals are highly compressible and can be effectively captured through explicit symbolic search.", "authors": ["Lipeng Xie", "Sen Huang", "Zhuo Zhang", "Anni Zou", "Yunpeng Zhai", "Dingchao Ren", "Kezun Zhang", "Haoyuan Hu", "Boyin Liu", "Haoran Chen", "Zhaoyang Liu", "Bolin Ding"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-20", "url": "https://arxiv.org/abs/2510.17314", "pdf_url": "https://arxiv.org/pdf/2510.17314v2", "arxiv_id": "2510.17314", "doi": null, "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3076}
{"id": "96f08c4422453cf7c153b6fe82cbe99deb9eb5cfe747c801d30a088826b59fa4", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Correlated Reward Models: Statistical Barriers and Opportunities", "abstract": "Random Utility Models (RUMs) are a classical framework for modeling user preferences and play a key role in reward modeling for Reinforcement Learning from Human Feedback (RLHF). However, a crucial shortcoming of many of these techniques is the Independence of Irrelevant Alternatives (IIA) assumption, which collapses \\emph{all} human preferences to a universal underlying utility function, yielding a coarse approximation of the range of human preferences. On the other hand, statistical and computational guarantees for models avoiding this assumption are scarce. In this paper, we investigate the statistical and computational challenges of learning a \\emph{correlated} probit model, a fundamental RUM that avoids the IIA assumption. First, we establish that the classical data collection paradigm of pairwise preference data is \\emph{fundamentally insufficient} to learn correlational information, explaining the lack of statistical and computational guarantees in this setting. Next, we demonstrate that \\emph{best-of-three} preference data provably overcomes these shortcomings, and devise a statistically and computationally efficient estimator with near-optimal performance. These results highlight the benefits of higher-order preference data in learning correlated utilities, allowing for more fine-grained modeling of human preferences. Finally, we validate these theoretical guarantees on several real-world datasets, demonstrating improved personalization of human preferences.", "authors": ["Yeshwanth Cherapanamjeri", "Constantinos Daskalakis", "Gabriele Farina", "Sobhan Mohammadpour"], "categories": ["cs.LG", "econ.EM", "stat.ML"], "fields_of_study": ["Computer Science", "Economics", "Mathematics"], "published_date": "2025-10-17", "url": "https://arxiv.org/abs/2510.15839", "pdf_url": "https://arxiv.org/pdf/2510.15839v2", "arxiv_id": "2510.15839", "doi": "10.48550/arXiv.2510.15839", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.338}
{"id": "00de8e8d4cf4b485f144a656c107a4e206080b2875bf5ecd5588be0fa939d356", "sources": ["arxiv", "semantic_scholar"], "title": "Reinforcement Learning with Stochastic Reward Machines", "abstract": "Reward machines are an established tool for dealing with reinforcement learning problems in which rewards are sparse and depend on complex sequences of actions. However, existing algorithms for learning reward machines assume an overly idealized setting where rewards have to be free of noise. To overcome this practical limitation, we introduce a novel type of reward machines, called stochastic reward machines, and an algorithm for learning them. Our algorithm, based on constraint solving, learns minimal stochastic reward machines from the explorations of a reinforcement learning agent. This algorithm can easily be paired with existing reinforcement learning algorithms for reward machines and guarantees to converge to an optimal policy in the limit. We demonstrate the effectiveness of our algorithm in two case studies and show that it outperforms both existing methods and a naive approach for handling noisy reward functions.", "authors": ["Jan Corazza", "Ivan Gavran", "Daniel Neider"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-16", "url": "https://arxiv.org/abs/2510.14837", "pdf_url": "https://arxiv.org/pdf/2510.14837v1", "arxiv_id": "2510.14837", "doi": "10.1609/aaai.v36i6.20594", "citation_count": 37, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/corazza/srm", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.5206}
{"id": "e56b1b6cbaaea933438182ff059046b34c2f406f22e68cf5fad1dc637af1f7aa", "sources": ["arxiv", "semantic_scholar"], "title": "Information-Theoretic Reward Modeling for Stable RLHF: Detecting and Mitigating Reward Hacking", "abstract": "Despite the success of Reinforcement Learning from Human Feedback (RLHF) in aligning language models with human values, reward hacking-or reward over-optimization-remains a major challenge. We identify two key obstacles to its mitigation: (1) reward misgeneralization in reward modeling, where reward models overfit to spurious, preference-irrelevant features; and (2) the lack of suitable regularization during RL optimization, as existing token-level constraints often over-restrict the policy space. To address these issues, we propose InfoRM, an information-theoretic reward modeling framework based on the Information Bottleneck (IB) principle, which filters out preference-irrelevant information to alleviate reward misgeneralization. We further observe that reward-hacked responses manifest as pronounced outliers in InfoRM's IB latent space, measured by Mahalanobis distance from the SFT-induced distribution. Motivated by this, we introduce IBL, a distribution-level regularization that penalizes such deviations, effectively expanding the optimization landscape while maintaining alignment. We prove that IBL is theoretically equivalent to the pessimistic RL objective within the IB latent space. Finally, we present Mahalanobis Outlier Probability (MOP), a statistical metric for quantifying reward hacking severity, enabling principled hyperparameter tuning and online mitigation such as early stopping. Extensive experiments across diverse LLMs and datasets confirm the generality of our findings, the effectiveness of InfoRM and IBL, and the reliability of MOP as a diagnostic tool-collectively advancing the state of RLHF.", "authors": ["Yuchun Miao", "Liang Ding", "Sen Zhang", "Rong Bao", "Lefei Zhang", "Dacheng Tao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-15", "url": "https://arxiv.org/abs/2510.13694", "pdf_url": "https://arxiv.org/pdf/2510.13694v1", "arxiv_id": "2510.13694", "doi": "10.48550/arXiv.2510.13694", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3357}
{"id": "64ea1df719f224643cf9e9195b5105dfc1317abcd4d662f99fb0e3f92909daf8", "sources": ["arxiv", "semantic_scholar"], "title": "Repairing Reward Functions with Feedback to Mitigate Reward Hacking", "abstract": "Human-designed reward functions for reinforcement learning (RL) agents are frequently misaligned with the humans' true, unobservable objectives, and thus act only as proxies. Optimizing for a misspecified proxy reward function often induces reward hacking, resulting in a policy misaligned with the human's true objectives. An alternative is to perform RL from human feedback, which involves learning a reward function from scratch by collecting human preferences over pairs of trajectories. However, building such datasets is costly. To address the limitations of both approaches, we propose Preference-Based Reward Repair (PBRR): an automated iterative framework that repairs a human-specified proxy reward function by learning an additive, transition-dependent correction term from preferences. A manually specified reward function can yield policies that are highly suboptimal under the ground-truth objective, yet corrections on only a few transitions may suffice to recover optimal performance. To identify and correct for those transitions, PBRR uses a targeted exploration strategy and a new preference-learning objective. We prove in tabular domains PBRR has a cumulative regret that matches, up to constants, that of prior preference-based RL methods. In addition, on a suite of reward-hacking benchmarks, PBRR consistently outperforms baselines that learn a reward function from scratch from preferences or modify the proxy reward function using other approaches, requiring substantially fewer preferences to learn high performing policies.", "authors": ["Stephane Hatgis-Kessell", "Logan Mondal Bhamidipaty", "Emma Brunskill"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-14", "url": "https://arxiv.org/abs/2510.13036", "pdf_url": "https://arxiv.org/pdf/2510.13036v2", "arxiv_id": "2510.13036", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2129}
{"id": "fbba9c49fbdca28416d16c7bf8e57d8d347c1cf9d77608543082a1f869ddb0f7", "sources": ["arxiv", "semantic_scholar"], "title": "APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport", "abstract": "The reward model (RM) plays a crucial role in aligning Large Language Models (LLMs) with human preferences through Reinforcement Learning, where the Bradley-Terry (BT) objective has been recognized as simple yet powerful, specifically for pairwise preference learning. However, BT-based RMs often struggle to effectively distinguish between similar preference responses, leading to insufficient separation between preferred and non-preferred outputs. Consequently, they may easily overfit easy samples and cannot generalize well to Out-Of-Distribution (OOD) samples, resulting in suboptimal performance. To address these challenges, this paper introduces an effective enhancement to BT-based RMs through an adaptive margin mechanism. Specifically, we design to dynamically adjust the RM focus on more challenging samples through margins, based on both semantic similarity and model-predicted reward differences, which is approached from a distributional perspective solvable with Optimal Transport (OT). By incorporating these factors into a principled OT cost matrix design, our adaptive margin enables the RM to better capture distributional differences between chosen and rejected responses, yielding significant improvements in performance, convergence speed, and generalization capabilities. Experimental results across multiple benchmarks demonstrate that our method outperforms several existing RM techniques, showcasing enhanced performance in both In-Distribution (ID) and OOD settings. Moreover, RLHF experiments support our practical effectiveness in better aligning LLMs with human preferences. Our code is available at https://github.com/BIRlz/APLOT", "authors": ["Zhuo Li", "Yuege Feng", "Dandan Guo", "Jinpeng Hu", "Anningzhe Gao", "Xiang Wan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-13", "url": "https://arxiv.org/abs/2510.10963", "pdf_url": "https://arxiv.org/pdf/2510.10963v1", "arxiv_id": "2510.10963", "doi": "10.18653/v1/2025.emnlp-main.281", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/BIRlz/APLOT", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.5153}
{"id": "d6fc5e2cdeba9ad1434fe8b175796eaee28a8cb82177135a4928bebb698b19e5", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Model Perspectives: Whose Opinions Do Reward Models Reward?", "abstract": "Reward models (RMs) are central to the alignment of language models (LMs). An RM often serves as a proxy for human preferences to guide downstream LM behavior. However, our understanding of RM behavior is limited. Our work (i) formalizes a framework for measuring the alignment of opinions captured by RMs, (ii) investigates the extent to which RMs demonstrate sociodemographic biases, and (iii) explores the effects of prompting to steer rewards towards the preferences of a target group. We study the subjective and diverse perspectives on controversial topics, which allows us to quantify RM perspectives in terms of their opinions, attitudes, and values. We show that RMs are poorly aligned with several demographic groups and can systematically reward harmful stereotypes, and steering alone is not enough to overcome these limitations. Our findings underscore the need for more careful consideration of RM behavior in model alignment during preference learning to prevent the propagation of unwanted social biases in the language technologies that we use.", "authors": [" Elle"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-07", "url": "https://arxiv.org/abs/2510.06391", "pdf_url": "https://arxiv.org/pdf/2510.06391v1", "arxiv_id": "2510.06391", "doi": "10.18653/v1/2025.emnlp-main.754", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3266}
{"id": "2742adcbf6ac3f20d94253aa313855099aef63d713d48f1f07bf358570660e32", "sources": ["arxiv", "semantic_scholar"], "title": "Margin Adaptive DPO: Leveraging Reward Model for Granular Control in Preference Optimization", "abstract": "Direct Preference Optimization (DPO) has emerged as a simple and effective method for aligning large language models. However, its reliance on a fixed temperature parameter leads to suboptimal training on diverse preference data, causing overfitting on easy examples and under-learning from informative ones. Recent methods have emerged to counter this. While IPO addresses general overfitting, its uniform regularization can be overly conservative. The more targeted approach of $β$-DPO suffers from its own limitations: its batch-level adaptation applies a single, compromised temperature to mixed-margin pairs, its linear update rule can produce unstable negative $β$ values, and its filtering mechanism discards potentially useful training signals. In this work, we introduce Margin-Adaptive Direct Preference Optimization (MADPO), a method that provides a stable, data-preserving, and instance-level solution. MADPO employs a practical two-step approach: it first trains a reward model to estimate preference margins and then uses these margins to apply a continuous, adaptive weight to the DPO loss for each individual training sample. This re-weighting scheme creates an effective target margin that is amplified for hard pairs and dampened for easy pairs, allowing for granular control over the learning signal. We provide a comprehensive theoretical analysis, proving that MADPO has a well-behaved optimization landscape and is robust to reward model estimation errors. We validate our theory with experiments on a summarization task using human preference data. MADPO consistently outperforms strong baselines across a comprehensive sweep of decoding temperatures.", "authors": ["Hyung Gyu Rho"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-10-06", "url": "https://arxiv.org/abs/2510.05342", "pdf_url": "https://arxiv.org/pdf/2510.05342v2", "arxiv_id": "2510.05342", "doi": "10.48550/arXiv.2510.05342", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3254}
{"id": "44421c4efbd4afb81532a01e87ac77639c15c18f1a70922f1444ac43d3502ae5", "sources": ["arxiv", "semantic_scholar"], "title": "Limited Preference Data? Learning Better Reward Model with Latent Space Synthesis", "abstract": "Reward modeling, crucial for aligning large language models (LLMs) with human preferences, is often bottlenecked by the high cost of preference data. Existing textual data synthesis methods are computationally expensive. We propose a novel framework LENS for synthesizing preference data directly in the LLM's latent embedding space. Our method employs a Variational Autoencoder (VAE) to learn a structured latent representation of response embeddings. By performing controlled perturbations in this latent space and decoding back to the embedding space, we efficiently generate diverse, semantically consistent synthetic preference pairs, bypassing costly text generation and annotation. We provide theoretical guarantees that our synthesized pairs approximately preserve original preference ordering and improve reward model generalization. Empirically, our latent-space synthesis significantly outperforms text-based augmentation on standard benchmarks, achieving superior results while being 18x faster in generation and using a 16,000x smaller model. Our work offers a scalable and effective alternative for enhancing reward modeling through efficient data augmentation. Code is publicly available at https://github.com/deeplearning-wisc/lens", "authors": ["Leitian Tao", "Xuefeng Du", "Sharon Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2509.26074", "pdf_url": "https://arxiv.org/pdf/2509.26074v2", "arxiv_id": "2509.26074", "doi": "10.48550/arXiv.2509.26074", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/deeplearning-wisc/lens", "venue": "arXiv.org", "quality_score": 0.4923}
{"id": "97b17b0db8aafe2780f2e3cde78953c05642ecc1538919580446e008547876d8", "sources": ["arxiv", "semantic_scholar"], "title": "Linking Process to Outcome: Conditional Reward Modeling for LLM Reasoning", "abstract": "Process Reward Models (PRMs) have emerged as a promising approach to enhance the reasoning capabilities of large language models (LLMs) by guiding their step-by-step reasoning toward a final answer. However, existing PRMs either treat each reasoning step in isolation, failing to capture inter-step dependencies, or struggle to align process rewards with the final outcome. Consequently, the reward signal fails to respect temporal causality in sequential reasoning and faces ambiguous credit assignment. These limitations make downstream models vulnerable to reward hacking and lead to suboptimal performance. In this work, we propose Conditional Reward Modeling (CRM) that frames LLM reasoning as a temporal process leading to a correct answer. The reward of each reasoning step is not only conditioned on the preceding steps but also explicitly linked to the final outcome of the reasoning trajectory. By enforcing conditional probability rules, our design captures the causal relationships among reasoning steps, with the link to the outcome allowing precise attribution of each intermediate step, thereby resolving credit assignment ambiguity. Further, through this consistent probabilistic modeling, the rewards produced by CRM enable more reliable cross-sample comparison. Experiments across Best-of-N sampling, beam search and reinforcement learning demonstrate that CRM consistently outperforms existing reward models, offering a principled framework for enhancing LLM reasoning. In particular, CRM is more robust to reward hacking and delivers stable downstream improvements without relying on verifiable rewards derived from ground truth.", "authors": ["Zheng Zhang", "Ziwei Shan", "Kaitao Song", "Yexin Li", "Kan Ren"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2509.26578", "pdf_url": "https://arxiv.org/pdf/2509.26578v2", "arxiv_id": "2509.26578", "doi": "10.48550/arXiv.2509.26578", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3185}
{"id": "a3abf5ec0161333c11f86fc6d8348b14ce74f8a686fc4dfd812ee751d45485e1", "sources": ["arxiv", "semantic_scholar"], "title": "Which Rewards Matter? Reward Selection for Reinforcement Learning under Limited Feedback", "abstract": "The ability of reinforcement learning algorithms to learn effective policies is determined by the rewards available during training. However, for practical problems, obtaining large quantities of reward labels is often infeasible due to computational or financial constraints, particularly when relying on human feedback. When reinforcement learning must proceed with limited feedback -- only a fraction of samples get rewards labeled -- a fundamental question arises: which samples should be labeled to maximize policy performance? We formalize this problem of reward selection for reinforcement learning from limited feedback (RLLF), introducing a new problem formulation that facilitates the study of strategies for selecting impactful rewards. Two types of selection strategies are investigated: (i) heuristics that rely on reward-free information such as state visitation and partial value functions, and (ii) strategies pre-trained using auxiliary evaluative feedback. We find that critical subsets of rewards are those that (1) guide the agent along optimal trajectories, and (2) support recovery toward near-optimal behavior after deviations. Effective selection methods yield near-optimal policies with significantly fewer reward labels than full supervision, establishing reward selection as a powerful paradigm for scaling reinforcement learning in feedback-limited settings.", "authors": ["Shreyas Chaudhari", "Renhao Zhang", "Philip S. Thomas", "Bruno Castro da Silva"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-30", "url": "https://arxiv.org/abs/2510.00144", "pdf_url": "https://arxiv.org/pdf/2510.00144v1", "arxiv_id": "2510.00144", "doi": "10.48550/arXiv.2510.00144", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3185}
{"id": "b4e4814223c78ec4d3f532999bf446809e8f8fbd81abf639fb155d21091ab271", "sources": ["arxiv", "semantic_scholar"], "title": "Circuit-Aware Reward Training: A Mechanistic Framework for Longtail Robustness in RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) reward models exhibit systematic failures on longtail distributions, leading to reward hacking and misalignment. We propose a mechanistic interpretability framework that identifies specialized neural circuits responsible for rare-event processing in reward models. Drawing from recent advances showing distributed specialization for rare tokens in language models\\citep{liu2025no, liu2025emergent}, we hypothesize that reward models also develop functionally distinct circuits for longtail scenarios. Our theoretical framework establishes formal connections between circuit specialization, reward generalization bounds, and longtail performance. We introduce \\textbf{Circuit-Aware Reward Training (CART)}, which uses circuit analysis to guide data augmentation, regularization, and ensemble strategies. This approach provides both theoretical insights into reward model failures and practical interventions for improving longtail robustness.", "authors": ["Jing Liu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-29", "url": "https://arxiv.org/abs/2509.24713", "pdf_url": "https://arxiv.org/pdf/2509.24713v1", "arxiv_id": "2509.24713", "doi": "10.48550/arXiv.2509.24713", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3174}
{"id": "b12d905b0f75627e05b4b0636ced579315ca39c244de46580c9d80d6e9d713d0", "sources": ["arxiv", "semantic_scholar"], "title": "Trust Region Reward Optimization and Proximal Inverse Reward Optimization Algorithm", "abstract": "Inverse Reinforcement Learning (IRL) learns a reward function to explain expert demonstrations. Modern IRL methods often use the adversarial (minimax) formulation that alternates between reward and policy optimization, which often lead to unstable training. Recent non-adversarial IRL approaches improve stability by jointly learning reward and policy via energy-based formulations but lack formal guarantees. This work bridges this gap. We first present a unified view showing canonical non-adversarial methods explicitly or implicitly maximize the likelihood of expert behavior, which is equivalent to minimizing the expected return gap. This insight leads to our main contribution: Trust Region Reward Optimization (TRRO), a framework that guarantees monotonic improvement in this likelihood via a Minorization-Maximization process. We instantiate TRRO into Proximal Inverse Reward Optimization (PIRO), a practical and stable IRL algorithm. Theoretically, TRRO provides the IRL counterpart to the stability guarantees of Trust Region Policy Optimization (TRPO) in forward RL. Empirically, PIRO matches or surpasses state-of-the-art baselines in reward recovery, policy imitation with high sample efficiency on MuJoCo and Gym-Robotics benchmarks and a real-world animal behavior modeling task.", "authors": ["Yang Chen", "Menglin Zou", "Jiaqi Zhang", "Yitan Zhang", "Junyi Yang", "Gael Gendron", "Libo Zhang", "Jiamou Liu", "Michael J. Witbrock"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-27", "url": "https://arxiv.org/abs/2509.23135", "pdf_url": "https://arxiv.org/pdf/2509.23135v3", "arxiv_id": "2509.23135", "doi": "10.48550/arXiv.2509.23135", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3151}
{"id": "dbab9c939c62ed8876fa1925a191c564d561fdcda9f4e6f6af5771f03492667a", "sources": ["arxiv", "semantic_scholar"], "title": "Preference-Guided Learning for Sparse-Reward Multi-Agent Reinforcement Learning", "abstract": "We study the problem of online multi-agent reinforcement learning (MARL) in environments with sparse rewards, where reward feedback is not provided at each interaction but only revealed at the end of a trajectory. This setting, though realistic, presents a fundamental challenge: the lack of intermediate rewards hinders standard MARL algorithms from effectively guiding policy learning. To address this issue, we propose a novel framework that integrates online inverse preference learning with multi-agent on-policy optimization into a unified architecture. At its core, our approach introduces an implicit multi-agent reward learning model, built upon a preference-based value-decomposition network, which produces both global and local reward signals. These signals are further used to construct dual advantage streams, enabling differentiated learning targets for the centralized critic and decentralized actors. In addition, we demonstrate how large language models (LLMs) can be leveraged to provide preference labels that enhance the quality of the learned reward model. Empirical evaluations on state-of-the-art benchmarks, including MAMuJoCo and SMACv2, show that our method achieves superior performance compared to existing baselines, highlighting its effectiveness in addressing sparse-reward challenges in online MARL.", "authors": ["The Viet Bui", "Tien Mai", "Hong Thanh Nguyen"], "categories": ["cs.LG", "cs.MA"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.21828", "pdf_url": "https://arxiv.org/pdf/2509.21828v1", "arxiv_id": "2509.21828", "doi": "10.48550/arXiv.2509.21828", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.314}
{"id": "5c0c37b90e9c180a75fa7d09218f3930b3b33de642aee8c39728246017d3d554", "sources": ["arxiv", "semantic_scholar"], "title": "SPARK: Synergistic Policy And Reward Co-Evolving Framework", "abstract": "Recent Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) increasingly use Reinforcement Learning (RL) for post-pretraining, such as RL with Verifiable Rewards (RLVR) for objective tasks and RL from Human Feedback (RLHF) for subjective tasks. However, RLHF incurs high costs and potential reward-policy mismatch due to reliance on human preferences, while RLVR still wastes supervision by discarding rollouts and correctness signals after each update. To address these challenges, we introduce the Synergistic Policy And Reward Co-Evolving Framework (SPARK), an efficient, on-policy, and stable method that builds on RLVR. Instead of discarding rollouts and correctness data, SPARK recycles this valuable information to simultaneously train the model itself as a generative reward model. This auxiliary training uses a mix of objectives, such as pointwise reward score, pairwise comparison, and evaluation conditioned on further-reflection responses, to teach the model to evaluate and improve its own responses. Our process eliminates the need for a separate reward model and costly human preference data. SPARK creates a positive co-evolving feedback loop: improved reward accuracy yields better policy gradients, which in turn produce higher-quality rollouts that further refine the reward model. Our unified framework supports test-time scaling via self-reflection without external reward models and their associated costs. We show that SPARK achieves significant performance gains on multiple LLM and LVLM models and multiple reasoning, reward models, and general benchmarks. For example, SPARK-VL-7B achieves an average 9.7% gain on 7 reasoning benchmarks, 12.1% on 2 reward benchmarks, and 1.5% on 8 general benchmarks over the baselines, demonstrating robustness and broad generalization.", "authors": ["Ziyu Liu", "Yuhang Zang", "Shengyuan Ding", "Yuhang Cao", "Xiaoyi Dong", "Haodong Duan", "Dahua Lin", "Jiaqi Wang"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-26", "url": "https://arxiv.org/abs/2509.22624", "pdf_url": "https://arxiv.org/pdf/2509.22624v1", "arxiv_id": "2509.22624", "doi": "10.48550/arXiv.2509.22624", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/InternLM/Spark", "venue": "arXiv.org", "quality_score": 0.4852}
{"id": "d131ab1547596287013b6a4eab582f3b9a4045e49e21e5bb8284225ae0393785", "sources": ["arxiv", "semantic_scholar"], "title": "GRPO is Secretly a Process Reward Model", "abstract": "Process reward models (PRMs) allow for fine-grained credit assignment in reinforcement learning (RL), and seemingly contrast with outcome reward models (ORMs), which assign a single reward to an entire trajectory. However, we provide theoretical proof in this work that the Group Relative Policy Optimization (GRPO) RL algorithm equipped with an ORM is in fact equivalent to a PRM-aware RL objective equipped with a non-trivial, Monte-Carlo-based PRM (given mild assumptions). Leveraging the framework of GRPO-as-a-PRM, we identify a flaw in the GRPO objective that interacts with imbalanced process steps and rewards to hinder both exploration and exploitation (under different conditions). We propose a simple modification to the algorithm to mitigate this defect ($λ$-GRPO), and show that LLMs tuned with $λ$-GRPO outperform LLMs tuned with standard GRPO on downstream reasoning tasks\\textemdash and reach peak performance more rapidly. These results show that we can leverage the hidden, built-in PRM structure within the vanilla GRPO algorithm to boost model performance without employing an explicit PRM, and with a negligible impact on training time and cost.", "authors": ["Michael Sullivan", "Alexander Koller"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-25", "url": "https://arxiv.org/abs/2509.21154", "pdf_url": "https://arxiv.org/pdf/2509.21154v4", "arxiv_id": "2509.21154", "doi": "10.48550/arXiv.2509.21154", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3128}
{"id": "01c59db9ecbe6fa44caf0f4469c0124517a18284f1618a8f5f2798d6a2f86fd3", "sources": ["arxiv", "semantic_scholar"], "title": "What Fundamental Structure in Reward Functions Enables Efficient Sparse-Reward Learning?", "abstract": "Sparse-reward reinforcement learning (RL) remains fundamentally hard: without structure, any agent needs $Ω(|\\mathcal{S}||\\mathcal{A}|/p)$ samples to recover rewards. We introduce Policy-Aware Matrix Completion (PAMC) as a first concrete step toward a structural reward learning framework. Our key idea is to exploit approximate low-rank + sparse structure in the reward matrix, under policy-biased (MNAR) sampling. We prove recovery guarantees with inverse-propensity weighting, and establish a visitation-weighted error-to-regret bound linking completion error to control performance. Importantly, when assumptions weaken, PAMC degrades gracefully: confidence intervals widen and the algorithm abstains, ensuring safe fallback to exploration. Empirically, PAMC improves sample efficiency across Atari-26 (10M steps), DM Control, MetaWorld MT50, D4RL offline RL, and preference-based RL benchmarks, outperforming DrQ-v2, DreamerV3, Agent57, T-REX/D-REX, and PrefPPO under compute-normalized comparisons. Our results highlight PAMC as a practical and principled tool when structural rewards exist, and as a concrete first instantiation of a broader structural reward learning perspective.", "authors": ["Ibne Farabi Shihab", "Sanjeda Akter", "Anuj Sharma"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-04", "url": "https://arxiv.org/abs/2509.03790", "pdf_url": "https://arxiv.org/pdf/2509.03790v2", "arxiv_id": "2509.03790", "doi": "10.48550/arXiv.2509.03790", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2888}
{"id": "3abdc739bc0b7070e2158b7650f636230fb2a86a0307fb2fb4b9ad7f12889c0a", "sources": ["arxiv", "semantic_scholar"], "title": "SharedRep-RLHF: A Shared Representation Approach to RLHF with Diverse Preferences", "abstract": "Uniform-reward reinforcement learning from human feedback (RLHF), which trains a single reward model to represent the preferences of all annotators, fails to capture the diversity of opinions across sub-populations, inadvertently favoring dominant groups. The state-of-the-art, MaxMin-RLHF, addresses this by learning group-specific reward models, and by optimizing for the group receiving the minimum reward, thereby promoting fairness. However, we identify that a key limitation of MaxMin-RLHF is its poor performance when the minimum-reward group is a minority. To mitigate this drawback, we introduce a novel framework, termed {\\em SharedRep-RLHF}. At its core, SharedRep-RLHF learns and leverages {\\em shared traits} in annotations among various groups, in contrast to learning separate reward models across groups. We first show that MaxMin-RLHF is provably suboptimal in learning shared traits, and then quantify the sample complexity of SharedRep-RLHF. Experiments across diverse natural language tasks showcase the effectiveness of SharedRep-RLHF compared to MaxMin-RLHF with a gain of up to 20% in win rate.", "authors": ["Arpan Mukherjee", "Marcello Bullo", "Deniz Gündüz"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-09-03", "url": "https://arxiv.org/abs/2509.03672", "pdf_url": "https://arxiv.org/pdf/2509.03672v1", "arxiv_id": "2509.03672", "doi": "10.48550/arXiv.2509.03672", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2876}
{"id": "e5702c497b56606ff5504820dcad2171002f137dc8b7ee5166f9080f182eebd5", "sources": ["arxiv", "semantic_scholar"], "title": "GRAM-R$^2$: Self-Training Generative Foundation Reward Models for Reward Reasoning", "abstract": "Significant progress in reward modeling over recent years has been driven by a paradigm shift from task-specific designs towards generalist reward models. Despite this trend, developing effective reward models remains a fundamental challenge: the heavy reliance on large-scale labeled preference data. Pre-training on abundant unlabeled data offers a promising direction, but existing approaches fall short of instilling explicit reasoning into reward models. To bridge this gap, we propose a self-training approach that leverages unlabeled data to elicit reward reasoning in reward models. Based on this approach, we develop GRAM-R$^2$, a generative reward model trained to produce not only preference labels but also accompanying reward rationales. GRAM-R$^2$ can serve as a foundation model for reward reasoning and can be applied to a wide range of tasks with minimal or no additional fine-tuning. It can support downstream applications such as response ranking and task-specific reward tuning. Experiments on response ranking, task adaptation, and reinforcement learning from human feedback demonstrate that GRAM-R$^2$ consistently delivers strong performance, outperforming several strong discriminative and generative baselines.", "authors": ["Chenglong Wang", "Yongyu Mu", "Hang Zhou", "Yifu Huo", "Ziming Zhu", "Jiali Zeng", "Murun Yang", "Bei Li", "Xiaoyang Hao", "Chunliang Zhang", "Fandong Meng", "Jingbo Zhu", "Tong Xiao"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-09-02", "url": "https://arxiv.org/abs/2509.02492", "pdf_url": "https://arxiv.org/pdf/2509.02492v3", "arxiv_id": "2509.02492", "doi": "10.48550/arXiv.2509.02492", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2865}
{"id": "31f06e60ac7f248fc0e4e2c2f898fda43c13977dd2bc09119a679df15e8d2556", "sources": ["arxiv", "semantic_scholar"], "title": "Counterfactual Reward Model Training for Bias Mitigation in Multimodal Reinforcement Learning", "abstract": "In reinforcement learning with human feedback (RLHF), reward models can efficiently learn and amplify latent biases within multimodal datasets, which can lead to imperfect policy optimization through flawed reward signals and decreased fairness. Bias mitigation studies have often applied passive constraints, which can fail under causal confounding. Here, we present a counterfactual reward model that introduces causal inference with multimodal representation learning to provide an unsupervised, bias-resilient reward signal. The heart of our contribution is the Counterfactual Trust Score, an aggregated score consisting of four components: (1) counterfactual shifts that decompose political framing bias from topical bias; (2) reconstruction uncertainty during counterfactual perturbations; (3) demonstrable violations of fairness rules for each protected attribute; and (4) temporal reward shifts aligned with dynamic trust measures. We evaluated the framework on a multimodal fake versus true news dataset, which exhibits framing bias, class imbalance, and distributional drift. Following methodologies similar to unsupervised drift detection from representation-based distances [1] and temporal robustness benchmarking in language models [2], we also inject synthetic bias across sequential batches to test robustness. The resulting system achieved an accuracy of 89.12% in fake news detection, outperforming the baseline reward models. More importantly, it reduced spurious correlations and unfair reinforcement signals. This pipeline outlines a robust and interpretable approach to fairness-aware RLHF, offering tunable bias reduction thresholds and increasing reliability in dynamic real-time policy making.", "authors": ["Sheryl Mathew", "N Harshit"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-27", "url": "https://arxiv.org/abs/2508.19567", "pdf_url": "https://arxiv.org/pdf/2508.19567v1", "arxiv_id": "2508.19567", "doi": "10.48550/arXiv.2508.19567", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2796}
{"id": "c9b19869f8cfdbe093f3650876eb40396c1fc8bced8787cd323eafa178fbf66b", "sources": ["arxiv", "semantic_scholar"], "title": "Fusing Rewards and Preferences in Reinforcement Learning", "abstract": "We present Dual-Feedback Actor (DFA), a reinforcement learning algorithm that fuses both individual rewards and pairwise preferences (if available) into a single update rule. DFA uses the policy's log-probabilities directly to model the preference probability, avoiding a separate reward-modeling step. Preferences can be provided by human-annotators (at state-level or trajectory-level) or be synthesized online from Q-values stored in an off-policy replay buffer. Under a Bradley-Terry model, we prove that minimizing DFA's preference loss recovers the entropy-regularized Soft Actor-Critic (SAC) policy. Our simulation results show that DFA trained on generated preferences matches or exceeds SAC on six control environments and demonstrates a more stable training process. With only a semi-synthetic preference dataset under Bradley-Terry model, our algorithm outperforms reward-modeling reinforcement learning from human feedback (RLHF) baselines in a stochastic GridWorld and approaches the performance of an oracle with true rewards.", "authors": ["Sadegh Khorasani", "Saber Salehkaleybar", "Negar Kiyavash", "Matthias Grossglauser"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-15", "url": "https://arxiv.org/abs/2508.11363", "pdf_url": "https://arxiv.org/pdf/2508.11363v1", "arxiv_id": "2508.11363", "doi": "10.48550/arXiv.2508.11363", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2658}
{"id": "663b5ef7858511b7bc7a8fdaa54efef7ef36cde7c7cbdfa663908cc941b4bb2a", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Reward Model via Sparse Autoencoder", "abstract": "Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models. Our code is available at https://github.com/schrieffer-z/sarm.", "authors": ["Shuyi Zhang", "Wei Shi", "Sihang Li", "Jiayi Liao", "Hengxing Cai", "Xiang Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-12", "url": "https://arxiv.org/abs/2508.08746", "pdf_url": "https://arxiv.org/pdf/2508.08746v5", "arxiv_id": "2508.08746", "doi": "10.48550/arXiv.2508.08746", "citation_count": 14, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/schrieffer-z/sarm", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.4055}
{"id": "b799dd705b4715541bc6e7d2cf3a69a580dec74f033ad0c213b6877e76fb430a", "sources": ["arxiv", "semantic_scholar"], "title": "ReCode: Reinforcing Code Generation with Reasoning-Process Rewards", "abstract": "In practice, rigorous reasoning is often a key driver of correct code, while Reinforcement Learning (RL) for code generation often neglects optimizing reasoning quality. Bringing process-level supervision into RL is appealing, but it faces two challenges. First, training reliable reward models to assess reasoning quality is bottlenecked by the scarcity of fine-grained preference data. Second, naively incorporating such neural rewards may suffer from reward hacking. This work proposes ReCode (Reasoning-Reinforced Code Generation), a novel RL training framework comprising: (1) Contrastive Reasoning-Process Reward Learning (CRPL), which trains a reward model with synthesized optimized and degraded reasoning variants to assess the quality of reasoning process; and (2) Consistency-Gated GRPO (CG-GRPO), which integrates the reasoning-process reward model into RL by gating neural reasoning-process rewards with strict execution outcomes, using execution correctness as a hard gate to mitigate reward hacking. Additionally, to assess the reward model's discriminative capability in assessing reasoning-process quality, we introduce LiveCodeBench-RewardBench (LCB-RB), a new benchmark comprising preference pairs of superior and inferior reasoning processes tailored for code generation. Experimental results across HumanEval(+), MBPP(+), LiveCodeBench, and BigCodeBench show that a 7B model trained with ReCode outperforms the base version by 16.1% and reaches performance comparable to GPT-4-Turbo. We further demonstrate the generalizability of ReCode by extending it to the math domain.", "authors": ["Lishui Fan", "Yu Zhang", "Mouxiang Chen", "Zhongxin Liu"], "categories": ["cs.SE", "cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-07", "url": "https://arxiv.org/abs/2508.05170", "pdf_url": "https://arxiv.org/pdf/2508.05170v3", "arxiv_id": "2508.05170", "doi": null, "citation_count": 21, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3356}
{"id": "79c008908f6d3ea0df9c8d715d23a747f4a412f187eccbe80eee972cdeb29bf4", "sources": ["arxiv", "semantic_scholar"], "title": "Difficulty-Based Preference Data Selection by DPO Implicit Reward Gap", "abstract": "Aligning large language models (LLMs) with human preferences is a critical challenge in AI research. While methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) are widely used, they often rely on large, costly preference datasets. The current work lacks methods for high-quality data selection specifically for preference data. In this work, we introduce a novel difficulty-based data selection strategy for preference datasets, grounded in the DPO implicit reward mechanism. By selecting preference data examples with smaller DPO implicit reward gaps, which are indicative of more challenging cases, we improve data efficiency and model alignment. Our approach consistently outperforms five strong baselines across multiple datasets and alignment tasks, achieving superior performance with only 10\\% of the original data. This principled, efficient selection method offers a promising solution for scaling LLM alignment with limited resources.", "authors": ["Xuan Qi", "Rongwu Xu", "Zhijing Jin"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-06", "url": "https://arxiv.org/abs/2508.04149", "pdf_url": "https://arxiv.org/pdf/2508.04149v2", "arxiv_id": "2508.04149", "doi": "10.48550/arXiv.2508.04149", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Difficulty-Based-Preference-Data-Select/Difficulty-Based-Preference-Data-Select", "venue": "arXiv.org", "quality_score": 0.3949}
{"id": "97da6da7b3c1d7f1733121c49ef497c79a0355e043107cc37e555d528e7c87fd", "sources": ["arxiv", "semantic_scholar"], "title": "Inferring Reward Machines and Transition Machines from Partially Observable Markov Decision Processes", "abstract": "Partially Observable Markov Decision Processes (POMDPs) are fundamental to many real-world applications. Although reinforcement learning (RL) has shown success in fully observable domains, learning policies from traces in partially observable environments remains challenging due to non-Markovian observations. Inferring an automaton to handle the non-Markovianity is a proven effective approach, but faces two limitations: 1) existing automaton representations focus only on reward-based non-Markovianity, leading to unnatural problem formulations; 2) inference algorithms face enormous computational costs. For the first limitation, we introduce Transition Machines (TMs) to complement existing Reward Machines (RMs). To develop a unified inference algorithm for both automata types, we propose the Dual Behavior Mealy Machine (DBMM) that subsumes both TMs and RMs. We then introduce DB-RPNI, a passive automata learning algorithm that efficiently infers DBMMs while avoiding the costly reductions required by prior work. We further develop optimization techniques and identify sufficient conditions for inferring the minimal correct automata. Experimentally, our inference method achieves speedups of up to three orders of magnitude over SOTA baselines.", "authors": ["Yuly Wu", "Jiamou Liu", "Libo Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-03", "url": "https://arxiv.org/abs/2508.01947", "pdf_url": "https://arxiv.org/pdf/2508.01947v1", "arxiv_id": "2508.01947", "doi": "10.48550/arXiv.2508.01947", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/sousoura/Inferring-Reward-Machines-and-Transition-Machines-from-POMDP.git", "venue": "arXiv.org", "quality_score": 0.3896}
{"id": "7491d5683c99bfeefe018fabf618c77bffcbc329edfcb0acaade8612acea86df", "sources": ["arxiv", "semantic_scholar"], "title": "The Bidirectional Process Reward Model", "abstract": "Process Reward Models (PRMs), which assign fine-grained scores to intermediate reasoning steps within a solution trajectory, have emerged as a promising approach to enhance the reasoning quality of Large Language Models (LLMs). However, most existing PRMs rely on a unidirectional left-to-right (L2R) evaluation scheme, which restricts their utilization of global context. In light of this challenge, we propose a novel bidirectional evaluation paradigm, named Bidirectional Process Reward Model (BiPRM). BiPRM incorporates a parallel right-to-left (R2L) evaluation stream, implemented via prompt reversal, alongside the conventional L2R flow. Then a gating mechanism is introduced to adaptively fuse the reward scores from both streams to yield a holistic quality assessment. Remarkably, compared to the original PRM, BiPRM introduces only a 0.3% parameter increase for the gating module, and the parallel execution of two streams incurs merely 5% inference time latency. Our extensive empirical evaluations spanning diverse benchmarks, LLM backbones, PRM objectives and sampling policies demonstrate that BiPRM consistently surpasses unidirectional baselines, achieving an average relative gain of 10.6% over 54 solution-level configurations and 37.7% in 12 step-level error detection scenarios. Generally, our results highlight the effectiveness, robustness and general applicability of BiPRM, offering a promising new direction for process-based reward modeling.", "authors": ["Lingyin Zhang", "Jun Gao", "Xiaoxue Ren", "Ziqiang Cao"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-08-03", "url": "https://arxiv.org/abs/2508.01682", "pdf_url": "https://arxiv.org/pdf/2508.01682v2", "arxiv_id": "2508.01682", "doi": "10.48550/arXiv.2508.01682", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2521}
{"id": "9587517d1a250d78a2ea7100894b5ca504acac6037a31a55e63166552dc8afac", "sources": ["arxiv", "semantic_scholar"], "title": "Policy Learning from Large Vision-Language Model Feedback without Reward Modeling", "abstract": "Offline reinforcement learning (RL) provides a powerful framework for training robotic agents using pre-collected, suboptimal datasets, eliminating the need for costly, time-consuming, and potentially hazardous online interactions. This is particularly useful in safety-critical real-world applications, where online data collection is expensive and impractical. However, existing offline RL algorithms typically require reward labeled data, which introduces an additional bottleneck: reward function design is itself costly, labor-intensive, and requires significant domain expertise. In this paper, we introduce PLARE, a novel approach that leverages large vision-language models (VLMs) to provide guidance signals for agent training. Instead of relying on manually designed reward functions, PLARE queries a VLM for preference labels on pairs of visual trajectory segments based on a language task description. The policy is then trained directly from these preference labels using a supervised contrastive preference learning objective, bypassing the need to learn explicit reward models. Through extensive experiments on robotic manipulation tasks from the MetaWorld, PLARE achieves performance on par with or surpassing existing state-of-the-art VLM-based reward generation methods. Furthermore, we demonstrate the effectiveness of PLARE in real-world manipulation tasks with a physical robot, further validating its practical applicability.", "authors": ["Tung M. Luu", "Donghoon Lee", "Younghwan Lee", "Chang D. Yoo"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-31", "url": "https://arxiv.org/abs/2507.23391", "pdf_url": "https://arxiv.org/pdf/2507.23391v1", "arxiv_id": "2507.23391", "doi": "10.1109/IROS60139.2025.11246902", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE/RJS International Conference on Intelligent RObots and Systems", "quality_score": 0.2486}
{"id": "fa115d47a00a5a7d7933f94e8466bc657c98c35867aa7921e91d0fdaaf043be5", "sources": ["arxiv", "semantic_scholar"], "title": "Multimodal LLMs as Customized Reward Models for Text-to-Image Generation", "abstract": "We introduce LLaVA-Reward, an efficient reward model designed to automatically evaluate text-to-image (T2I) generations across multiple perspectives, leveraging pretrained multimodal large language models (MLLMs). Existing MLLM-based approaches require instruction-following data for supervised fine-tuning and evaluate generation quality on analyzing text response, which is time-consuming and difficult to train. To address this problem, we propose LLaVA-Reward, which directly utilizes the hidden states of MLLMs given text-image pairs. To enhance the bidirectional interaction between visual and textual representations in decoder-only MLLMs, we further propose adding a Skip-connection Cross Attention (SkipCA) module. This design enhances text-image correlation reasoning by connecting early-layer visual features with later-layer hidden representations. In addition, LLaVA-Reward supports different types of preference data for efficient fine-tuning, including paired preference data and unpaired data. We train LLaVA-Reward on four evaluation perspectives: text-image alignment, fidelity/artifact, safety, and overall ranking. Empirical results demonstrate that LLaVA-Reward outperforms conventional and MLLM-based methods in generating human-aligned scores for automatic evaluations and inference-time scaling in text-to-image generations.", "authors": ["Shijie Zhou", "Ruiyi Zhang", "Huaisheng Zhu", "Branislav Kveton", "Yufan Zhou", "Jiuxiang Gu", "Jian Chen", "Changyou Chen"], "categories": ["cs.CV", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-28", "url": "https://arxiv.org/abs/2507.21391", "pdf_url": "https://arxiv.org/pdf/2507.21391v2", "arxiv_id": "2507.21391", "doi": "10.1109/ICCV51701.2025.01826", "citation_count": 9, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/sjz5202/LLaVA-Reward", "venue": "IEEE International Conference on Computer Vision", "quality_score": 0.379}
{"id": "138ff77d0bcb41dc48703732472de43dd9ea35b6c8f28d78370fd9bf23d32cae", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic and Generalizable Process Reward Modeling", "abstract": "Process Reward Models (PRMs) are crucial for guiding Large Language Models (LLMs) in complex scenarios by providing dense reward signals. However, existing PRMs primarily rely on heuristic approaches, which struggle with cross-domain generalization. While LLM-as-judge has been proposed to provide generalized rewards, current research has focused mainly on feedback results, overlooking the meaningful guidance embedded within the text. Additionally, static and coarse-grained evaluation criteria struggle to adapt to complex process supervision. To tackle these challenges, we propose Dynamic and Generalizable Process Reward Modeling (DG-PRM), which features a reward tree to capture and store fine-grained, multi-dimensional reward criteria. DG-PRM dynamically selects reward signals for step-wise reward scoring. To handle multifaceted reward signals, we pioneeringly adopt Pareto dominance estimation to identify discriminative positive and negative pairs. Experimental results show that DG-PRM achieves stunning performance on prevailing benchmarks, significantly boosting model performance across tasks with dense rewards. Further analysis reveals that DG-PRM adapts well to out-of-distribution scenarios, demonstrating exceptional generalizability.", "authors": ["Zhangyue Yin", "Qiushi Sun", "Zhiyuan Zeng", "Qinyuan Cheng", "Xipeng Qiu", "Xuanjing Huang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-23", "url": "https://arxiv.org/abs/2507.17849", "pdf_url": "https://arxiv.org/pdf/2507.17849v1", "arxiv_id": "2507.17849", "doi": "10.18653/v1/2025.acl-long.212", "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3076}
{"id": "2b71b255fd249af8aec4039afdd3d9827364095cd0504aa3b0f21f45a32a09bd", "sources": ["arxiv", "semantic_scholar"], "title": "Off-Policy Corrected Reward Modeling for Reinforcement Learning from Human Feedback", "abstract": "Reinforcement Learning from Human Feedback (RLHF) allows us to train models, such as language models (LMs), to follow complex human preferences. In RLHF for LMs, we first train an LM using supervised fine-tuning, sample pairs of responses, obtain human feedback, and use the resulting data to train a reward model (RM). RL methods are then used to train the LM to maximize the reward given by the RM. As training progresses, the responses generated by the LM no longer resemble the responses seen by the RM during training, leading to the RM becoming inaccurate. The score given by the RM keeps increasing, but the learned behavior no longer matches the human preferences. This issue is known as overoptimization. We investigate overoptimization from the point of view of distribution shift and show that the shift results in an inconsistent estimate of the RM parameters, leading to an inconsistent estimate of the policy gradient. We propose Off-Policy Corrected Reward Modeling (OCRM), which iteratively off-policy corrects the RM using importance weighting, without requiring new labels or samples. This results in a more accurate RM, which empirically leads to an improved final policy. We validate our approach in experiments with summarization and chatbot datasets and show that it performs significantly better than standard RLHF methods and baselines. Our implementation is available at https://github.com/JohannesAck/OffPolicyCorrectedRewardModeling", "authors": ["Johannes Ackermann", "Takashi Ishida", "Masashi Sugiyama"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-21", "url": "https://arxiv.org/abs/2507.15507", "pdf_url": "https://arxiv.org/pdf/2507.15507v1", "arxiv_id": "2507.15507", "doi": "10.48550/arXiv.2507.15507", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/JohannesAck/OffPolicyCorrectedRewardModeling", "venue": "arXiv.org", "quality_score": 0.3666}
{"id": "e68027ef6d0bcf43e1562e068b88d7404d53e4e19c5a178ab7c3ab4970c4867c", "sources": ["arxiv", "semantic_scholar"], "title": "Tiny Reward Models", "abstract": "Large decoder-based language models have become the dominant architecture for reward modeling in reinforcement learning from human feedback (RLHF). However, as reward models are increasingly deployed in test-time strategies, their inference costs become a growing concern. We present TinyRM, a family of small, bidirectional masked language models (MLMs) with as few as 400 million parameters, that rival the capabilities of models over 175 times larger on reasoning and safety preference modeling tasks. TinyRM combines FLAN-style prompting, Directional Low-Rank Adaptation (DoRA), and layer freezing to achieve strong performance on RewardBench, despite using significantly fewer resources. Our experiments suggest that small models benefit from domain-specific tuning strategies, particularly in reasoning, where lightweight finetuning methods are especially effective. While challenges remain in building generalist models and conversational preference modeling, our preliminary results highlight the promise of lightweight bidirectional architectures as efficient, scalable alternatives for preference modeling.", "authors": ["Sarah Pan"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-14", "url": "https://arxiv.org/abs/2507.09973", "pdf_url": "https://arxiv.org/pdf/2507.09973v1", "arxiv_id": "2507.09973", "doi": "10.48550/arXiv.2507.09973", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2292}
{"id": "bbbf858be2ae361da6e2a4fb1bfe371efe19ee7defcdf3a280189dd5d35224f7", "sources": ["arxiv", "semantic_scholar"], "title": "Recursive Reward Aggregation", "abstract": "In reinforcement learning (RL), aligning agent behavior with specific objectives typically requires careful design of the reward function, which can be challenging when the desired objectives are complex. In this work, we propose an alternative approach for flexible behavior alignment that eliminates the need to modify the reward function by selecting appropriate reward aggregation functions. By introducing an algebraic perspective on Markov decision processes (MDPs), we show that the Bellman equations naturally emerge from the recursive generation and aggregation of rewards, allowing for the generalization of the standard discounted sum to other recursive aggregations, such as discounted max and Sharpe ratio. Our approach applies to both deterministic and stochastic settings and integrates seamlessly with value-based and actor-critic algorithms. Experimental results demonstrate that our approach effectively optimizes diverse objectives, highlighting its versatility and potential for real-world applications.", "authors": ["Yuting Tang", "Yivan Zhang", "Johannes Ackermann", "Yu-Jie Zhang", "Soichiro Nishimori", "Masashi Sugiyama"], "categories": ["cs.LG", "math.CT"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-07-11", "url": "https://arxiv.org/abs/2507.08537", "pdf_url": "https://arxiv.org/pdf/2507.08537v2", "arxiv_id": "2507.08537", "doi": "10.48550/arXiv.2507.08537", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2257}
{"id": "55d62268686e902f42d2faf95ac5361cec21cd6eb8a73ecdd214f22d48014d3e", "sources": ["arxiv", "semantic_scholar"], "title": "Bradley-Terry and Multi-Objective Reward Modeling Are Complementary", "abstract": "Reward models trained on human preference data have demonstrated strong effectiveness in aligning Large Language Models (LLMs) with human intent under the framework of Reinforcement Learning from Human Feedback (RLHF). However, RLHF remains vulnerable to reward hacking, where the policy exploits imperfections in the reward function rather than genuinely learning the intended behavior. Although significant efforts have been made to mitigate reward hacking, they predominantly focus on and evaluate in-distribution scenarios, where the training and testing data for the reward model share the same distribution. In this paper, we empirically show that state-of-the-art methods struggle in more challenging out-of-distribution (OOD) settings. We further demonstrate that incorporating fine-grained multi-attribute scores helps address this challenge. However, the limited availability of high-quality data often leads to weak performance of multi-objective reward functions, which can negatively impact overall performance and become the bottleneck. To address this issue, we propose a unified reward modeling framework that jointly trains Bradley--Terry (BT) single-objective and multi-objective regression-based reward functions using a shared embedding space. We theoretically establish a connection between the BT loss and the regression objective and highlight their complementary benefits. Specifically, the regression task enhances the single-objective reward function's ability to mitigate reward hacking in challenging OOD settings, while BT-based training improves the scoring capability of the multi-objective reward function, enabling a 7B model to outperform a 70B baseline. Extensive experimental results demonstrate that our framework significantly improves both the robustness and the scoring performance of reward models.", "authors": ["Zhiwei Zhang", "Hui Liu", "Xiaomin Li", "Zhenwei Dai", "Jingying Zeng", "Fali Wang", "Minhua Lin", "Ramraj Chandradevan", "Zhen Li", "Chen Luo", "Xianfeng Tang", "Qi He", "Suhang Wang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-10", "url": "https://arxiv.org/abs/2507.07375", "pdf_url": "https://arxiv.org/pdf/2507.07375v1", "arxiv_id": "2507.07375", "doi": "10.48550/arXiv.2507.07375", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "9b3ac64ce71b6b84af6b5c71c4334a3aa83a892c2371c83980a89c413e268135", "sources": ["arxiv", "semantic_scholar"], "title": "Teach a Reward Model to Correct Itself: Reward Guided Adversarial Failure Discovery for Robust Reward Modeling", "abstract": "Reward modeling (RM), which captures human preferences to align large language models (LLMs), is increasingly employed in tasks such as model finetuning, response filtering, and ranking. However, due to the inherent complexity of human preferences and the limited coverage of available datasets, reward models often fail under distributional shifts or adversarial perturbations. Existing approaches for identifying such failure modes typically rely on prior knowledge about preference distributions or failure attributes, limiting their practicality in real-world settings where such information is unavailable. In this work, we propose a tractable, preference-distribution agnostic method for discovering reward model failure modes via reward guided controlled decoding. Building on this, we introduce REFORM, a self-improving reward modeling framework that enhances robustness by using the reward model itself to guide the generation of falsely scored responses. These adversarial examples are then used to augment the training data and patch the reward model's misaligned behavior. We evaluate REFORM on two widely used preference datasets Anthropic Helpful Harmless (HH) and PKU Beavertails and demonstrate that it significantly improves robustness without sacrificing reward quality. Notably, REFORM preserves performance both in direct evaluation and in downstream policy training, and further improves alignment quality by removing spurious correlations.", "authors": ["Pankayaraj Pathmanathan", "Furong Huang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-08", "url": "https://arxiv.org/abs/2507.06419", "pdf_url": "https://arxiv.org/pdf/2507.06419v3", "arxiv_id": "2507.06419", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "ACL 2026 Main Conference [Oral]", "quality_score": 0.2223}
{"id": "d07a0744583b2b3a36917654c542f44f5d6e8941a21b0c911971da2fc2100c84", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Reward Modeling with Active Concept Bottlenecks", "abstract": "We introduce Concept Bottleneck Reward Models (CB-RM), a reward modeling framework that enables interpretable preference learning through selective concept annotation. Unlike standard RLHF methods that rely on opaque reward functions, CB-RM decomposes reward prediction into human-interpretable concepts. To make this framework efficient in low-supervision settings, we formalize an active learning strategy that dynamically acquires the most informative concept labels. We propose an acquisition function based on Expected Information Gain and show that it significantly accelerates concept learning without compromising preference accuracy. Evaluated on the UltraFeedback dataset, our method outperforms baselines in interpretability and sample efficiency, marking a step towards more transparent, auditable, and human-aligned reward models.", "authors": ["Sonia Laguna", "Katarzyna Kobalczyk", "Julia E. Vogt", "Mihaela Van der Schaar"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-07", "url": "https://arxiv.org/abs/2507.04695", "pdf_url": "https://arxiv.org/pdf/2507.04695v2", "arxiv_id": "2507.04695", "doi": "10.48550/arXiv.2507.04695", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2211}
{"id": "1643c07009d41a7ebac6763a7c0d6743e5c81eff96f3d28994761bc055c04600", "sources": ["arxiv", "semantic_scholar"], "title": "Pre-Trained Policy Discriminators are General Reward Models", "abstract": "We offer a novel perspective on reward modeling by formulating it as a policy discriminator, which quantifies the difference between two policies to generate a reward signal, guiding the training policy towards a target policy with desired behaviors. Based on this conceptual insight, we propose a scalable pre-training method named Policy Discriminative Learning (POLAR), which trains a reward model (RM) to discern identical policies and discriminate different ones. Unlike traditional reward modeling methods relying on absolute preferences, POLAR captures the relative difference between one policy and an arbitrary target policy, which is a scalable, high-level optimization objective suitable for modeling generic ranking relationships. Leveraging the POLAR pre-training paradigm, we present a series of RMs with parameter scales from 1.8B to 7B. Empirical results show that POLAR substantially outperforms traditional non-pre-trained methods, significantly enhancing RM performance. For instance, POLAR-7B could improve preference accuracy from 54.8% to 81.0% on STEM tasks and from 57.9% to 85.5% on creative writing tasks compared to SOTA baselines. POLAR also shows robust generalization capabilities in RLHF using Reinforcement Fine-tuning (RFT), providing reliable reward signals and markedly enhancing policy performance--improving LLaMa3.1-8B from an average of 47.36% to 56.33% and Qwen2.5-32B from 64.49% to 70.47% on 20 benchmarks. Moreover, scaling experiments reveal a clear power-law relationship between computation and performance, supported by linear correlation coefficients approaching 0.99. The impressive performance, strong generalization, and scaling properties suggest that POLAR is a promising direction for developing general and strong reward models.", "authors": ["Shihan Dou", "Shichun Liu", "Yuming Yang", "Yicheng Zou", "Yunhua Zhou", "Shuhao Xing", "Chenhao Huang", "Qiming Ge", "Demin Song", "Haijun Lv", "Songyang Gao", "Chengqi Lv", "Enyu Zhou", "Honglin Guo", "Zhiheng Xi", "Wenwei Zhang", "Qipeng Guo", "Qi Zhang", "Xipeng Qiu", "Xuanjing Huang", "Tao Gui", "Kai Chen"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-07", "url": "https://arxiv.org/abs/2507.05197", "pdf_url": "https://arxiv.org/pdf/2507.05197v2", "arxiv_id": "2507.05197", "doi": "10.48550/arXiv.2507.05197", "citation_count": 12, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.301}
{"id": "1d61232ac9a584be3218c1e8dddc1e3e6ab9c921fd9db0785acb3c64118334e1", "sources": ["arxiv", "semantic_scholar"], "title": "ARF-RLHF: Adaptive Reward-Following for RLHF through Emotion-Driven Self-Supervision and Trace-Biased Dynamic Optimization", "abstract": "Current RLHF methods such as PPO and DPO typically reduce human preferences to binary labels, which are costly to obtain and too coarse to reflect individual variation. We observe that expressions of satisfaction and dissatisfaction follow stable linguistic patterns across users, indicating that more informative supervisory signals can be extracted from free-form feedback. Building on this insight, we introduce Adaptive Reward-Following (ARF), which converts natural feedback into continuous preference trajectories and optimizes them using the novel TraceBias algorithm. Across diverse LLMs and preference domains, ARF consistently outperforms PPO and DPO, improving alignment by up to 7.6%. Our results demonstrate that continuous reward modeling provides a scalable path toward personalized and theoretically grounded RLHF.", "authors": ["YuXuan Zhang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-03", "url": "https://arxiv.org/abs/2507.03069", "pdf_url": "https://arxiv.org/pdf/2507.03069v3", "arxiv_id": "2507.03069", "doi": "10.48550/arXiv.2507.03069", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2166}
{"id": "1be2e97a46581870f7e0c192880e517f768ec40d59f45e67cd9feaad0058ddee", "sources": ["arxiv", "semantic_scholar"], "title": "Uncertainty-aware Reward Design Process", "abstract": "Designing effective reward functions is a cornerstone of reinforcement learning (RL), yet it remains a challenging process due to the inefficiencies and inconsistencies inherent in conventional reward engineering methodologies. Recent advances have explored leveraging large language models (LLMs) to automate reward function design. However, their suboptimal performance in numerical optimization often yields unsatisfactory reward quality, while the evolutionary search paradigm demonstrates inefficient utilization of simulation resources, resulting in prohibitively lengthy design cycles with disproportionate computational overhead. To address these challenges, we propose the Uncertainty-aware Reward Design Process (URDP), a novel framework that integrates large language models to streamline reward function design and evaluation in RL environments. URDP quantifies candidate reward function uncertainty based on self-consistency analysis, enabling simulation-free identification of ineffective reward components while discovering novel reward components. Furthermore, we introduce uncertainty-aware Bayesian optimization (UABO), which incorporates uncertainty estimation to significantly enhance hyperparameter configuration efficiency. Finally, we construct a bi-level optimization architecture by decoupling the reward component optimization and the hyperparameter tuning. URDP orchestrates synergistic collaboration between the reward logic reasoning of the LLMs and the numerical optimization strengths of the Bayesian Optimization. We conduct a comprehensive evaluation of URDP across 35 diverse tasks spanning three benchmark environments. Our experimental results demonstrate that URDP not only generates higher-quality reward functions but also achieves significant improvements in the efficiency of automated reward design compared to existing approaches.", "authors": ["Yang Yang", "Xiaolu Zhou", "Bosong Ding", "Miao Xin"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-03", "url": "https://arxiv.org/abs/2507.02256", "pdf_url": "https://arxiv.org/pdf/2507.02256v1", "arxiv_id": "2507.02256", "doi": "10.48550/arXiv.2507.02256", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1378}
{"id": "952e485cff5ec38f2a0471bb6aa7c33a69ba9276c49e56db4a5d9aa9e3925c68", "sources": ["arxiv", "semantic_scholar"], "title": "Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy", "abstract": "Despite the critical role of reward models (RMs) in Reinforcement Learning from Human Feedback (RLHF), current state-of-the-art open RMs perform poorly on most existing evaluation benchmarks, failing to capture nuanced human preferences. We hypothesize that this brittleness stems primarily from limitations in preference datasets, which are often narrowly scoped, synthetically labeled, or lack rigorous quality control. To address these challenges, we present SynPref-40M, a large-scale preference dataset comprising 40 million preference pairs. To enable data curation at scale, we design a human-AI synergistic two-stage pipeline that leverages the complementary strengths of human annotation quality and AI scalability. In this pipeline, humans provide verified annotations, while LLMs perform automatic curation based on human guidance. Training on this preference mixture, we introduce Skywork-Reward-V2, a suite of eight reward models ranging from 0.6B to 8B parameters, trained on a carefully curated subset of 26 million preference pairs from SynPref-40M. We demonstrate that Skywork-Reward-V2 is versatile across a wide range of capabilities, including alignment with human preferences, objective correctness, safety, resistance to stylistic biases, and best-of-N scaling. These reward models achieve state-of-the-art performance across seven major reward model benchmarks, outperform generative reward models, and demonstrate strong downstream performance. Ablation studies confirm that effectiveness stems not only from data scale but also from high-quality curation. The Skywork-Reward-V2 series represents substantial progress in open reward models, demonstrating how human-AI curation synergy can unlock significantly higher data quality.", "authors": ["Chris Yuhao Liu", "Liang Zeng", "Yuzhen Xiao", "Jujie He", "Jiacai Liu", "Chaojie Wang", "Rui Yan", "Wei Shen", "Fuxiang Zhang", "Jiacheng Xu", "Yang Liu", "Yahui Zhou"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-02", "url": "https://arxiv.org/abs/2507.01352", "pdf_url": "https://arxiv.org/pdf/2507.01352v3", "arxiv_id": "2507.01352", "doi": "10.48550/arXiv.2507.01352", "citation_count": 154, "influential_citation_count": 17, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.6276}
{"id": "8edd344b927941def2ef488153ca17681c0be3af8228070e2686718ae436371d", "sources": ["arxiv", "semantic_scholar"], "title": "Activation Reward Models for Few-Shot Model Alignment", "abstract": "Aligning Large Language Models (LLMs) and Large Multimodal Models (LMMs) to human preferences is a central challenge in improving the quality of the models' generative outputs for real-world applications. A common approach is to use reward modeling to encode preferences, enabling alignment via post-training using reinforcement learning. However, traditional reward modeling is not easily adaptable to new preferences because it requires a separate reward model, commonly trained on large preference datasets. To address this, we introduce Activation Reward Models (Activation RMs) -- a novel few-shot reward modeling method that leverages activation steering to construct well-aligned reward signals using minimal supervision and no additional model finetuning. Activation RMs outperform existing few-shot reward modeling approaches such as LLM-as-a-judge with in-context learning, voting-based scoring, and token probability scoring on standard reward modeling benchmarks. Furthermore, we demonstrate the effectiveness of Activation RMs in mitigating reward hacking behaviors, highlighting their utility for safety-critical applications. Toward this end, we propose PreferenceHack, a novel few-shot setting benchmark, the first to test reward models on reward hacking in a paired preference format. Finally, we show that Activation RM achieves state-of-the-art performance on this benchmark, surpassing even GPT-4o.", "authors": ["Tianning Chai", "Chancharik Mitra", "Brandon Huang", "Gautam Rajendrakumar Gare", "Zhiqiu Lin", "Assaf Arbelle", "Leonid Karlinsky", "Rogerio Feris", "Trevor Darrell", "Deva Ramanan", "Roei Herzig"], "categories": ["cs.CV", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-02", "url": "https://arxiv.org/abs/2507.01368", "pdf_url": "https://arxiv.org/pdf/2507.01368v1", "arxiv_id": "2507.01368", "doi": "10.48550/arXiv.2507.01368", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2154}
{"id": "cd095d3ab3942adbfdc57a73f504426c4ea61b26a96ac5972080c23a74fe32e5", "sources": ["arxiv", "semantic_scholar"], "title": "Residual Reward Models for Preference-based Reinforcement Learning", "abstract": "Preference-based Reinforcement Learning (PbRL) provides a way to learn high-performance policies in environments where the reward signal is hard to specify, avoiding heuristic and time-consuming reward design. However, PbRL can suffer from slow convergence speed since it requires training in a reward model. Prior work has proposed learning a reward model from demonstrations and fine-tuning it using preferences. However, when the model is a neural network, using different loss functions for pre-training and fine-tuning can pose challenges to reliable optimization. In this paper, we propose a method to effectively leverage prior knowledge with a Residual Reward Model (RRM). An RRM assumes that the true reward of the environment can be split into a sum of two parts: a prior reward and a learned reward. The prior reward is a term available before training, for example, a user's ``best guess'' reward function, or a reward function learned from inverse reinforcement learning (IRL), and the learned reward is trained with preferences. We introduce state-based and image-based versions of RRM and evaluate them on several tasks in the Meta-World environment suite. Experimental results show that our method substantially improves the performance of a common PbRL method. Our method achieves performance improvements for a variety of different types of prior rewards, including proxy rewards, a reward obtained from IRL, and even a negated version of the proxy reward. We also conduct experiments with a Franka Panda to show that our method leads to superior performance on a real robot. It significantly accelerates policy learning for different tasks, achieving success in fewer steps than the baseline. The videos are presented at https://sunlighted.github.io/RRM-web/.", "authors": ["Chenyang Cao", "Miguel Rogel-García", "Mohamed Nabail", "Xueqian Wang", "Nicholas Rhinehart"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-07-01", "url": "https://arxiv.org/abs/2507.00611", "pdf_url": "https://arxiv.org/pdf/2507.00611v1", "arxiv_id": "2507.00611", "doi": "10.48550/arXiv.2507.00611", "citation_count": 2, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2143}
{"id": "9b388b7ecad4064a32451de556f4482595179c1871ab5120dcd76c66dcb51cea", "sources": ["arxiv", "semantic_scholar"], "title": "AutoRule: Reasoning Chain-of-thought Extracted Rule-based Rewards Improve Preference Learning", "abstract": "Rule-based rewards offer a promising strategy for improving reinforcement learning from human feedback (RLHF), but current approaches often rely on manual rule engineering. We present AutoRule, a fully automated method for extracting rules from preference feedback and formulating them into rule-based rewards. AutoRule extraction operates in three stages: it leverages a reasoning model to interpret user preferences, identifies candidate rules from the reasoning chain of these interpretations, and synthesizes them into a unified rule set. Leveraging the finalized rule set, we employ language-model verifiers to compute the fraction of rules satisfied by each output, using this metric as an auxiliary reward alongside the learned reward model during policy optimization. Training a Llama-3-8B model with AutoRule results in a 28.6\\% relative improvement in length-controlled win rate on AlpacaEval2.0, and a 6.1\\% relative gain in second-turn performance on a held-out MT-Bench subset, compared to a GRPO baseline trained with the same learned reward model but without the rule-based auxiliary reward. Our analysis confirms that the extracted rules exhibit good agreement with dataset preference. We find that AutoRule demonstrates reduced reward hacking compared to a learned reward model when run over two episodes. Finally, our case study suggests that the extracted rules capture unique qualities valued in different datasets. The extracted rules are provided in the appendix, and the code is open-sourced at https://github.com/cxcscmu/AutoRule.", "authors": ["Tevin Wang", "Chenyan Xiong"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-18", "url": "https://arxiv.org/abs/2506.15651", "pdf_url": "https://arxiv.org/pdf/2506.15651v1", "arxiv_id": "2506.15651", "doi": "10.48550/arXiv.2506.15651", "citation_count": 9, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/cxcscmu/AutoRule", "venue": "arXiv.org", "quality_score": 0.3081}
{"id": "4645bdf06984e503f6d43074c19d8d3ac9dc59d6c1e0c26162c454d7aff85b06", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Models in Deep Reinforcement Learning: A Survey", "abstract": "In reinforcement learning (RL), agents continually interact with the environment and use the feedback to refine their behavior. To guide policy optimization, reward models are introduced as proxies of the desired objectives, such that when the agent maximizes the accumulated reward, it also fulfills the task designer's intentions. Recently, significant attention from both academic and industrial researchers has focused on developing reward models that not only align closely with the true objectives but also facilitate policy optimization. In this survey, we provide a comprehensive review of reward modeling techniques within the deep RL literature. We begin by outlining the background and preliminaries in reward modeling. Next, we present an overview of recent reward modeling approaches, categorizing them based on the source, the mechanism, and the learning paradigm. Building on this understanding, we discuss various applications of these reward modeling techniques and review methods for evaluating reward models. Finally, we conclude by highlighting promising research directions in reward modeling. Altogether, this survey includes both established and emerging methods, filling the vacancy of a systematic review of reward models in current literature.", "authors": ["Rui Yu", "Shenghua Wan", "Yucen Wang", "Chen-Xiao Gao", "Le Gan", "Zongzhang Zhang", "De-Chuan Zhan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-18", "url": "https://arxiv.org/abs/2506.15421", "pdf_url": "https://arxiv.org/pdf/2506.15421v1", "arxiv_id": "2506.15421", "doi": "10.48550/arXiv.2506.15421", "citation_count": 28, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Joint Conference on Artificial Intelligence", "quality_score": 0.3656}
{"id": "58d0bedfbae10d42cfd2510ce7effd7659cf04ad02b8ac104fa07687495c49ab", "sources": ["arxiv", "semantic_scholar"], "title": "A General Framework for Off-Policy Learning with Partially-Observed Reward", "abstract": "Off-policy learning (OPL) in contextual bandits aims to learn a decision-making policy that maximizes the target rewards by using only historical interaction data collected under previously developed policies. Unfortunately, when rewards are only partially observed, the effectiveness of OPL degrades severely. Well-known examples of such partial rewards include explicit ratings in content recommendations, conversion signals on e-commerce platforms that are partial due to delay, and the issue of censoring in medical problems. One possible solution to deal with such partial rewards is to use secondary rewards, such as dwelling time, clicks, and medical indicators, which are more densely observed. However, relying solely on such secondary rewards can also lead to poor policy learning since they may not align with the target reward. Thus, this work studies a new and general problem of OPL where the goal is to learn a policy that maximizes the expected target reward by leveraging densely observed secondary rewards as supplemental data. We then propose a new method called Hybrid Policy Optimization for Partially-Observed Reward (HyPeR), which effectively uses the secondary rewards in addition to the partially-observed target reward to achieve effective OPL despite the challenging scenario. We also discuss a case where we aim to optimize not only the expected target reward but also the expected secondary rewards to some extent; counter-intuitively, we will show that leveraging the two objectives is in fact advantageous also for the optimization of only the target reward. Along with statistical analysis of our proposed methods, empirical evaluations on both synthetic and real-world data show that HyPeR outperforms existing methods in various scenarios.", "authors": ["Rikiya Takehi", "Masahiro Asami", "Kosuke Kawakami", "Yuta Saito"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-17", "url": "https://arxiv.org/abs/2506.14439", "pdf_url": "https://arxiv.org/pdf/2506.14439v1", "arxiv_id": "2506.14439", "doi": "10.48550/arXiv.2506.14439", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.1982}
{"id": "da2eee8ece386cc92ee841673ef9c046a8be31977e6f7e91d756ac9356c0df81", "sources": ["arxiv", "semantic_scholar"], "title": "TGDPO: Harnessing Token-Level Reward Guidance for Enhancing Direct Preference Optimization", "abstract": "Recent advancements in reinforcement learning from human feedback have shown that utilizing fine-grained token-level reward models can substantially enhance the performance of Proximal Policy Optimization (PPO) in aligning large language models. However, it is challenging to leverage such token-level reward as guidance for Direct Preference Optimization (DPO), since DPO is formulated as a sequence-level bandit problem. To address this challenge, this work decomposes the sequence-level PPO into a sequence of token-level proximal policy optimization problems and then frames the problem of token-level PPO with token-level reward guidance, from which closed-form optimal token-level policy and the corresponding token-level reward can be derived. Using the obtained reward and Bradley-Terry model, this work establishes a framework of computable loss functions with token-level reward guidance for DPO, and proposes a practical reward guidance based on the induced DPO reward. This formulation enables different tokens to exhibit varying degrees of deviation from reference policy based on their respective rewards. Experiment results demonstrate that our method achieves substantial performance improvements over DPO, with win rate gains of up to 7.5 points on MT-Bench, 6.2 points on AlpacaEval 2, and 4.3 points on Arena-Hard. Code is available at https://github.com/dvlab-research/TGDPO.", "authors": ["Mingkang Zhu", "Xi Chen", "Zhongdao Wang", "Bei Yu", "Hengshuang Zhao", "Jiaya Jia"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-17", "url": "https://arxiv.org/abs/2506.14574", "pdf_url": "https://arxiv.org/pdf/2506.14574v1", "arxiv_id": "2506.14574", "doi": "10.48550/arXiv.2506.14574", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/dvlab-research/TGDPO", "venue": "International Conference on Machine Learning", "quality_score": 0.3064}
{"id": "155b7e74e3282ba27ef05598bed14b68c7c23386d62407e3e91574eed0ecca66", "sources": ["arxiv", "semantic_scholar"], "title": "Learning to Explore in Diverse Reward Settings via Temporal-Difference-Error Maximization", "abstract": "Numerous heuristics and advanced approaches have been proposed for exploration in different settings for deep reinforcement learning. Noise-based exploration generally fares well with dense-shaped rewards and bonus-based exploration with sparse rewards. However, these methods usually require additional tuning to deal with undesirable reward settings by adjusting hyperparameters and noise distributions. Rewards that actively discourage exploration, i.e., with an action cost and no other dense signal to follow, can pose a major challenge. We propose a novel exploration method, Stable Error-seeking Exploration (SEE), that is robust across dense, sparse, and exploration-adverse reward settings. To this endeavor, we revisit the idea of maximizing the TD-error as a separate objective. Our method introduces three design choices to mitigate instability caused by far-off-policy learning, the conflict of interest of maximizing the cumulative TD-error in an episodic setting, and the non-stationary nature of TD-errors. SEE can be combined with off-policy algorithms without modifying the optimization pipeline of the original objective. In our experimental analysis, we show that a Soft-Actor Critic agent with the addition of SEE performs robustly across three diverse reward settings in a variety of tasks without hyperparameter adjustments.", "authors": ["Sebastian Griesbach", "Carlo D'Eramo"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-16", "url": "https://arxiv.org/abs/2506.13345", "pdf_url": "https://arxiv.org/pdf/2506.13345v2", "arxiv_id": "2506.13345", "doi": "10.48550/arXiv.2506.13345", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1971}
{"id": "3c6cdc1b7059053ba080edb0b2ad867579667988676be916a089956c890c7eba", "sources": ["arxiv", "semantic_scholar"], "title": "Similarity as Reward Alignment: Robust and Versatile Preference-based Reinforcement Learning", "abstract": "Preference-based Reinforcement Learning (PbRL) entails a variety of approaches for aligning models with human intent to alleviate the burden of reward engineering. However, most previous PbRL work has not investigated the robustness to labeler errors, inevitable with labelers who are non-experts or operate under time constraints. Additionally, PbRL algorithms often target very specific settings (e.g. pairwise ranked preferences or purely offline learning). We introduce Similarity as Reward Alignment (SARA), a simple contrastive framework that is both resilient to noisy labels and adaptable to diverse feedback formats and training paradigms. SARA learns a latent representation of preferred samples and computes rewards as similarities to the learned latent. We demonstrate strong performance compared to baselines on continuous control offline RL benchmarks. We further demonstrate SARA's versatility in applications such as trajectory filtering for downstream tasks, cross-task preference transfer, and reward shaping in online learning.", "authors": ["Sara Rajaram", "R. James Cotton", "Fabian H. Sinz"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-06-14", "url": "https://arxiv.org/abs/2506.12529", "pdf_url": "https://arxiv.org/pdf/2506.12529v1", "arxiv_id": "2506.12529", "doi": "10.48550/arXiv.2506.12529", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1948}
{"id": "3a51464484b8fb38ab68447771f1f83aa81d46ba950875d5a2f98a51c4659d40", "sources": ["arxiv", "semantic_scholar"], "title": "Multi-Task Reward Learning from Human Ratings", "abstract": "Reinforcement learning from human feedback (RLHF) has become a key factor in aligning model behavior with users' goals. However, while humans integrate multiple strategies when making decisions, current RLHF approaches often simplify this process by modeling human reasoning through isolated tasks such as classification or regression. In this paper, we propose a novel reinforcement learning (RL) method that mimics human decision-making by jointly considering multiple tasks. Specifically, we leverage human ratings in reward-free environments to infer a reward function, introducing learnable weights that balance the contributions of both classification and regression models. This design captures the inherent uncertainty in human decision-making and allows the model to adaptively emphasize different strategies. We conduct several experiments using synthetic human ratings to validate the effectiveness of the proposed approach. Results show that our method consistently outperforms existing rating-based RL methods, and in some cases, even surpasses traditional RL approaches.", "authors": ["Mingkang Wu", "Devin White", "Evelyn Rose", "Vernon Lawhern", "Nicholas R Waytowich", "Yongcan Cao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-10", "url": "https://arxiv.org/abs/2506.09183", "pdf_url": "https://arxiv.org/pdf/2506.09183v2", "arxiv_id": "2506.09183", "doi": "10.48550/arXiv.2506.09183", "citation_count": 1, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1902}
{"id": "e44c85e3e81318c8d3ca27c07416fc8688ff6c361ea1c45bc49bb145cbfcd92d", "sources": ["arxiv", "semantic_scholar"], "title": "Intra-Trajectory Consistency for Reward Modeling", "abstract": "Reward models are critical for improving large language models (LLMs), particularly in reinforcement learning from human feedback (RLHF) or inference-time verification. Current reward modeling typically relies on scores of overall responses to learn the outcome rewards for the responses. However, since the response-level scores are coarse-grained supervision signals, the reward model struggles to identify the specific components within a response trajectory that truly correlate with the scores, leading to poor generalization on unseen responses. In this paper, we propose to leverage generation probabilities to establish reward consistency between processes in the response trajectory, which allows the response-level supervisory signal to propagate across processes, thereby providing additional fine-grained signals for reward learning. Building on analysis under the Bayesian framework, we develop an intra-trajectory consistency regularization to enforce that adjacent processes with higher next-token generation probability maintain more consistent rewards. We apply the proposed regularization to the advanced outcome reward model, improving its performance on RewardBench. Besides, we show that the reward model trained with the proposed regularization induces better DPO-aligned policies and achieves better best-of-N (BON) inference-time verification results. Our code is provided in https://github.com/chaoyang101/ICRM.", "authors": ["Chaoyang Zhou", "Shunyu Liu", "Zengmao Wang", "Di Wang", "Rong-Cheng Tu", "Bo Du", "Dacheng Tao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-10", "url": "https://arxiv.org/abs/2506.09096", "pdf_url": "https://arxiv.org/pdf/2506.09096v3", "arxiv_id": "2506.09096", "doi": "10.48550/arXiv.2506.09096", "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/chaoyang101/ICRM", "venue": "arXiv.org", "quality_score": 0.294}
{"id": "acdb1c67eb8e501130bb9adc082d54b00cbdf3de0f74130edca567b6494a6eba", "sources": ["arxiv", "semantic_scholar"], "title": "GFRIEND: Generative Few-shot Reward Inference through EfficieNt DPO", "abstract": "The ability to train high-performing reward models with few-shot data is critical for enhancing the efficiency and scalability of Reinforcement Learning from Human Feedback (RLHF). We propose a data augmentation and expansion framework that enables generative reward models trained on small datasets to achieve comparable performance to those trained on large-scale datasets. Traditional methods to train a generative reward model, such as Direct Preference Optimization (DPO), are constrained by inefficiencies in sample pairing and limited data diversity. This work introduces preference refinement, which employs Chain-of-Thought (CoT) sampling to uncover diverse and high-quality preference relationships. It also incorporates a perplexity-based scoring mechanism to assign nuanced preference levels and utilizes Multi-level Direct Preference Optimization (M-DPO) to enable the model to capture finer-grained preference differences between samples. Experimental results demonstrate that the proposed method significantly enhances data efficiency and model performance, enabling reward models trained in a few-shot setting to achieve results on par with those trained on large-scale datasets. This study underscores the potential of data-efficient strategies in advancing reward model optimization, offering a robust solution for low-resource RLHF applications.", "authors": ["Yiyang Zhao", "Huiyu Bai", "Xuejiao Zhao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-10", "url": "https://arxiv.org/abs/2506.08965", "pdf_url": "https://arxiv.org/pdf/2506.08965v1", "arxiv_id": "2506.08965", "doi": "10.48550/arXiv.2506.08965", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1902}
{"id": "e42d4253c2f889bbbc4d97782b8fd8f6078211cb41b1638e0d158f28d5dca451", "sources": ["arxiv", "semantic_scholar"], "title": "Explicit Preference Optimization: No Need for an Implicit Reward Model", "abstract": "The generated responses of large language models (LLMs) are often fine-tuned to human preferences through a process called reinforcement learning from human feedback (RLHF). As RLHF relies on a challenging training sequence, whereby a separate reward model is independently learned and then later applied to LLM policy updates, ongoing research effort has targeted more straightforward alternatives. In this regard, direct preference optimization (DPO) and its many offshoots circumvent the need for a separate reward training step. Instead, through the judicious use of a reparameterization trick that induces an \\textit{implicit} reward, DPO and related methods consolidate learning to the minimization of a single loss function. And yet despite demonstrable success in some real-world settings, we prove that DPO-based objectives are nonetheless subject to sub-optimal regularization and counter-intuitive interpolation behaviors, underappreciated artifacts of the reparameterizations upon which they are based. To this end, we introduce an \\textit{explicit} preference optimization framework termed EXPO that requires no analogous reparameterization to achieve an implicit reward. Quite differently, we merely posit intuitively-appealing regularization factors from scratch that transparently avoid the potential pitfalls of key DPO variants, provably satisfying regularization desiderata that prior methods do not. Empirical results serve to corroborate our analyses and showcase the efficacy of EXPO.", "authors": ["Xiangkun Hu", "Lemin Kong", "Tong He", "David Wipf"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-06-09", "url": "https://arxiv.org/abs/2506.07492", "pdf_url": "https://arxiv.org/pdf/2506.07492v1", "arxiv_id": "2506.07492", "doi": "10.48550/arXiv.2506.07492", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1945}
{"id": "2ad7e7671e1fe9c422089ebc327a17cf009e2ff72aed5940a6eb3417f3ce1d0a", "sources": ["arxiv", "semantic_scholar"], "title": "AMoPO: Adaptive Multi-objective Preference Optimization without Reward Models and Reference Models", "abstract": "Existing multi-objective preference alignment methods for large language models (LLMs) face limitations: (1) the inability to effectively balance various preference dimensions, and (2) reliance on auxiliary reward/reference models introduces computational complexity. To address these challenges, we propose Adaptive Multi-objective Preference Optimization (AMoPO), a novel framework that achieves dynamic balance across preference dimensions. By introducing the multi-objective optimization paradigm to use the dimension-aware generation metrics as implicit rewards, AMoPO aligns LLMs with diverse preferences without additional reward models or reference models. We introduce an adaptive weight assignment mechanism that models the generation space as a Gaussian distribution, allowing dynamic prioritization of preference dimensions. Empirical results demonstrate that AMoPO outperforms state-of-the-art baselines by 28.5%, and the experiments on 7B, 14B, and 32B models reveal the scaling ability of AMoPO. Moreover, additional analysis of multiple dimensions verifies its adaptability and effectiveness. These findings validate AMoPO's capability to achieve dimension-aware preference alignment, highlighting its superiority. Our codes and datasets are available at https://github.com/Javkonline/AMoPO.", "authors": ["Qi Liu", "Jingqing Ruan", "Hao Li", "Haodong Zhao", "Desheng Wang", "Jiansong Chen", "Wan Guanglu", "Xunliang Cai", "Zhi Zheng", "Tong Xu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-08", "url": "https://arxiv.org/abs/2506.07165", "pdf_url": "https://arxiv.org/pdf/2506.07165v1", "arxiv_id": "2506.07165", "doi": "10.48550/arXiv.2506.07165", "citation_count": 8, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Javkonline/AMoPO", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2904}
{"id": "6a3f5aac88fd55555fee8029ecbbe11f6e1c4bc935d4b9c267cff40932520e37", "sources": ["arxiv", "semantic_scholar"], "title": "BadReward: Clean-Label Poisoning of Reward Models in Text-to-Image RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning text-to-image (T2I) models with human preferences. However, RLHF's feedback mechanism also opens new pathways for adversaries. This paper demonstrates the feasibility of hijacking T2I models by poisoning a small fraction of preference data with natural-appearing examples. Specifically, we propose BadReward, a stealthy clean-label poisoning attack targeting the reward model in multi-modal RLHF. BadReward operates by inducing feature collisions between visually contradicted preference data instances, thereby corrupting the reward model and indirectly compromising the T2I model's integrity. Unlike existing alignment poisoning techniques focused on single (text) modality, BadReward is independent of the preference annotation process, enhancing its stealth and practical threat. Extensive experiments on popular T2I models show that BadReward can consistently guide the generation towards improper outputs, such as biased or violent imagery, for targeted concepts. Our findings underscore the amplified threat landscape for RLHF in multi-modal systems, highlighting the urgent need for robust defenses. Disclaimer. This paper contains uncensored toxic content that might be offensive or disturbing to the readers.", "authors": ["Kaiwen Duan", "Hongwei Yao", "Yufei Chen", "Ziyun Li", "Tong Qiao", "Zhan Qin", "Cong Wang"], "categories": ["cs.LG", "cs.AI", "cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-03", "url": "https://arxiv.org/abs/2506.03234", "pdf_url": "https://arxiv.org/pdf/2506.03234v1", "arxiv_id": "2506.03234", "doi": "10.48550/arXiv.2506.03234", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1822}
{"id": "c67feb08b1efbdee3f5245fcd6a488fa38981befd75d780ac97b540ceb4e9200", "sources": ["arxiv", "semantic_scholar"], "title": "RewardBench 2: Advancing Reward Model Evaluation", "abstract": "Reward models are used throughout the post-training of language models to capture nuanced signals from preference data and provide a training target for optimization across instruction following, reasoning, safety, and more domains. The community has begun establishing best practices for evaluating reward models, from the development of benchmarks that test capabilities in specific skill areas to others that test agreement with human preferences. At the same time, progress in evaluation has not been mirrored by the effectiveness of reward models in downstream tasks -- simpler direct alignment algorithms are reported to work better in many cases. This paper introduces RewardBench 2, a new multi-skill reward modeling benchmark designed to bring new, challenging data for accuracy-based reward model evaluation -- models score about 20 points on average lower on RewardBench 2 compared to the first RewardBench -- while being highly correlated with downstream performance. Compared to most other benchmarks, RewardBench 2 sources new human prompts instead of existing prompts from downstream evaluations, facilitating more rigorous evaluation practices. In this paper, we describe our benchmark construction process and report how existing models perform on it, while quantifying how performance on the benchmark correlates with downstream use of the models in both inference-time scaling algorithms, like best-of-N sampling, and RLHF training algorithms like proximal policy optimization.", "authors": ["Saumya Malik", "Valentina Pyatkin", "Sander Land", "Jacob Morrison", "Noah A. Smith", "Hannaneh Hajishirzi", "Nathan Lambert"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-06-02", "url": "https://arxiv.org/abs/2506.01937", "pdf_url": "https://arxiv.org/pdf/2506.01937v2", "arxiv_id": "2506.01937", "doi": "10.48550/arXiv.2506.01937", "citation_count": 100, "influential_citation_count": 22, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.6809}
{"id": "17a35f64c9d38f0ffae023cb4c2a6bcc39d7caec9dbfbca99de868440badd091", "sources": ["arxiv", "semantic_scholar"], "title": "Accelerating RLHF Training with Reward Variance Increase", "abstract": "Reinforcement learning from human feedback (RLHF) is an essential technique for ensuring that large language models (LLMs) are aligned with human values and preferences during the post-training phase. As an effective RLHF approach, group relative policy optimization (GRPO) has demonstrated success in many LLM-based applications. However, efficient GRPO-based RLHF training remains a challenge. Recent studies reveal that a higher reward variance of the initial policy model leads to faster RLHF training. Inspired by this finding, we propose a practical reward adjustment model to accelerate RLHF training by provably increasing the reward variance and preserving the relative preferences and reward expectation. Our reward adjustment method inherently poses a nonconvex optimization problem, which is NP-hard to solve in general. To overcome the computational challenges, we design a novel $O(n \\log n)$ algorithm to find a global solution of the nonconvex reward adjustment model by explicitly characterizing the extreme points of the feasible set. As an important application, we naturally integrate this reward adjustment model into the GRPO algorithm, leading to a more efficient GRPO with reward variance increase (GRPOVI) algorithm for RLHF training. As an interesting byproduct, we provide an indirect explanation for the empirical effectiveness of GRPO with rule-based reward for RLHF training, as demonstrated in DeepSeek-R1. Experiment results demonstrate that the GRPOVI algorithm can significantly improve the RLHF training efficiency compared to the original GRPO algorithm.", "authors": ["Zonglin Yang", "Zhexuan Gu", "Houduo Qi", "Yancheng Yuan"], "categories": ["cs.LG", "cs.AI", "math.OC"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-29", "url": "https://arxiv.org/abs/2505.23247", "pdf_url": "https://arxiv.org/pdf/2505.23247v2", "arxiv_id": "2505.23247", "doi": "10.48550/arXiv.2505.23247", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1765}
{"id": "3ebf66805ee3bd1222f64bbb6b45f239de68a480cf0a1f63c41585b5d80316c9", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Reward Fairness in RLHF: From a Resource Allocation Perspective", "abstract": "Rewards serve as proxies for human preferences and play a crucial role in Reinforcement Learning from Human Feedback (RLHF). However, if these rewards are inherently imperfect, exhibiting various biases, they can adversely affect the alignment of large language models (LLMs). In this paper, we collectively define the various biases present in rewards as the problem of reward unfairness. We propose a bias-agnostic method to address the issue of reward fairness from a resource allocation perspective, without specifically designing for each type of bias, yet effectively mitigating them. Specifically, we model preference learning as a resource allocation problem, treating rewards as resources to be allocated while considering the trade-off between utility and fairness in their distribution. We propose two methods, Fairness Regularization and Fairness Coefficient, to achieve fairness in rewards. We apply our methods in both verification and reinforcement learning scenarios to obtain a fairness reward model and a policy model, respectively. Experiments conducted in these scenarios demonstrate that our approach aligns LLMs with human preferences in a more fair manner.", "authors": ["Sheng Ouyang", "Yulan Hu", "Ge Chen", "Qingyang Li", "Fuzheng Zhang", "Yong Liu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-29", "url": "https://arxiv.org/abs/2505.23349", "pdf_url": "https://arxiv.org/pdf/2505.23349v1", "arxiv_id": "2505.23349", "doi": "10.18653/v1/2025.acl-long.163", "citation_count": 12, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2785}
{"id": "dcb5997c76565e63cb72d094b02260523faef7fded02f5fc0df61da80e0f17d6", "sources": ["arxiv", "semantic_scholar"], "title": "Learning a Pessimistic Reward Model in RLHF", "abstract": "This work proposes `PET', a novel pessimistic reward fine-tuning method, to learn a pessimistic reward model robust against reward hacking in offline reinforcement learning from human feedback (RLHF). Traditional reward modeling techniques in RLHF train an imperfect reward model, on which a KL regularization plays a pivotal role in mitigating reward hacking when optimizing a policy. Such an intuition-based method still suffers from reward hacking, and the policies with large KL divergence from the dataset distribution are excluded during learning. In contrast, we show that when optimizing a policy on a pessimistic reward model fine-tuned through PET, reward hacking can be prevented without relying on any regularization. We test our methods on the standard TL;DR summarization dataset. We find that one can learn a high-quality policy on our pessimistic reward without using any regularization. Such a policy has a high KL divergence from the dataset distribution while having high performance in practice. In summary, our work shows the feasibility of learning a pessimistic reward model against reward hacking. The agent can greedily search for the policy with a high pessimistic reward without suffering from reward hacking.", "authors": ["Yinglun Xu", "Hangoo Kang", "Tarun Suresh", "Yuxuan Wan", "Gagandeep Singh"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.20556", "pdf_url": "https://arxiv.org/pdf/2505.20556v1", "arxiv_id": "2505.20556", "doi": "10.48550/arXiv.2505.20556", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "9080ca3d8309373118e20454a75d9c0b7497a98e4c553d80e2a8551c00b7a024", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding the Performance Gap in Preference Learning: A Dichotomy of RLHF and DPO", "abstract": "We present a fine-grained theoretical analysis of the performance gap between two-stage reinforcement learning from human feedback~(RLHF) and direct preference optimization~(DPO). Our study decomposes this gap into two sources: the explicit representation gap under exact optimization and the implicit representation gap under finite samples. In the exact optimization setting, we characterize how the relative capacities of the reward and policy model classes influence the final policy qualities. We show that RLHF, DPO, or online DPO can outperform one another depending on type of model mis-specifications. Notably, online DPO can outperform both RLHF and standard DPO when the reward and policy model classes are isomorphic and both mis-specified. In the approximate optimization setting, we provide a concrete construction where the ground-truth reward is sparse and show that RLHF requires significantly fewer samples than DPO to recover an effective reward model, highlighting a statistical advantage of two-stage learning. Together, these results provide a comprehensive understanding of the performance gap between RLHF and DPO under various settings, and offer practical insights into when each method is preferred.", "authors": ["Ruizhe Shi", "Minhak Song", "Runlong Zhou", "Zihan Zhang", "Maryam Fazel", "Simon S. Du"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-26", "url": "https://arxiv.org/abs/2505.19770", "pdf_url": "https://arxiv.org/pdf/2505.19770v5", "arxiv_id": "2505.19770", "doi": "10.48550/arXiv.2505.19770", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2603}
{"id": "61e2106a41124afb7efd442f3ceb89f3cead9caff1455a8988827998e3f53b70", "sources": ["arxiv", "semantic_scholar"], "title": "MOSLIM:Align with diverse preferences in prompts through reward classification", "abstract": "The multi-objective alignment of Large Language Models (LLMs) is essential for ensuring foundational models conform to diverse human preferences. Current research in this field typically involves either multiple policies or multiple reward models customized for various preferences, or the need to train a preference-specific supervised fine-tuning (SFT) model. In this work, we introduce a novel multi-objective alignment method, MOSLIM, which utilizes a single reward model and policy model to address diverse objectives. MOSLIM provides a flexible way to control these objectives through prompting and does not require preference training during SFT phase, allowing thousands of off-the-shelf models to be directly utilized within this training framework. MOSLIM leverages a multi-head reward model that classifies question-answer pairs instead of scoring them and then optimize policy model with a scalar reward derived from a mapping function that converts classification results from reward model into reward scores. We demonstrate the efficacy of our proposed method across several multi-objective benchmarks and conduct ablation studies on various reward model sizes and policy optimization methods. The MOSLIM method outperforms current multi-objective approaches in most results while requiring significantly fewer GPU computing resources compared with existing policy optimization methods.", "authors": ["Yu Zhang", "Wanli Jiang", "Zhengyu Yang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-24", "url": "https://arxiv.org/abs/2505.20336", "pdf_url": "https://arxiv.org/pdf/2505.20336v1", "arxiv_id": "2505.20336", "doi": "10.48550/arXiv.2505.20336", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1707}
{"id": "d18be7733c0fd29534dea4d41883e67e738dd9f9de135f39b096bbeb3954deb1", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Model Overoptimisation in Iterated RLHF", "abstract": "Reinforcement learning from human feedback (RLHF) is a widely used method for aligning large language models with human preferences. However, RLHF often suffers from reward model overoptimisation, in which models overfit to the reward function, resulting in non-generalisable policies that exploit the idiosyncrasies and peculiarities of the reward function. A common mitigation is iterated RLHF, in which reward models are repeatedly retrained with updated human feedback and policies are re-optimised. Despite its increasing adoption, the dynamics of overoptimisation in this setting remain poorly understood. In this work, we present the first comprehensive study of overoptimisation in iterated RLHF. We systematically analyse key design choices - how reward model training data is transferred across iterations, which reward function is used for optimisation, and how policies are initialised. Using the controlled AlpacaFarm benchmark, we observe that overoptimisation tends to decrease over successive iterations, as reward models increasingly approximate ground-truth preferences. However, performance gains diminish over time, and while reinitialising from the base policy is robust, it limits optimisation flexibility. Other initialisation strategies often fail to recover from early overoptimisation. These findings offer actionable insights for building more stable and generalisable RLHF pipelines.", "authors": ["Lorenz Wolf", "Robert Kirk", "Mirco Musolesi"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-23", "url": "https://arxiv.org/abs/2505.18126", "pdf_url": "https://arxiv.org/pdf/2505.18126v2", "arxiv_id": "2505.18126", "doi": "10.48550/arXiv.2505.18126", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "d1f29db2306123334fc5447c67bf4c3698cc491dfe17f5f30f05a2c17f5ec6ce", "sources": ["arxiv", "semantic_scholar"], "title": "Think-RM: Enabling Long-Horizon Reasoning in Generative Reward Models", "abstract": "Reinforcement learning from human feedback (RLHF) has become a powerful post-training paradigm for aligning large language models with human preferences. A core challenge in RLHF is constructing accurate reward signals, where the conventional Bradley-Terry reward models (BT RMs) often suffer from sensitivity to data size and coverage, as well as vulnerability to reward hacking. Generative reward models (GenRMs) offer a more robust alternative by generating chain-of-thought (CoT) rationales followed by a final reward. However, existing GenRMs rely on shallow, vertically scaled reasoning, limiting their capacity to handle nuanced or complex (e.g., reasoning-intensive) tasks. Moreover, their pairwise preference outputs are incompatible with standard RLHF algorithms that require pointwise reward signals. In this work, we introduce Think-RM, a training framework that enables long-horizon reasoning in GenRMs by modeling an internal thinking process. Rather than producing structured, externally provided rationales, Think-RM generates flexible, self-guided reasoning traces that support advanced capabilities such as self-reflection, hypothetical reasoning, and divergent reasoning. To elicit these reasoning abilities, we first warm-up the models by supervised fine-tuning (SFT) over long CoT data. We then further improve the model's long-horizon abilities by rule-based reinforcement learning (RL). In addition, we propose a novel pairwise RLHF pipeline that directly optimizes policies using pairwise preference rewards, eliminating the need for pointwise reward conversion and enabling more effective use of Think-RM outputs. Experiments show that Think-RM achieves state-of-the-art results on RM-Bench, outperforming both BT RM and vertically scaled GenRM by 8%. When combined with our pairwise RLHF pipeline, it demonstrates superior end-policy performance compared to traditional approaches.", "authors": ["Ilgee Hong", "Changlong Yu", "Liang Qiu", "Weixiang Yan", "Zhenghao Xu", "Haoming Jiang", "Qingru Zhang", "Qin Lu", "Xin Liu", "Chao Zhang", "Tuo Zhao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-22", "url": "https://arxiv.org/abs/2505.16265", "pdf_url": "https://arxiv.org/pdf/2505.16265v1", "arxiv_id": "2505.16265", "doi": "10.48550/arXiv.2505.16265", "citation_count": 11, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.301}
{"id": "947b81b4f5e9316484a0d8d5370949384e4fd7316f009f451303b29bc2147a7d", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Reasoning Model", "abstract": "Reward models play a critical role in guiding large language models toward outputs that align with human expectations. However, an open challenge remains in effectively utilizing test-time compute to enhance reward model performance. In this work, we introduce Reward Reasoning Models (RRMs), which are specifically designed to execute a deliberate reasoning process before generating final rewards. Through chain-of-thought reasoning, RRMs leverage additional test-time compute for complex queries where appropriate rewards are not immediately apparent. To develop RRMs, we implement a reinforcement learning framework that fosters self-evolved reward reasoning capabilities without requiring explicit reasoning traces as training data. Experimental results demonstrate that RRMs achieve superior performance on reward modeling benchmarks across diverse domains. Notably, we show that RRMs can adaptively exploit test-time compute to further improve reward accuracy. The pretrained reward reasoning models are available at https://huggingface.co/Reward-Reasoning.", "authors": ["Jiaxin Guo", "Zewen Chi", "Li Dong", "Qingxiu Dong", "Xun Wu", "Shaohan Huang", "Furu Wei"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-20", "url": "https://arxiv.org/abs/2505.14674", "pdf_url": "https://arxiv.org/pdf/2505.14674v1", "arxiv_id": "2505.14674", "doi": "10.48550/arXiv.2505.14674", "citation_count": 35, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3891}
{"id": "319da1303360bd98715a2d3605071061a190aa7792a308950cdf0253f15631e0", "sources": ["arxiv", "semantic_scholar"], "title": "Bias Fitting to Mitigate Length Bias of Reward Model in RLHF", "abstract": "Reinforcement Learning from Human Feedback relies on reward models to align large language models with human preferences. However, RLHF often suffers from reward hacking, wherein policy learning exploits flaws in the trained reward model to maximize reward scores without genuinely aligning with human preferences. A significant example of such reward hacking is length bias, where reward models usually favor longer responses irrespective of actual response quality. Previous works on length bias have notable limitations, these approaches either mitigate bias without characterizing the bias form, or simply assume a linear length-reward relation. To accurately model the intricate nature of length bias and facilitate more effective bias mitigation, we propose FiMi-RM (Bias Fitting to Mitigate Length Bias of Reward Model in RLHF), a framework that autonomously learns and corrects underlying bias patterns. Our approach consists of three stages: First, we train a standard reward model which inherently contains length bias. Next, we deploy a lightweight fitting model to explicitly capture the non-linear relation between length and reward. Finally, we incorporate this learned relation into the reward model to debias. Experimental results demonstrate that FiMi-RM achieves a more balanced length-reward distribution. Furthermore, when applied to alignment algorithms, our debiased reward model improves length-controlled win rate and reduces verbosity without compromising its performance.", "authors": ["Kangwen Zhao", "Jianfeng Cai", "Jinhua Zhu", "Ruopei Sun", "Dongyun Xue", "Wengang Zhou", "Li Li", "Houqiang Li"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-19", "url": "https://arxiv.org/abs/2505.12843", "pdf_url": "https://arxiv.org/pdf/2505.12843v1", "arxiv_id": "2505.12843", "doi": "10.48550/arXiv.2505.12843", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "ec39b36546472882af42521050f29e6acbff0edde3de1206625b34c2044c7ed6", "sources": ["arxiv", "semantic_scholar"], "title": "Rethinking Reward Model Evaluation Through the Lens of Reward Overoptimization", "abstract": "Reward models (RMs) play a crucial role in reinforcement learning from human feedback (RLHF), aligning model behavior with human preferences. However, existing benchmarks for reward models show a weak correlation with the performance of optimized policies, suggesting that they fail to accurately assess the true capabilities of RMs. To bridge this gap, we explore several evaluation designs through the lens of reward overoptimization\\textemdash a phenomenon that captures both how well the reward model aligns with human preferences and the dynamics of the learning signal it provides to the policy. The results highlight three key findings on how to construct a reliable benchmark: (i) it is important to minimize differences between chosen and rejected responses beyond correctness, (ii) evaluating reward models requires multiple comparisons across a wide range of chosen and rejected responses, and (iii) given that reward models encounter responses with diverse representations, responses should be sourced from a variety of models. However, we also observe that a extremely high correlation with degree of overoptimization leads to comparatively lower correlation with certain downstream performance. Thus, when designing a benchmark, it is desirable to use the degree of overoptimization as a useful tool, rather than the end goal.", "authors": ["Sunghwan Kim", "Dongjin Kang", "Taeyoon Kwon", "Hyungjoo Chae", "Dongha Lee", "Jinyoung Yeo"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-19", "url": "https://arxiv.org/abs/2505.12763", "pdf_url": "https://arxiv.org/pdf/2505.12763v1", "arxiv_id": "2505.12763", "doi": "10.48550/arXiv.2505.12763", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2258}
{"id": "bc8c6db98eb94e802239b6901ef44a6fd6be8d3b5500d0260a84699df6e623f7", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Pareto-Optimal Rewards from Noisy Preferences: A Framework for Multi-Objective Inverse Reinforcement Learning", "abstract": "As generative agents become increasingly capable, alignment of their behavior with complex human values remains a fundamental challenge. Existing approaches often simplify human intent through reduction to a scalar reward, overlooking the multi-faceted nature of human feedback. In this work, we introduce a theoretical framework for preference-based Multi-Objective Inverse Reinforcement Learning (MO-IRL), where human preferences are modeled as latent vector-valued reward functions. We formalize the problem of recovering a Pareto-optimal reward representation from noisy preference queries and establish conditions for identifying the underlying multi-objective structure. We derive tight sample complexity bounds for recovering $ε$-approximations of the Pareto front and introduce a regret formulation to quantify suboptimality in this multi-objective setting. Furthermore, we propose a provably convergent algorithm for policy optimization using preference-inferred reward cones. Our results bridge the gap between practical alignment techniques and theoretical guarantees, providing a principled foundation for learning aligned behaviors in a high-dimension and value-pluralistic environment.", "authors": ["Kalyan Cherukuri", "Aarav Lala"], "categories": ["cs.LG", "cs.AI", "cs.CG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-17", "url": "https://arxiv.org/abs/2505.11864", "pdf_url": "https://arxiv.org/pdf/2505.11864v3", "arxiv_id": "2505.11864", "doi": "10.48550/arXiv.2505.11864", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1627}
{"id": "f52fd260499c3a19e4e32b879fbf872f179a1a334e51c318c2cfe207e013737e", "sources": ["arxiv", "semantic_scholar"], "title": "Detecting Prefix Bias in LLM-based Reward Models", "abstract": "Reinforcement Learning with Human Feedback (RLHF) has emerged as a key paradigm for task-specific fine-tuning of language models using human preference data. While numerous publicly available preference datasets provide pairwise comparisons of responses, the potential for biases in the resulting reward models remains underexplored. In this work, we introduce novel methods to detect and evaluate prefix bias -- a systematic shift in model preferences triggered by minor variations in query prefixes -- in LLM-based reward models trained on such datasets. We leverage these metrics to reveal significant biases in preference models across racial and gender dimensions. Our comprehensive evaluation spans diverse open-source preference datasets and reward model architectures, demonstrating susceptibility to this kind of bias regardless of the underlying model architecture. Furthermore, we propose a data augmentation strategy to mitigate these biases, showing its effectiveness in reducing the impact of prefix bias. Our findings highlight the critical need for bias-aware dataset design and evaluation in developing fair and reliable reward models, contributing to the broader discourse on fairness in AI.", "authors": ["Ashwin Kumar", "Yuzi He", "Aram H. Markosyan", "Bobbie Chern", "Imanol Arrieta-Ibarra"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-13", "url": "https://arxiv.org/abs/2505.13487", "pdf_url": "https://arxiv.org/pdf/2505.13487v2", "arxiv_id": "2505.13487", "doi": "10.1145/3715275.3732204", "citation_count": 13, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": "Conference on Fairness, Accountability and Transparency", "quality_score": 0.2865}
{"id": "9177fd7c623e81899daaa21d3a2583d6809ec941b4f845c3bb7a0433b993d6d4", "sources": ["arxiv", "semantic_scholar"], "title": "On the Robustness of Reward Models for Language Model Alignment", "abstract": "The Bradley-Terry (BT) model is widely practiced in reward modeling for reinforcement learning with human feedback (RLHF). Despite its effectiveness, reward models (RMs) trained with BT model loss are prone to over-optimization, losing generalizability to unseen input distributions. In this paper, we study the cause of over-optimization in RM training and its downstream effects on the RLHF procedure, accentuating the importance of distributional robustness of RMs in unseen data. First, we show that the excessive dispersion of hidden state norms is the main source of over-optimization. Then, we propose batch-wise sum-to-zero regularization (BSR) to enforce zero-centered reward sum per batch, constraining the rewards with extreme magnitudes. We assess the impact of BSR in improving robustness in RMs through four scenarios of over-optimization, where BSR consistently manifests better robustness. Subsequently, we compare the plain BT model and BSR on RLHF training and empirically show that robust RMs better align the policy to the gold preference model. Finally, we apply BSR to high-quality data and models, which surpasses state-of-the-art RMs in the 8B scale by adding more than 5% in complex preference prediction tasks. By conducting RLOO training with 8B RMs, AlpacaEval 2.0 reduces generation length by 40% while adding a 7% increase in win rate, further highlighting that robustness in RMs induces robustness in RLHF training. We release the code, data, and models: https://github.com/LinkedIn-XFACT/RM-Robustness.", "authors": ["Jiwoo Hong", "Noah Lee", "Eunki Kim", "Guijin Son", "Woojin Chung", "Aman Gupta", "Shao Tang", "James Thorne"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-12", "url": "https://arxiv.org/abs/2505.07271", "pdf_url": "https://arxiv.org/pdf/2505.07271v1", "arxiv_id": "2505.07271", "doi": "10.48550/arXiv.2505.07271", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/LinkedIn-XFACT/RM-Robustness", "venue": "International Conference on Machine Learning", "quality_score": 0.2698}
{"id": "c65e399c70be1f592947a35b249bf97aa5a7c65dea4131bd6bad673fe27f41db", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Guarantee of Reward Modeling Using Deep Neural Networks", "abstract": "In this work, we study the learning theory of reward modeling with pairwise comparison data using deep neural networks. We establish a novel non-asymptotic regret bound for deep reward estimators in a non-parametric setting, which depends explicitly on the network architecture. Furthermore, to underscore the critical importance of clear human beliefs, we introduce a margin-type condition that assumes the conditional winning probability of the optimal action in pairwise comparisons is significantly distanced from 1/2. This condition enables a sharper regret bound, which substantiates the empirical efficiency of Reinforcement Learning from Human Feedback and highlights clear human beliefs in its success. Notably, this improvement stems from high-quality pairwise comparison data implied by the margin-type condition, is independent of the specific estimators used, and thus applies to various learning algorithms and models.", "authors": ["Yuanhang Luo", "Yeheng Ge", "Ruijian Han", "Guohao Shen"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-05-10", "url": "https://arxiv.org/abs/2505.06601", "pdf_url": "https://arxiv.org/pdf/2505.06601v1", "arxiv_id": "2505.06601", "doi": "10.1145/3770854.3780316", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Knowledge Discovery and Data Mining", "quality_score": 0.1547}
{"id": "666b2c5f57a8c9ecbb18268f706a919ebc5f6e81a213b549b8ea8eb6e6e753ae", "sources": ["arxiv", "semantic_scholar"], "title": "DMRL: Data- and Model-aware Reward Learning for Data Extraction", "abstract": "Large language models (LLMs) are inherently vulnerable to unintended privacy breaches. Consequently, systematic red-teaming research is essential for developing robust defense mechanisms. However, current data extraction methods suffer from several limitations: (1) rely on dataset duplicates (addressable via deduplication), (2) depend on prompt engineering (now countered by detection and defense), and (3) rely on random-search adversarial generation. To address these challenges, we propose DMRL, a Data- and Model-aware Reward Learning approach for data extraction. This technique leverages inverse reinforcement learning to extract sensitive data from LLMs. Our method consists of two main components: (1) constructing an introspective reasoning dataset that captures leakage mindsets to guide model behavior, and (2) training reward models with Group Relative Policy Optimization (GRPO), dynamically tuning optimization based on task difficulty at both the data and model levels. Comprehensive experiments across various LLMs demonstrate that DMRL outperforms all baseline methods in data extraction performance.", "authors": ["Zhiqiang Wang", "Ruoxi Cheng"], "categories": ["cs.LG", "cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-07", "url": "https://arxiv.org/abs/2505.06284", "pdf_url": "https://arxiv.org/pdf/2505.06284v1", "arxiv_id": "2505.06284", "doi": "10.48550/arXiv.2505.06284", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1513}
{"id": "6e030a5fd3aa9ced9849c76019f2493a53b3f913b3bd6566a1aa72a632a75923", "sources": ["arxiv", "semantic_scholar"], "title": "Policy-labeled Preference Learning: Is Preference Enough for RLHF?", "abstract": "To design rewards that align with human goals, Reinforcement Learning from Human Feedback (RLHF) has emerged as a prominent technique for learning reward functions from human preferences and optimizing policies via reinforcement learning algorithms. However, existing RLHF methods often misinterpret trajectories as being generated by an optimal policy, causing inaccurate likelihood estimation and suboptimal learning. Inspired by Direct Preference Optimization framework which directly learns optimal policy without explicit reward, we propose policy-labeled preference learning (PPL), to resolve likelihood mismatch issues by modeling human preferences with regret, which reflects behavior policy information. We also provide a contrastive KL regularization, derived from regret-based principles, to enhance RLHF in sequential decision making. Experiments in high-dimensional continuous control tasks demonstrate PPL's significant improvements in offline RLHF performance and its effectiveness in online settings.", "authors": ["Taehyun Cho", "Seokhun Ju", "Seungyub Han", "Dohyeong Kim", "Kyungjae Lee", "Jungwoo Lee"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-06", "url": "https://arxiv.org/abs/2505.06273", "pdf_url": "https://arxiv.org/pdf/2505.06273v2", "arxiv_id": "2505.06273", "doi": "10.48550/arXiv.2505.06273", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1505}
{"id": "3c658dc8a481b46e4da551de8411b504d0dc36323c24d2c2612fd751b59162c0", "sources": ["arxiv", "semantic_scholar"], "title": "Sailing by the Stars: A Survey on Reward Models and Learning Strategies for Learning from Rewards", "abstract": "Recent developments in Large Language Models (LLMs) have shifted from pre-training scaling to post-training and test-time scaling. Across these developments, a key unified paradigm has arisen: Learning from Rewards, where reward signals act as the guiding stars to steer LLM behavior. It has underpinned a wide range of prevalent techniques, such as reinforcement learning (RLHF, RLAIF, DPO, and GRPO), reward-guided decoding, and post-hoc correction. Crucially, this paradigm enables the transition from passive learning from static data to active learning from dynamic feedback. This endows LLMs with aligned preferences and deep reasoning capabilities for diverse tasks. In this survey, we present a comprehensive overview of learning from rewards, from the perspective of reward models and learning strategies across training, inference, and post-inference stages. We further discuss the benchmarks for reward models and the primary applications. Finally we highlight the challenges and future directions. We maintain a paper collection at https://github.com/bobxwu/learning-from-rewards-llm-papers.", "authors": ["Xiaobao Wu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-05", "url": "https://arxiv.org/abs/2505.02686", "pdf_url": "https://arxiv.org/pdf/2505.02686v2", "arxiv_id": "2505.02686", "doi": null, "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/bobxwu/learning-from-rewards-llm-papers", "venue": null, "quality_score": 0.2258}
{"id": "647e2b573a55bc08bcb6e55eef68b9e7c8d35f3ecc19b634a916b994768c66fd", "sources": ["arxiv", "semantic_scholar"], "title": "R1-Reward: Training Multimodal Reward Model Through Stable Reinforcement Learning", "abstract": "Multimodal Reward Models (MRMs) play a crucial role in enhancing the performance of Multimodal Large Language Models (MLLMs). While recent advancements have primarily focused on improving the model structure and training data of MRMs, there has been limited exploration into the effectiveness of long-term reasoning capabilities for reward modeling and how to activate these capabilities in MRMs. In this paper, we explore how Reinforcement Learning (RL) can be used to improve reward modeling. Specifically, we reformulate the reward modeling problem as a rule-based RL task. However, we observe that directly applying existing RL algorithms, such as Reinforce++, to reward modeling often leads to training instability or even collapse due to the inherent limitations of these algorithms. To address this issue, we propose the StableReinforce algorithm, which refines the training loss, advantage estimation strategy, and reward design of existing RL methods. These refinements result in more stable training dynamics and superior performance. To facilitate MRM training, we collect 200K preference data from diverse datasets. Our reward model, R1-Reward, trained using the StableReinforce algorithm on this dataset, significantly improves performance on multimodal reward modeling benchmarks. Compared to previous SOTA models, R1-Reward achieves a $8.4\\%$ improvement on the VL Reward-Bench and a $14.3\\%$ improvement on the Multimodal Reward Bench. Moreover, with more inference compute, R1-Reward's performance is further enhanced, highlighting the potential of RL algorithms in optimizing MRMs.", "authors": ["Yi-Fan Zhang", "Xingyu Lu", "Xiao Hu", "Chaoyou Fu", "Bin Wen", "Tianke Zhang", "Changyi Liu", "Kaiyu Jiang", "Kaibing Chen", "Kaiyu Tang", "Haojie Ding", "Jiankang Chen", "Fan Yang", "Zhang Zhang", "Tingting Gao", "Liang Wang"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-05", "url": "https://arxiv.org/abs/2505.02835", "pdf_url": "https://arxiv.org/pdf/2505.02835v2", "arxiv_id": "2505.02835", "doi": "10.48550/arXiv.2505.02835", "citation_count": 60, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/yfzhang114/r1_reward", "venue": "arXiv.org", "quality_score": 0.4463}
{"id": "4112a6e5900033e5067279fa930bb6d6dc68b3b1fd85be38c58b82f149846151", "sources": ["arxiv", "semantic_scholar"], "title": "RM-R1: Reward Modeling as Reasoning", "abstract": "Reward modeling is essential for aligning large language models with human preferences through reinforcement learning. To provide accurate reward signals, a reward model (RM) should stimulate deep thinking and conduct interpretable reasoning before assigning a score or a judgment. Inspired by recent advances of long chain-of-thought on reasoning-intensive tasks, we hypothesize and validate that integrating reasoning into reward modeling significantly enhances RM's interpretability and performance. We introduce a new class of generative reward models, Reasoning Reward Models (ReasRMs), which formulate reward modeling as a reasoning task. We propose a reasoning-oriented training pipeline and train a family of ReasRMs, RM-R1. RM-R1 features a chain-of-rubrics (CoR) mechanism -- self-generating sample-level chat rubrics or math/code solutions, and evaluating candidate responses against them. The training of RM-R1 consists of two key stages: (1) distillation of high-quality reasoning chains and (2) reinforcement learning with verifiable rewards. Empirically, our models achieve superior performance across three reward model benchmarks on average, outperforming much larger open-weight models (e.g., INF-ORM-Llama3.1-70B) and proprietary ones (e.g., GPT-4o) by up to 4.9%. Beyond final performance, we perform thorough analyses to understand the key ingredients of successful ReasRM training.", "authors": ["Xiusi Chen", "Gaotang Li", "Ziqi Wang", "Bowen Jin", "Cheng Qian", "Yu Wang", "Hongru Wang", "Yu Zhang", "Denghui Zhang", "Tong Zhang", "Hanghang Tong", "Heng Ji"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-05-05", "url": "https://arxiv.org/abs/2505.02387", "pdf_url": "https://arxiv.org/pdf/2505.02387v4", "arxiv_id": "2505.02387", "doi": "10.48550/arXiv.2505.02387", "citation_count": 129, "influential_citation_count": 24, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.699}
{"id": "94e00a7af75a5865ec9b5cd109304dbad8efd77b2fd7718dd8e1d48ccfb13ea7", "sources": ["arxiv", "semantic_scholar"], "title": "CaRL: Learning Scalable Planning Policies with Simple Rewards", "abstract": "We investigate reinforcement learning (RL) for privileged planning in autonomous driving. State-of-the-art approaches for this task are rule-based, but these methods do not scale to the long tail. RL, on the other hand, is scalable and does not suffer from compounding errors like imitation learning. Contemporary RL approaches for driving use complex shaped rewards that sum multiple individual rewards, \\eg~progress, position, or orientation rewards. We show that PPO fails to optimize a popular version of these rewards when the mini-batch size is increased, which limits the scalability of these approaches. Instead, we propose a new reward design based primarily on optimizing a single intuitive reward term: route completion. Infractions are penalized by terminating the episode or multiplicatively reducing route completion. We find that PPO scales well with higher mini-batch sizes when trained with our simple reward, even improving performance. Training with large mini-batch sizes enables efficient scaling via distributed data parallelism. We scale PPO to 300M samples in CARLA and 500M samples in nuPlan with a single 8-GPU node. The resulting model achieves 64 DS on the CARLA longest6 v2 benchmark, outperforming other RL methods with more complex rewards by a large margin. Requiring only minimal adaptations from its use in CARLA, the same method is the best learning-based approach on nuPlan. It scores 91.3 in non-reactive and 90.6 in reactive traffic on the Val14 benchmark while being an order of magnitude faster than prior work.", "authors": ["Bernhard Jaeger", "Daniel Dauner", "Jens Beißwenger", "Simon Gerstenecker", "Kashyap Chitta", "Andreas Geiger"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-24", "url": "https://arxiv.org/abs/2504.17838", "pdf_url": "https://arxiv.org/pdf/2504.17838v3", "arxiv_id": "2504.17838", "doi": "10.48550/arXiv.2504.17838", "citation_count": 26, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3578}
{"id": "c276d6b2a94d6af40c70bf1aafdd1a083beb61d089c31c57794573fbdd8c0fc6", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Explainable Dense Reward Shapes via Bayesian Optimization", "abstract": "Current reinforcement learning from human feedback (RLHF) pipelines for large language model (LLM) alignment typically assign scalar rewards to sequences, using the final token as a surrogate indicator for the quality of the entire sequence. However, this leads to sparse feedback and suboptimal token-level credit assignment. In this work, we frame reward shaping as an optimization problem focused on token-level credit assignment. We propose a reward-shaping function leveraging explainability methods such as SHAP and LIME to estimate per-token rewards from the reward model. To learn parameters of this shaping function, we employ a bilevel optimization framework that integrates Bayesian Optimization and policy training to handle noise from the token reward estimates. Our experiments show that achieving a better balance of token-level reward attribution leads to performance improvements over baselines on downstream tasks and finds an optimal policy faster during training. Furthermore, we show theoretically that explainability methods that are feature additive attribution functions maintain the optimal policy as the original reward.", "authors": ["Ryan Koo", "Ian Yang", "Vipul Raheja", "Mingyi Hong", "Kwang-Sung Jun", "Dongyeop Kang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-22", "url": "https://arxiv.org/abs/2504.16272", "pdf_url": "https://arxiv.org/pdf/2504.16272v1", "arxiv_id": "2504.16272", "doi": "10.48550/arXiv.2504.16272", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1341}
{"id": "be2bf58c15f704843c7932ba5702c4a6fc5c21dcc5ad589fa915e7a49ce2a53e", "sources": ["arxiv", "semantic_scholar"], "title": "LoRe: Personalizing LLMs via Low-Rank Reward Modeling", "abstract": "Personalizing large language models (LLMs) to accommodate diverse user preferences is essential for enhancing alignment and user satisfaction. Traditional reinforcement learning from human feedback (RLHF) approaches often rely on monolithic value representations, limiting their ability to adapt to individual preferences. We introduce a novel framework that leverages low-rank preference modeling to efficiently learn and generalize user-specific reward functions. By representing reward functions in a low-dimensional subspace and modeling individual preferences as weighted combinations of shared basis functions, our approach avoids rigid user categorization while enabling scalability and few-shot adaptation. We validate our method on multiple preference datasets, demonstrating superior generalization to unseen users and improved accuracy in preference prediction tasks.", "authors": ["Avinandan Bose", "Zhihan Xiong", "Yuejie Chi", "Simon Shaolei Du", "Lin Xiao", "Maryam Fazel"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-20", "url": "https://arxiv.org/abs/2504.14439", "pdf_url": "https://arxiv.org/pdf/2504.14439v1", "arxiv_id": "2504.14439", "doi": "10.48550/arXiv.2504.14439", "citation_count": 24, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "3b218257623054ec67457c9ae710bc0e0449e57d1bc8a6b1c6176b3b2677f41c", "sources": ["arxiv", "semantic_scholar"], "title": "CHARM: Calibrating Reward Models With Chatbot Arena Scores", "abstract": "Reward models (RMs) play a crucial role in Reinforcement Learning from Human Feedback by serving as proxies for human preferences in aligning large language models. However, they suffer from various biases which could lead to reward hacking. In this paper, we identify a model preference bias in RMs, where they systematically assign disproportionately high scores to responses from certain policy models, leading to unfair judgments. To mitigate this bias, we propose a calibration method named CHatbot Arena calibrated Reward Modeling (CHARM) that leverages Elo scores from the Chatbot Arena to construct debiased preference datasets and adjust reward model scoring. We conduct extensive experiments on reward model benchmarks and human preference alignment. Results demonstrate that our calibrated RMs achieve improved evaluation accuracy on RM-Bench and the Chat-Hard domain of RewardBench, exhibit a stronger correlation with human preferences by producing scores more closely aligned with Elo rankings and improve downstream post-training performance. These results demonstrate that CHARM provides a simple, effective, and broadly applicable approach to building more reliable and fair reward models.", "authors": ["Xiao Zhu", "Chenmien Tan", "Pinzhen Chen", "Rico Sennrich", "Huiming Wang", "Yanlin Zhang", "Hanxu Hu"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-14", "url": "https://arxiv.org/abs/2504.10045", "pdf_url": "https://arxiv.org/pdf/2504.10045v2", "arxiv_id": "2504.10045", "doi": "10.48550/arXiv.2504.10045", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "23edfa7e4921483e120223b3fb315f50a55dec10bf5387ddd0180d9ff21be781", "sources": ["arxiv", "semantic_scholar"], "title": "FLoRA: Sample-Efficient Preference-based RL via Low-Rank Style Adaptation of Reward Functions", "abstract": "Preference-based reinforcement learning (PbRL) is a suitable approach for style adaptation of pre-trained robotic behavior: adapting the robot's policy to follow human user preferences while still being able to perform the original task. However, collecting preferences for the adaptation process in robotics is often challenging and time-consuming. In this work we explore the adaptation of pre-trained robots in the low-preference-data regime. We show that, in this regime, recent adaptation approaches suffer from catastrophic reward forgetting (CRF), where the updated reward model overfits to the new preferences, leading the agent to become unable to perform the original task. To mitigate CRF, we propose to enhance the original reward model with a small number of parameters (low-rank matrices) responsible for modeling the preference adaptation. Our evaluation shows that our method can efficiently and effectively adjust robotic behavior to human preferences across simulation benchmark tasks and multiple real-world robotic tasks.", "authors": ["Daniel Marta", "Simon Holk", "Miguel Vasco", "Jens Lundell", "Timon Homberger", "Finn Busch", "Olov Andersson", "Danica Kragic", "Iolanda Leite"], "categories": ["cs.RO", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-14", "url": "https://arxiv.org/abs/2504.10002", "pdf_url": "https://arxiv.org/pdf/2504.10002v1", "arxiv_id": "2504.10002", "doi": "10.1109/ICRA55743.2025.11127633", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Robotics and Automation", "quality_score": 0.1249}
{"id": "7500e5c17e13d3a53adac64ffcdc54de62b1b93e57148722c19896df552ab46e", "sources": ["arxiv", "semantic_scholar"], "title": "Information-Theoretic Reward Decomposition for Generalizable RLHF", "abstract": "A generalizable reward model is crucial in Reinforcement Learning from Human Feedback (RLHF) as it enables correctly evaluating unseen prompt-response pairs. However, existing reward models lack this ability, as they are typically trained by increasing the reward gap between chosen and rejected responses, while overlooking the prompts that the responses are conditioned on. Consequently, when the trained reward model is evaluated on prompt-response pairs that lie outside the data distribution, neglecting the effect of prompts may result in poor generalization of the reward model. To address this issue, we decompose the reward value into two independent components: prompt-free reward and prompt-related reward. Prompt-free reward represents the evaluation that is determined only by responses, while the prompt-related reward reflects the reward that derives from both the prompt and the response. We extract these two components from an information-theoretic perspective, which requires no extra models. Subsequently, we propose a new reward learning algorithm by prioritizing data samples based on their prompt-free reward values. Through toy examples, we demonstrate that the extracted prompt-free and prompt-related rewards effectively characterize two parts of the reward model. Further, standard evaluations show that our method improves both the alignment performance and the generalization capability of the reward model.", "authors": ["Liyuan Mao", "Haoran Xu", "Amy Zhang", "Weinan Zhang", "Chenjia Bai"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-08", "url": "https://arxiv.org/abs/2504.06020", "pdf_url": "https://arxiv.org/pdf/2504.06020v2", "arxiv_id": "2504.06020", "doi": "10.48550/arXiv.2504.06020", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "c89b51520214f8e5f022b8b095e83c12f4b1dfd83457ea6e687d15bb62df98b5", "sources": ["arxiv", "semantic_scholar"], "title": "Adversarial Training of Reward Models", "abstract": "Reward modeling has emerged as a promising approach for the scalable alignment of language models. However, contemporary reward models (RMs) often lack robustness, awarding high rewards to low-quality, out-of-distribution (OOD) samples. This can lead to reward hacking, where policies exploit unintended shortcuts to maximize rewards, undermining alignment. To address this challenge, we introduce Adv-RM, a novel adversarial training framework that automatically identifies adversarial examples -- responses that receive high rewards from the target RM but are OOD and of low quality. By leveraging reinforcement learning, Adv-RM trains a policy to generate adversarial examples that reliably expose vulnerabilities in large state-of-the-art reward models such as Nemotron 340B RM. Incorporating these adversarial examples into the reward training process improves the robustness of RMs, mitigating reward hacking and enhancing downstream performance in RLHF. We demonstrate that Adv-RM significantly outperforms conventional RM training, increasing stability and enabling more effective RLHF training in both synthetic and real-data settings.", "authors": ["Alexander Bukharin", "Haifeng Qian", "Shengyang Sun", "Adithya Renduchintala", "Soumye Singhal", "Zhilin Wang", "Oleksii Kuchaiev", "Olivier Delalleau", "Tuo Zhao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-08", "url": "https://arxiv.org/abs/2504.06141", "pdf_url": "https://arxiv.org/pdf/2504.06141v2", "arxiv_id": "2504.06141", "doi": "10.48550/arXiv.2504.06141", "citation_count": 13, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2865}
{"id": "b5f392da3195630890228649bb006c9b911f99958a5359c4217b7a6b7c7f954c", "sources": ["arxiv", "semantic_scholar"], "title": "A Unified Pairwise Framework for RLHF: Bridging Generative Reward Modeling and Policy Optimization", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has emerged as a important paradigm for aligning large language models (LLMs) with human preferences during post-training. This framework typically involves two stages: first, training a reward model on human preference data, followed by optimizing the language model using reinforcement learning algorithms. However, current RLHF approaches may constrained by two limitations. First, existing RLHF frameworks often rely on Bradley-Terry models to assign scalar rewards based on pairwise comparisons of individual responses. However, this approach imposes significant challenges on reward model (RM), as the inherent variability in prompt-response pairs across different contexts demands robust calibration capabilities from the RM. Second, reward models are typically initialized from generative foundation models, such as pre-trained or supervised fine-tuned models, despite the fact that reward models perform discriminative tasks, creating a mismatch. This paper introduces Pairwise-RL, a RLHF framework that addresses these challenges through a combination of generative reward modeling and a pairwise proximal policy optimization (PPO) algorithm. Pairwise-RL unifies reward model training and its application during reinforcement learning within a consistent pairwise paradigm, leveraging generative modeling techniques to enhance reward model performance and score calibration. Experimental evaluations demonstrate that Pairwise-RL outperforms traditional RLHF frameworks across both internal evaluation datasets and standard public benchmarks, underscoring its effectiveness in improving alignment and model behavior.", "authors": ["Wenyuan Xu", "Xiaochen Zuo", "Chao Xin", "Yu Yue", "Lin Yan", "Yonghui Wu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-07", "url": "https://arxiv.org/abs/2504.04950", "pdf_url": "https://arxiv.org/pdf/2504.04950v1", "arxiv_id": "2504.04950", "doi": "10.48550/arXiv.2504.04950", "citation_count": 20, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3306}
{"id": "9b31d18a84bc4511fa9b579531827f8d4a998552f229b4fac0cf4b84441a86ee", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Generation via Large Vision-Language Model in Offline Reinforcement Learning", "abstract": "In offline reinforcement learning (RL), learning from fixed datasets presents a promising solution for domains where real-time interaction with the environment is expensive or risky. However, designing dense reward signals for offline dataset requires significant human effort and domain expertise. Reinforcement learning with human feedback (RLHF) has emerged as an alternative, but it remains costly due to the human-in-the-loop process, prompting interest in automated reward generation models. To address this, we propose Reward Generation via Large Vision-Language Models (RG-VLM), which leverages the reasoning capabilities of LVLMs to generate rewards from offline data without human involvement. RG-VLM improves generalization in long-horizon tasks and can be seamlessly integrated with the sparse reward signals to enhance task performance, demonstrating its potential as an auxiliary reward signal.", "authors": ["Younghwan Lee", "Tung M. Luu", "Donghoon Lee", "Chang D. Yoo"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-04-03", "url": "https://arxiv.org/abs/2504.08772", "pdf_url": "https://arxiv.org/pdf/2504.08772v1", "arxiv_id": "2504.08772", "doi": "10.1109/ICASSP49660.2025.10889042", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Acoustics, Speech, and Signal Processing", "quality_score": 0.1505}
{"id": "45d08ff3375d1b91919777a58de6d5577947e0c1a3c487f43aeb2e460edc838c", "sources": ["arxiv", "semantic_scholar"], "title": "Probabilistic Uncertain Reward Model", "abstract": "Reinforcement learning from human feedback (RLHF) is a critical technique for training large language models. However, conventional reward models based on the Bradley-Terry model (BTRM) often suffer from overconfidence when faced with inconsistent labels or out-of-distribution samples, leading to reward hacking, where the policy model blindly optimizes for proxy rewards while degrading true performance. This paper proposes the Probabilistic Uncertain Reward Model (PURM), which generalizes the Bradley-Terry model to learn the reward distributions that emerged from the preference data. We theoretically derive the loss function of PURM and introduce a novel method that uses the overlap between distributions to quantify uncertainty. Empirical results show that PURM outperforms existing methods with more accurate reward and sound uncertainty estimations, and sustains effective learning for more optimization steps and obtain higher maximum win rate in RLHF. The data and code of this paper are released at https://anonymous.4open.science/r/Probabilistic-Uncertain-Reward-Model/", "authors": ["Wangtao Sun", "Xiang Cheng", "Xing Yu", "Haotian Xu", "Zhao Yang", "Shizhu He", "Jun Zhao", "Kang Liu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-28", "url": "https://arxiv.org/abs/2503.22480", "pdf_url": "https://arxiv.org/pdf/2503.22480v6", "arxiv_id": "2503.22480", "doi": "10.48550/arXiv.2503.22480", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1054}
{"id": "44ad06fffad9ac7d2af1123140133961da71de87819ad8d2dfde121c6d78f789", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Design for Reinforcement Learning Agents", "abstract": "Reward functions are central in reinforcement learning (RL), guiding agents towards optimal decision-making. The complexity of RL tasks requires meticulously designed reward functions that effectively drive learning while avoiding unintended consequences. Effective reward design aims to provide signals that accelerate the agent's convergence to optimal behavior. Crafting rewards that align with task objectives, foster desired behaviors, and prevent undesirable actions is inherently challenging. This thesis delves into the critical role of reward signals in RL, highlighting their impact on the agent's behavior and learning dynamics and addressing challenges such as delayed, ambiguous, or intricate rewards. In this thesis work, we tackle different aspects of reward shaping. First, we address the problem of designing informative and interpretable reward signals from a teacher's/expert's perspective (teacher-driven). Here, the expert, equipped with the optimal policy and the corresponding value function, designs reward signals that expedite the agent's convergence to optimal behavior. Second, we build on this teacher-driven approach by introducing a novel method for adaptive interpretable reward design. In this scenario, the expert tailors the rewards based on the learner's current policy, ensuring alignment and optimal progression. Third, we propose a meta-learning approach, enabling the agent to self-design its reward signals online without expert input (agent-driven). This self-driven method considers the agent's learning and exploration to establish a self-improving feedback loop.", "authors": ["Rati Devidze"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-27", "url": "https://arxiv.org/abs/2503.21949", "pdf_url": "https://arxiv.org/pdf/2503.21949v1", "arxiv_id": "2503.21949", "doi": "10.48550/arXiv.2503.21949", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "d971ae1199ec2e89cd83e2da45084a13c1f73d0ec8fd59b4f47d7cc050bd46a7", "sources": ["arxiv", "semantic_scholar"], "title": "ViLBench: A Suite for Vision-Language Process Reward Modeling", "abstract": "Process-supervised reward models serve as a fine-grained function that provides detailed step-wise feedback to model responses, facilitating effective selection of reasoning trajectories for complex tasks. Despite its advantages, evaluation on PRMs remains less explored, especially in the multimodal domain. To address this gap, this paper first benchmarks current vision large language models (VLLMs) as two types of reward models: output reward models (ORMs) and process reward models (PRMs) on multiple vision-language benchmarks, which reveal that neither ORM nor PRM consistently outperforms across all tasks, and superior VLLMs do not necessarily yield better rewarding performance. To further advance evaluation, we introduce ViLBench, a vision-language benchmark designed to require intensive process reward signals. Notably, OpenAI's GPT-4o with Chain-of-Thought (CoT) achieves only 27.3% accuracy, indicating the benchmark's challenge for current VLLMs. Lastly, we preliminarily showcase a promising pathway towards bridging the gap between general VLLMs and reward models -- by collecting 73.6K vision-language process reward data using an enhanced tree-search algorithm, our 3B model is able to achieve an average improvement of 3.3% over standard CoT and up to 2.5% compared to its untrained counterpart on ViLBench by selecting OpenAI o1's generations. We release the implementations at https://ucsc-vlaa.github.io/ViLBench with our code, model, and data.", "authors": ["Haoqin Tu", "Weitao Feng", "Hardy Chen", "Hui Liu", "Xianfeng Tang", "Cihang Xie"], "categories": ["cs.CV", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-26", "url": "https://arxiv.org/abs/2503.20271", "pdf_url": "https://arxiv.org/pdf/2503.20271v1", "arxiv_id": "2503.20271", "doi": "10.48550/arXiv.2503.20271", "citation_count": 19, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3253}
{"id": "a80334dc8886133768d587f17fed54544c0b28e9624004e414ee7182900f7e4f", "sources": ["arxiv", "semantic_scholar"], "title": "Mitigating Reward Over-Optimization in RLHF via Behavior-Supported Regularization", "abstract": "Reinforcement learning from human feedback (RLHF) is an effective method for aligning large language models (LLMs) with human values. However, reward over-optimization remains an open challenge leading to discrepancies between the performance of LLMs under the reward model and the true human objectives. A primary contributor to reward over-optimization is the extrapolation error that arises when the reward model evaluates out-of-distribution (OOD) responses. However, current methods still fail to prevent the increasing frequency of OOD response generation during the reinforcement learning (RL) process and are not effective at handling extrapolation errors from OOD responses. In this work, we propose the Behavior-Supported Policy Optimization (BSPO) method to mitigate the reward over-optimization issue. Specifically, we define behavior policy as the next token distribution of the reward training dataset to model the in-distribution (ID) region of the reward model. Building on this, we introduce the behavior-supported Bellman operator to regularize the value function, penalizing all OOD values without impacting the ID ones. Consequently, BSPO reduces the generation of OOD responses during the RL process, thereby avoiding overestimation caused by the reward model's extrapolation errors. Theoretically, we prove that BSPO guarantees a monotonic improvement of the supported policy until convergence to the optimal behavior-supported policy. Empirical results from extensive experiments show that BSPO outperforms baselines in preventing reward over-optimization due to OOD evaluation and finding the optimal ID policy.", "authors": ["Juntao Dai", "Taiye Chen", "Yaodong Yang", "Qian Zheng", "Gang Pan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-23", "url": "https://arxiv.org/abs/2503.18130", "pdf_url": "https://arxiv.org/pdf/2503.18130v1", "arxiv_id": "2503.18130", "doi": "10.48550/arXiv.2503.18130", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2603}
{"id": "ccda508663d7898f7a49fde81ccd7f3b2217d75195d5ac3f34cba0f1a5c3d632", "sources": ["arxiv", "semantic_scholar"], "title": "Capturing Individual Human Preferences with Reward Features", "abstract": "Reinforcement learning from human feedback usually models preferences using a reward function that does not distinguish between people. We argue that this is unlikely to be a good design choice in contexts with high potential for disagreement, like in the training of large language models. We formalise and analyse the problem of learning a reward model that can be specialised to a user. Using the principle of empirical risk minimisation, we derive a probably approximately correct (PAC) bound showing the dependency of the approximation error on the number of training examples, as usual, and also on the number of human raters who provided feedback on them. Based on our theoretical findings, we discuss how to best collect pairwise preference data and argue that adaptive reward models should be beneficial when there is considerable disagreement among users. We also propose a concrete architecture for an adaptive reward model. Our approach leverages the observation that individual preferences can be captured as a linear combination of a set of general reward features. We show how to learn such features and subsequently use them to quickly adapt the reward model to a specific individual, even if their preferences are not reflected in the training data. We present experiments with large language models illustrating our theoretical results and comparing the proposed architecture with a non-adaptive baseline. Consistent with our analysis, the benefits provided by our model increase with the number of raters and the heterogeneity of their preferences. We also show that our model compares favourably to adaptive counterparts, including those performing in-context personalisation.", "authors": ["André Barreto", "Vincent Dumoulin", "Yiran Mao", "Mark Rowland", "Nicolas Perez-Nieves", "Bobak Shahriari", "Yann Dauphin", "Doina Precup", "Hugo Larochelle"], "categories": ["cs.AI", "cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-03-21", "url": "https://arxiv.org/abs/2503.17338", "pdf_url": "https://arxiv.org/pdf/2503.17338v2", "arxiv_id": "2503.17338", "doi": "10.48550/arXiv.2503.17338", "citation_count": 11, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "f0b43246e955aca73fd63ed015c14d8290c37be52d543ece2081e812b5c92c8d", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Redistribution via Gaussian Process Likelihood Estimation", "abstract": "In many practical reinforcement learning tasks, feedback is only provided at the end of a long horizon, leading to sparse and delayed rewards. Existing reward redistribution methods typically assume that per-step rewards are independent, thus overlooking interdependencies among state-action pairs. In this paper, we propose a Gaussian process based Likelihood Reward Redistribution (GP-LRR) framework that addresses this issue by modeling the reward function as a sample from a Gaussian process, which explicitly captures dependencies between state-action pairs through the kernel function. By maximizing the likelihood of the observed episodic return via a leave-one-out strategy that leverages the entire trajectory, our framework inherently introduces uncertainty regularization. Moreover, we show that conventional mean-squared-error (MSE) based reward redistribution arises as a special case of our GP-LRR framework when using a degenerate kernel without observation noise. When integrated with an off-policy algorithm such as Soft Actor-Critic, GP-LRR yields dense and informative reward signals, resulting in superior sample efficiency and policy performance on several MuJoCo benchmarks.", "authors": ["Minheng Xiao", "Xian Yu"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-20", "url": "https://arxiv.org/abs/2503.17409", "pdf_url": "https://arxiv.org/pdf/2503.17409v2", "arxiv_id": "2503.17409", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0612}
{"id": "87061b5f8f9c20d43470a9a2b6a78f22ac05261df71bdf913d95e5feb9b5a99c", "sources": ["arxiv", "semantic_scholar"], "title": "What Makes a Reward Model a Good Teacher? An Optimization Perspective", "abstract": "The success of Reinforcement Learning from Human Feedback (RLHF) critically depends on the quality of the reward model. However, while this quality is primarily evaluated through accuracy, it remains unclear whether accuracy fully captures what makes a reward model an effective teacher. We address this question from an optimization perspective. First, we prove that regardless of how accurate a reward model is, if it induces low reward variance, then the RLHF objective suffers from a flat landscape. Consequently, even a perfectly accurate reward model can lead to extremely slow optimization, underperforming less accurate models that induce higher reward variance. We additionally show that a reward model that works well for one language model can induce low reward variance, and thus a flat objective landscape, for another. These results establish a fundamental limitation of evaluating reward models solely based on accuracy or independently of the language model they guide. Experiments using models of up to 8B parameters corroborate our theory, demonstrating the interplay between reward variance, accuracy, and reward maximization rate. Overall, our findings highlight that beyond accuracy, a reward model needs to induce sufficient variance for efficient optimization.", "authors": ["Noam Razin", "Zixuan Wang", "Hubert Strauss", "Stanley Wei", "Jason D. Lee", "Sanjeev Arora"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-03-19", "url": "https://arxiv.org/abs/2503.15477", "pdf_url": "https://arxiv.org/pdf/2503.15477v4", "arxiv_id": "2503.15477", "doi": "10.48550/arXiv.2503.15477", "citation_count": 64, "influential_citation_count": 8, "has_code": true, "code_url": "https://github.com/princeton-pli/what-makes-good-rm", "venue": "arXiv.org", "quality_score": 0.4771}
{"id": "ee30f7d9c4c841ffd3657072f5d8157dc8cdb493dd373ea97b4ce6f858e2b3b2", "sources": ["arxiv", "semantic_scholar"], "title": "Provably Efficient Reward Transfer in Reinforcement Learning with Discrete Markov Decision Processes", "abstract": "In this paper, we propose a new solution to reward adaptation (RA) in reinforcement learning, where the agent adapts to a target reward function based on one or more existing source behaviors learned a priori under the same domain dynamics but different reward functions. While learning the target behavior from scratch is possible, it is often inefficient given the available source behaviors. Our work introduces a new approach to RA through the manipulation of Q-functions. Assuming the target reward function is a known function of the source reward functions, we compute bounds on the Q-function and present an iterative process (akin to value iteration) to tighten these bounds. Such bounds enable action pruning in the target domain before learning even starts. We refer to this method as \"Q-Manipulation\" (Q-M). The iteration process assumes access to a lite-model, which is easy to provide or learn. We formally prove that Q-M, under discrete domains, does not affect the optimality of the returned policy and show that it is provably efficient in terms of sample complexity in a probabilistic sense. Q-M is evaluated in a variety of synthetic and simulation domains to demonstrate its effectiveness, generalizability, and practicality.", "authors": ["Kevin Vora", "Yu Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-17", "url": "https://arxiv.org/abs/2503.13414", "pdf_url": "https://arxiv.org/pdf/2503.13414v3", "arxiv_id": "2503.13414", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "548ee58eb4a0c184cd4d2255620a3b14d571d5f1736dc1810201e41ba1a26995", "sources": ["arxiv", "semantic_scholar"], "title": "From Demonstrations to Rewards: Alignment Without Explicit Human Preferences", "abstract": "One of the challenges of aligning large models with human preferences lies in both the data requirements and the technical complexities of current approaches. Predominant methods, such as RLHF, involve multiple steps, each demanding distinct types of data, including demonstration data and preference data. In RLHF, human preferences are typically modeled through a reward model, which serves as a proxy to guide policy learning during the reinforcement learning stage, ultimately producing a policy aligned with human preferences. However, in this paper, we propose a fresh perspective on learning alignment based on inverse reinforcement learning principles, where the optimal policy is still derived from reward maximization. However, instead of relying on preference data, we directly learn the reward model from demonstration data. This new formulation offers the flexibility to be applied even when only demonstration data is available, a capability that current RLHF methods lack, and it also shows that demonstration data offers more utility than what conventional wisdom suggests. Our extensive evaluation, based on public reward benchmark, HuggingFace Open LLM Leaderboard and MT-Bench, demonstrates that our approach compares favorably to state-of-the-art methods that rely solely on demonstration data.", "authors": ["Siliang Zeng", "Yao Liu", "Huzefa Rangwala", "George Karypis", "Mingyi Hong", "Rasool Fakoor"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-15", "url": "https://arxiv.org/abs/2503.13538", "pdf_url": "https://arxiv.org/pdf/2503.13538v1", "arxiv_id": "2503.13538", "doi": "10.48550/arXiv.2503.13538", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "859549dd753d63b73b5a5b798bfe0562a21a2e9b81f54f431907168d25782e88", "sources": ["arxiv", "semantic_scholar"], "title": "Right Reward Right Time for Federated Learning", "abstract": "Critical learning periods (CLPs) in federated learning (FL) refer to early stages during which low-quality contributions (e.g., sparse training data availability) can permanently impair the performance of the global model owned by the cloud server. However, existing incentive mechanisms typically assume temporal homogeneity, treating all training rounds as equally important, thereby failing to prioritize and attract high-quality contributions during CLPs. This inefficiency is compounded by information asymmetry due to privacy regulations, where the cloud lacks knowledge of client training capabilities, leading to adverse selection and moral hazard. Thus, in this article, we propose a time-aware contract-theoretic incentive framework, named Right Reward Right Time (R3T), to encourage client involvement, especially during CLPs, to maximize the utility of the cloud server. We formulate a cloud utility function that captures the trade-off between the achieved model performance and rewards allocated for clients' contributions, explicitly accounting for client heterogeneity in time and system capabilities, effort, and joining time. Then, we devise a CLP-aware incentive mechanism deriving an optimal contract design that satisfies individual rationality, incentive compatibility, and budget feasibility constraints, motivating rational clients to participate early and contribute efforts. By providing the right reward at the right time, our approach can attract the highest-quality contributions during CLPs. Simulation and proof-of-concept studies show that R3T mitigates information asymmetry, increases cloud utility, and yields superior economic efficiency compared to conventional incentive mechanisms. Our proof-of-concept results demonstrate up to a 47.6% reduction in the total number of clients and up to a 300% improvement in convergence time while achieving competitive test accuracy.", "authors": ["Thanh Linh Nguyen", "Dinh Thai Hoang", "Diep N. Nguyen", "Quoc-Viet Pham"], "categories": ["cs.LG", "cs.AI", "cs.DC", "cs.GT"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-10", "url": "https://arxiv.org/abs/2503.07869", "pdf_url": "https://arxiv.org/pdf/2503.07869v3", "arxiv_id": "2503.07869", "doi": "10.48550/arXiv.2503.07869", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "ecf9b9cdee2c1b17d877580efd168cd9dc614f7a372387773a7a071b608f8f5b", "sources": ["arxiv", "semantic_scholar"], "title": "Language Model Personalization via Reward Factorization", "abstract": "Modern large language models (LLMs) are optimized for human-aligned responses using Reinforcement Learning from Human Feedback (RLHF). However, existing RLHF approaches assume a universal preference model and fail to account for individual user preferences, limiting their effectiveness in personalized applications. We introduce a framework that extends RLHF to enable user personalization by leveraging the assumption that user preferences lie in a low-dimensional space. Instead of training a separate model per user, we represent user-specific rewards as a linear combination of base reward functions. Using only ~10 user responses, our method can infer user-specific rewards and align LLM outputs accordingly. We validate our approach through experiments with both synthetic and real users, demonstrating significant personalization achieved by our method. In human evaluations, our method achieves a 67% win rate over default GPT-4o responses.", "authors": ["Idan Shenfeld", "Felix Faltings", "Pulkit Agrawal", "Aldo Pacchiano"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-08", "url": "https://arxiv.org/abs/2503.06358", "pdf_url": "https://arxiv.org/pdf/2503.06358v1", "arxiv_id": "2503.06358", "doi": "10.48550/arXiv.2503.06358", "citation_count": 23, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3495}
{"id": "91da2ec8eff57c25f9754497ab772339ab0dab5858165d4df843389d745955ea", "sources": ["arxiv", "semantic_scholar"], "title": "Implicit Cross-Lingual Rewarding for Efficient Multilingual Preference Alignment", "abstract": "Direct Preference Optimization (DPO) has become a prominent method for aligning Large Language Models (LLMs) with human preferences. While DPO has enabled significant progress in aligning English LLMs, multilingual preference alignment is hampered by data scarcity. To address this, we propose a novel approach that $\\textit{captures}$ learned preferences from well-aligned English models by implicit rewards and $\\textit{transfers}$ them to other languages through iterative training. Specifically, we derive an implicit reward model from the logits of an English DPO-aligned model and its corresponding reference model. This reward model is then leveraged to annotate preference relations in cross-lingual instruction-following pairs, using English instructions to evaluate multilingual responses. The annotated data is subsequently used for multilingual DPO fine-tuning, facilitating preference knowledge transfer from English to other languages. Fine-tuning Llama3 for two iterations resulted in a 12.72% average improvement in Win Rate and a 5.97% increase in Length Control Win Rate across all training languages on the X-AlpacaEval leaderboard. Our findings demonstrate that leveraging existing English-aligned models can enable efficient and effective multilingual preference alignment, significantly reducing the need for extensive multilingual preference data. The code is available at https://github.com/ZNLP/Implicit-Cross-Lingual-Rewarding", "authors": ["Wen Yang", "Junhong Wu", "Chen Wang", "Chengqing Zong", "Jiajun Zhang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-06", "url": "https://arxiv.org/abs/2503.04647", "pdf_url": "https://arxiv.org/pdf/2503.04647v2", "arxiv_id": "2503.04647", "doi": "10.48550/arXiv.2503.04647", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/ZNLP/Implicit-Cross-Lingual-Rewarding", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.2113}
{"id": "04b8859e44d134c56a5b30b9c2a59ea8c10f08797bbde2cd069c5b2210d130e9", "sources": ["arxiv", "semantic_scholar"], "title": "Diffusion Classifier-Driven Reward for Offline Preference-based Reinforcement Learning", "abstract": "Offline preference-based reinforcement learning (PbRL) mitigates the need for reward definition, aligning with human preferences via preference-driven reward feedback without interacting with the environment. However, trajectory-wise preference labels are difficult to meet the precise learning of step-wise reward, thereby affecting the performance of downstream algorithms. To alleviate the insufficient step-wise reward caused by trajectory-wise preferences, we propose a novel preference-based reward acquisition method: Diffusion Preference-based Reward (DPR). DPR directly treats step-wise preference-based reward acquisition as a binary classification and utilizes the robustness of diffusion classifiers to infer step-wise rewards discriminatively. In addition, to further utilize trajectory-wise preference information, we propose Conditional Diffusion Preference-based Reward (C-DPR), which conditions on trajectory-wise preference labels to enhance reward inference. We apply the above methods to existing offline RL algorithms, and a series of experimental results demonstrate that the diffusion classifier-driven reward outperforms the previous reward acquisition method with the Bradley-Terry model.", "authors": ["Teng Pang", "Bingzheng Wang", "Guoqiang Wu", "Yilong Yin"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-03", "url": "https://arxiv.org/abs/2503.01143", "pdf_url": "https://arxiv.org/pdf/2503.01143v3", "arxiv_id": "2503.01143", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0489}
{"id": "364037cda029563f86122dc266f4dcdb67e2adb4c8b1b9313200ef0c8205cd60", "sources": ["arxiv", "semantic_scholar"], "title": "Sentence-level Reward Model can Generalize Better for Aligning LLM from Human Preference", "abstract": "Learning reward models from human preference datasets and subsequently optimizing language models via reinforcement learning has emerged as a fundamental paradigm for aligning LLMs with human preferences. The performance of the reward model plays a crucial role in the effectiveness of alignment. Previous reward models operate at a coarse-grained level, requiring the generation of a complete response to obtain a reward value. The sparse reward may present challenges for downstream reinforcement learning. While recent efforts have attempted to learn token-level reward models, the lack of explicit semantic information makes it difficult to model the credit of every individual token. In this paper, we propose assigning scores to every sentence, introducing an intermediate-grained reward model. By segmenting the complete response into sentences and applying differential operations to reward output at the start and end positions of each sentence, we can effectively model the rewards of sentences. Moreover, a novel attention mechanism is introduced to aggregate the scores of all sentences into a response-level score, which allows it to be trained using the Bradley-Terry model. On common benchmarks, our method outperforms the response-level reward model by 2.7% on RewardBench (for reward modeling evaluation) and surpasses all baselines on AlpacaEval (for alignment evaluation).", "authors": ["Wenjie Qiu", "Yi-Chen Li", "Xuqin Zhang", "Tianyi Zhang", "Yihang Zhang", "Zongzhang Zhang", "Yang Yu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-03-01", "url": "https://arxiv.org/abs/2503.04793", "pdf_url": "https://arxiv.org/pdf/2503.04793v4", "arxiv_id": "2503.04793", "doi": "10.48550/arXiv.2503.04793", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "7a3f222ec9f4669af8212aee7125b5a05f40b366a397719706fb868c9b729e9c", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Learning from Multiple Feedback Types", "abstract": "Learning rewards from preference feedback has become an important tool in the alignment of agentic models. Preference-based feedback, often implemented as a binary comparison between multiple completions, is an established method to acquire large-scale human feedback. However, human feedback in other contexts is often much more diverse. Such diverse feedback can better support the goals of a human annotator, and the simultaneous use of multiple sources might be mutually informative for the learning process or carry type-dependent biases for the reward learning process. Despite these potential benefits, learning from different feedback types has yet to be explored extensively. In this paper, we bridge this gap by enabling experimentation and evaluating multi-type feedback in a broad set of environments. We present a process to generate high-quality simulated feedback of six different types. Then, we implement reward models and downstream RL training for all six feedback types. Based on the simulated feedback, we investigate the use of types of feedback across ten RL environments and compare them to pure preference-based baselines. We show empirically that diverse types of feedback can be utilized and lead to strong reward modeling performance. This work is the first strong indicator of the potential of multi-type feedback for RLHF.", "authors": ["Yannick Metz", "András Geiszl", "Raphaël Baur", "Mennatallah El-Assady"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-28", "url": "https://arxiv.org/abs/2502.21038", "pdf_url": "https://arxiv.org/pdf/2502.21038v1", "arxiv_id": "2502.21038", "doi": "10.48550/arXiv.2502.21038", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2258}
{"id": "2db0ec7c2ab39bd180362a9a7937a17b0466d559caa39240ab823ef2d13dca27", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Shaping to Mitigate Reward Hacking in RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is essential for aligning large language models (LLMs) with human values. However, RLHF is susceptible to \\emph{reward hacking}, where the agent exploits flaws in the reward function rather than learning the intended behavior, thus degrading alignment. Although reward shaping helps stabilize RLHF and partially mitigate reward hacking, a systematic investigation into shaping techniques and their underlying principles remains lacking. To bridge this gap, we present a comprehensive study of the prevalent reward shaping methods. Our analysis suggests two key design principles: (1) the RL reward should be bounded, and (2) the RL reward benefits from rapid initial growth followed by gradual convergence. Guided by these insights, we propose Preference As Reward (PAR), a novel approach that leverages the latent preferences embedded within the reward model as the signal for reinforcement learning. Moreover, PAR exhibits two critical variance-reduction properties that contribute to stabilizing the RLHF training process and effectively extending the tolerance window for early stopping. We evaluated PAR on the base model Gemma2-2B using two datasets, Ultrafeedback-Binarized and HH-RLHF. Experimental results demonstrate PAR's superior performance over other reward shaping methods. On the AlpacaEval 2.0 benchmark, PAR achieves a win rate of at least 5 percentage points higher than competing approaches. Furthermore, PAR exhibits remarkable data efficiency, requiring only a single reference reward for optimal performance, and maintains robustness against reward hacking even after two full epochs of training. The code is available at https://github.com/PorUna-byte/PAR.", "authors": ["Jiayi Fu", "Xuandong Zhao", "Chengyuan Yao", "Heng Wang", "Qi Han", "Yanghua Xiao"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-26", "url": "https://arxiv.org/abs/2502.18770", "pdf_url": "https://arxiv.org/pdf/2502.18770v5", "arxiv_id": "2502.18770", "doi": "10.48550/arXiv.2502.18770", "citation_count": 78, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/PorUna-byte/PAR", "venue": "arXiv.org", "quality_score": 0.4744}
{"id": "503deaa5c3620362933f180c29e6c7911be811104e328a301dac5e3411863e3b", "sources": ["arxiv", "semantic_scholar"], "title": "Can RLHF be More Efficient with Imperfect Reward Models? A Policy Coverage Perspective", "abstract": "Sample efficiency is critical for online Reinforcement Learning from Human Feedback (RLHF). While existing works investigate sample-efficient online exploration strategies, the potential of utilizing misspecified yet relevant reward models to accelerate learning remains underexplored. This paper studies how to transfer knowledge from those imperfect reward models in online RLHF. We start by identifying a novel property due to KL-regularization in the RLHF objective: \\emph{a policy's coverability of the optimal policy is captured by its sub-optimality}. Building on this insight, we propose novel transfer learning principles and a theoretical algorithm -- \\emph{\\textbf{T}ransfer \\textbf{P}olicy \\textbf{O}ptimization (\\textbf{TPO})} -- with provable benefits compared to standard online learning. Empirically, inspired by our theoretical findings, we develop a win-rate-based transfer policy selection strategy with improved computational efficiency. Moreover, our empirical transfer learning technique is modular and can be integrated with various policy optimization methods, such as DPO, IPO and XPO, to further enhance their performance. We validate the effectiveness of our method through experiments on summarization tasks.", "authors": ["Jiawei Huang", "Bingcong Li", "Christoph Dann", "Niao He"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-02-26", "url": "https://arxiv.org/abs/2502.19255", "pdf_url": "https://arxiv.org/pdf/2502.19255v3", "arxiv_id": "2502.19255", "doi": "10.48550/arXiv.2502.19255", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1747}
{"id": "44c02d9d0d7891d15271536a56ca7bde2f7aaeceedd5a8e104a9fdb348863b60", "sources": ["arxiv", "semantic_scholar"], "title": "Agentic Reward Modeling: Integrating Human Preferences with Verifiable Correctness Signals for Reliable Reward Systems", "abstract": "Reward models (RMs) are crucial for the training and inference-time scaling up of large language models (LLMs). However, existing reward models primarily focus on human preferences, neglecting verifiable correctness signals which have shown strong potential in training LLMs. In this paper, we propose agentic reward modeling, a reward system that combines reward models with verifiable correctness signals from different aspects to provide reliable rewards. We empirically implement a reward agent, named RewardAgent, that combines human preference rewards with two verifiable signals: factuality and instruction following, to provide more reliable rewards. We conduct comprehensive experiments on existing reward model benchmarks and inference time best-of-n searches on real-world downstream tasks. RewardAgent significantly outperforms vanilla reward models, demonstrating its effectiveness. We further construct training preference pairs using RewardAgent and train an LLM with the DPO objective, achieving superior performance on various NLP benchmarks compared to conventional reward models. Our codes are publicly released to facilitate further research (https://github.com/THU-KEG/Agentic-Reward-Modeling).", "authors": ["Hao Peng", "Yunjia Qi", "Xiaozhi Wang", "Zijun Yao", "Bin Xu", "Lei Hou", "Juanzi Li"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-26", "url": "https://arxiv.org/abs/2502.19328", "pdf_url": "https://arxiv.org/pdf/2502.19328v1", "arxiv_id": "2502.19328", "doi": "10.48550/arXiv.2502.19328", "citation_count": 49, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/THU-KEG/Agentic-Reward-Modeling", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.4247}
{"id": "00e6bc7efdcdc603d11974e7e193e026eb6db7fcbdd67a71b436fa380922410b", "sources": ["arxiv", "semantic_scholar"], "title": "FSPO: Few-Shot Optimization of Synthetic Preferences Personalizes to Real Users", "abstract": "Effective personalization of LLMs is critical for a broad range of user-interfacing applications such as virtual assistants and content curation. Inspired by the strong in-context capabilities of LLMs, we propose few-shot preference optimization (FSPO), an algorithm for LLM personalization that reframes reward modeling as a meta-learning problem. Under FSPO, an LLM learns to quickly infer a personalized reward function for a user via a few labeled preferences. FSPO also utilizes user description rationalization (RAT) to encourage better reward modeling and instruction following, recovering performance with the oracle user description. Since real-world preference data is challenging to collect at scale, we propose careful design choices to construct synthetic preference datasets for personalization, generating over 1M synthetic personalized preferences using publicly available LLMs. To successfully transfer from synthetic data to real users, we find it crucial for the data to exhibit both high diversity and coherent, self-consistent structure. We evaluate FSPO on personalized open-ended generation for up to 1,500 synthetic users across three domains: movie reviews, education, and open-ended question answering. We also run a controlled human study. Overall, FSPO achieves an 87% Alpaca Eval winrate in generating responses that are personalized to synthetic users and a 70% winrate with real human users in open-ended question answering.", "authors": ["Anikait Singh", "Sheryl Hsu", "Kyle Hsu", "Eric Mitchell", "Stefano Ermon", "Tatsunori Hashimoto", "Archit Sharma", "Chelsea Finn"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.HC", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-02-26", "url": "https://arxiv.org/abs/2502.19312", "pdf_url": "https://arxiv.org/pdf/2502.19312v2", "arxiv_id": "2502.19312", "doi": null, "citation_count": 20, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3306}
{"id": "e4cd76c36faf75ae1c8dc71ea566101507bf7aeb25b96998ccc282fd4695ac5d", "sources": ["arxiv", "semantic_scholar"], "title": "Larger or Smaller Reward Margins to Select Preferences for Alignment?", "abstract": "Preference learning is critical for aligning large language models (LLMs) with human values, with the quality of preference datasets playing a crucial role in this process. While existing metrics primarily assess data quality based on either explicit or implicit reward margins, they often provide contradictory evaluations for the same data. To address this issue, we introduce the alignment potential metric, which quantifies the gap from the model's current implicit reward margin to the target explicit reward margin, thereby estimating the model's potential to align with the preference data. Empirical results demonstrate that training on data selected by this metric consistently enhances alignment performance, surpassing existing metrics across different base models and optimization objectives. Furthermore, our method extends to self-play data generation frameworks, where the metric is used to identify high-quality data within the self-generated content by LLMs. Under this data generation scenario, our method surpasses current state-of-the-art (SOTA) results across various training settings and demonstrates continuous improvements in alignment performance as dataset size and training iterations increase.", "authors": ["Kexin Huang", "Junkang Wu", "Ziqian Chen", "Xue Wang", "Jinyang Gao", "Bolin Ding", "Jiancan Wu", "Xiangnan He", "Xiang Wang"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-25", "url": "https://arxiv.org/abs/2503.01864", "pdf_url": "https://arxiv.org/pdf/2503.01864v1", "arxiv_id": "2503.01864", "doi": "10.48550/arXiv.2503.01864", "citation_count": 9, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.25}
{"id": "d1045fea6a5cd25c5c9072c7e21ef8c500c2e70b31ed8596a6381768c709e211", "sources": ["arxiv", "semantic_scholar"], "title": "Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data", "abstract": "Supervised fine-tuning (SFT) has become a crucial step for aligning pretrained large language models (LLMs) using supervised datasets of input-output pairs. However, despite being supervised, SFT is inherently limited by its generative training objective. To address its limitations, the existing common strategy is to follow SFT with a separate phase of preference optimization (PO), which relies on either human-labeled preference data or a strong reward model to guide the learning process. In this paper, we address the limitations of SFT by exploring one of the most successful techniques in conventional supervised learning: discriminative learning. We introduce Discriminative Fine-Tuning (DFT), an improved variant of SFT, which mitigates the burden of collecting human-labeled preference data or training strong reward models. Unlike SFT that employs a generative approach and overlooks negative data, DFT adopts a discriminative paradigm that increases the probability of positive answers while suppressing potentially negative ones, aiming for data prediction instead of token prediction. Our contributions include: (i) a discriminative probabilistic framework for fine-tuning LLMs by explicitly modeling the discriminative likelihood of an answer among all possible outputs given an input; (ii) efficient algorithms to optimize this discriminative likelihood; and (iii) extensive experiments demonstrating DFT's effectiveness, achieving performance better than SFT and comparable to if not better than SFT$\\rightarrow$PO. The code can be found at https://github.com/Optimization-AI/DFT.", "authors": ["Siqi Guo", "Ilgee Hong", "Vicente Balmaseda", "Changlong Yu", "Liang Qiu", "Xin Liu", "Haoming Jiang", "Tuo Zhao", "Tianbao Yang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-25", "url": "https://arxiv.org/abs/2502.18679", "pdf_url": "https://arxiv.org/pdf/2502.18679v3", "arxiv_id": "2502.18679", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Optimization-AI/DFT", "venue": "International Conference on Machine Learning", "quality_score": 0.108}
{"id": "546c7440afa4f5ecdd641a331f9c8167d0924faf13930c14421bd8a9e3c4c2bc", "sources": ["arxiv", "semantic_scholar"], "title": "Pretrain Value, Not Reward: Decoupled Value Policy Optimization", "abstract": "In this paper, we explore how directly pretraining a value model simplifies and stabilizes reinforcement learning from human feedback (RLHF). In reinforcement learning, value estimation is the key to policy optimization, distinct from reward supervision. The value function predicts the \\emph{return-to-go} of a partial answer, that is, how promising the partial answer is if it were continued to completion. In RLHF, however, the standard pipeline first pretrains a reward model and then learns a value function online, even though no new reward signals are available once preference data is collected. This makes critic learning redundant, as the process of training a reward model and then deriving a value model is informationally equivalent to directly pretraining a value model. Importantly, this requires no additional supervision, and our value model is trained on exactly the same data used for reward modeling. Building on this insight, we introduce \\emph{Decoupled Value Policy Optimization} (DVPO), a framework that pretrains a \\emph{Global Value Model} (GVM) offline and freezes it as a universal critic for policy learning. The GVM provides stable, fine-grained credit assignment without critic drift or trajectory sampling. Experiments across MT-Bench, Alpaca-Eval, and Arena-Hard demonstrate that DVPO matches or surpasses state-of-the-art RLHF methods. These results highlight RLHF can be reframed as policy-only optimization guided by a single pretrained value model.", "authors": ["Chenghua Huang", "Lu Wang", "Fangkai Yang", "Pu Zhao", "Zhixu Li", "Qingwei Lin", "Dongmei Zhang", "Saravan Rajmohan", "Qi Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-24", "url": "https://arxiv.org/abs/2502.16944", "pdf_url": "https://arxiv.org/pdf/2502.16944v2", "arxiv_id": "2502.16944", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "659b98252045b8b988bfeaa938420b40d200899bdfedf0ebcfdb670a7ff7088b", "sources": ["arxiv", "semantic_scholar"], "title": "RLHF in an SFT Way: From Optimal Solution to Reward-Weighted Alignment", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning Large Language Models (LLMs) with human values. However, RLHF has been continuously challenged by its high complexity in implementation and computation consumption, specifically for online sampling-based methods like Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO). Even with recent simplifications, such as Direct Preference Optimization (DPO) that designs an offline implicit reward learning objective relying on pre-collected preference datasets, the problems of over-fitting and training instability remain hindering the alignment process from the expected optimal performance. To address the existing challenges, we propose a novel simplification of RLHF from the perspective of variational inference, called Variational Alignment with Re-weighting (VAR). Specifically, by directly minimizing the distribution gap between the learning LLM policy and the optimal solution of RLHF, we transform the alignment objective into an offline reward-driven re-weighted supervised fine-tuning (SFT) form, which only requires minor adjustment on the SFT loss to obtain noticeable improvement on training stability and effectiveness. In comprehensive evaluation benchmarks, our objective empowers LLMs to outperform offline alignments, demonstrating superior performance in both helpfulness and harmlessness metrics (avg. $\\uparrow7.16\\%$ than DPO). Meanwhile, when compared to online sampling methods, our method is also comparable even better while significantly reducing computational overhead and accelerating convergence speed (over $5\\times$ faster than GRPO), suggesting our approach as an efficient and effective solution in bridging the gap between efficiency and performance in LLM alignment.", "authors": ["Yuhao Du", "Zhuo Li", "Pengyu Cheng", "Zhihong Chen", "Yuejiao Xie", "Xiang Wan", "Anningzhe Gao"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-16", "url": "https://arxiv.org/abs/2502.11026", "pdf_url": "https://arxiv.org/pdf/2502.11026v3", "arxiv_id": "2502.11026", "doi": null, "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.294}
{"id": "a45c279e3a6d9db75715153b9d5987808b52c3d225ae14a2b4770ac3c2f66810", "sources": ["arxiv", "semantic_scholar"], "title": "Process Reward Models for LLM Agents: Practical Framework and Directions", "abstract": "We introduce Agent Process Reward Models (AgentPRM), a simple and scalable framework for training LLM agents to continually improve through interactions. AgentPRM follows a lightweight actor-critic paradigm, using Monte Carlo rollouts to compute reward targets and optimize policies. It requires minimal modifications to existing RLHF pipelines, making it easy to integrate at scale. Beyond AgentPRM, we propose InversePRM, which learns process rewards directly from demonstrations without explicit outcome supervision. We also explore key challenges and opportunities, including exploration, process reward shaping, and model-predictive reasoning. We evaluate on ALFWorld benchmark, show that small 3B models trained with AgentPRM and InversePRM outperform strong GPT-4o baselines, and analyze test-time scaling, reward hacking, and more. Our code is available at: https://github.com/sanjibanc/agent_prm.", "authors": ["Sanjiban Choudhury"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-14", "url": "https://arxiv.org/abs/2502.10325", "pdf_url": "https://arxiv.org/pdf/2502.10325v1", "arxiv_id": "2502.10325", "doi": "10.48550/arXiv.2502.10325", "citation_count": 66, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/sanjibanc/agent_prm", "venue": "arXiv.org", "quality_score": 0.4565}
{"id": "05be952385e975fc233797ffd481e9a4a7b665ef16ac4a543017ebe747792ef7", "sources": ["arxiv", "semantic_scholar"], "title": "Provably Efficient Online RLHF with One-Pass Reward Modeling", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has shown remarkable success in aligning Large Language Models (LLMs) with human preferences. Traditional RLHF methods rely on a fixed dataset, which often suffers from limited coverage. To this end, online RLHF has emerged as a promising direction, enabling iterative data collection and refinement. Despite its potential, this paradigm faces a key bottleneck: the requirement to continuously integrate new data into the dataset and re-optimize the model from scratch at each iteration, resulting in computational and storage costs that grow linearly with the number of iterations. In this work, we address this challenge by proposing a one-pass reward modeling method that eliminates the need to store historical data and achieves constant-time updates per iteration. Specifically, we first formalize RLHF as a contextual preference bandit and develop a new algorithm based on online mirror descent with a tailored local norm, replacing the standard maximum likelihood estimation for reward modeling. We then apply it to various online RLHF settings, including passive data collection, active data collection, and deployment-time adaptation. We provide theoretical guarantees showing that our method enhances both statistical and computational efficiency. Finally, we design practical algorithms for LLMs and conduct experiments with the Llama-3-8B-Instruct and Qwen2.5-7B-Instruct models on Ultrafeedback and Mixture2 datasets, validating the effectiveness of our approach.", "authors": ["Long-Fei Li", "Yu-Yang Qian", "Peng Zhao", "Zhi-Hua Zhou"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-02-11", "url": "https://arxiv.org/abs/2502.07193", "pdf_url": "https://arxiv.org/pdf/2502.07193v3", "arxiv_id": "2502.07193", "doi": null, "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2113}
{"id": "84f4bc71d45cf2dbfb72b536d632b54bd111ecab05f02e2814d24a7def6f70b5", "sources": ["arxiv", "semantic_scholar"], "title": "Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks", "abstract": "Collaborative learning enables multiple participants to learn a single global model by exchanging focused updates instead of sharing data. One of the core challenges in collaborative learning is ensuring that participants are rewarded fairly for their contributions, which entails two key sub-problems: contribution assessment and reward allocation. This work focuses on fair reward allocation, where the participants are incentivized through model rewards - differentiated final models whose performance is commensurate with the contribution. In this work, we leverage the concept of slimmable neural networks to collaboratively learn a shared global model whose performance degrades gracefully with a reduction in model width. We also propose a post-training fair allocation algorithm that determines the model width for each participant based on their contributions. We theoretically study the convergence of our proposed approach and empirically validate it using extensive experiments on different datasets and architectures. We also extend our approach to enable training-time model reward allocation.", "authors": ["Nurbek Tastan", "Samuel Horvath", "Karthik Nandakumar"], "categories": ["cs.LG", "cs.DC"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-07", "url": "https://arxiv.org/abs/2502.04850", "pdf_url": "https://arxiv.org/pdf/2502.04850v2", "arxiv_id": "2502.04850", "doi": "10.48550/arXiv.2502.04850", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2113}
{"id": "4840c97328c5a423ef7e9fe324012617d39b1ed2a8a4d81e3d7cf6e7b73fdef0", "sources": ["arxiv", "semantic_scholar"], "title": "PILAF: Optimal Human Preference Sampling for Reward Modeling", "abstract": "As large language models increasingly drive real-world applications, aligning them with human values becomes paramount. Reinforcement Learning from Human Feedback (RLHF) has emerged as a key technique, translating preference data into reward models when oracle human values remain inaccessible. In practice, RLHF mostly relies on approximate reward models, which may not consistently guide the policy toward maximizing the underlying human values. We propose Policy-Interpolated Learning for Aligned Feedback (PILAF), a novel response sampling strategy for preference labeling that explicitly aligns preference learning with maximizing the underlying oracle reward. PILAF is theoretically grounded, demonstrating optimality from both an optimization and a statistical perspective. The method is straightforward to implement and demonstrates strong performance in iterative and online RLHF settings where feedback curation is critical.", "authors": ["Yunzhen Feng", "Ariel Kwiatkowski", "Kunhao Zheng", "Julia Kempe", "Yaqi Duan"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-02-06", "url": "https://arxiv.org/abs/2502.04270", "pdf_url": "https://arxiv.org/pdf/2502.04270v1", "arxiv_id": "2502.04270", "doi": "10.48550/arXiv.2502.04270", "citation_count": 18, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3197}
{"id": "2066d042cc4261d2f1f01f49f10e698c8844f43abbe1725df679d6897784f218", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Cost-Effective Reward Guided Text Generation", "abstract": "Reward-guided text generation (RGTG) has emerged as a viable alternative to offline reinforcement learning from human feedback (RLHF). RGTG methods can align baseline language models to human preferences without further training like in standard RLHF methods. However, they rely on a reward model to score each candidate token generated by the language model at inference, incurring significant test-time overhead. Additionally, the reward model is usually only trained to score full sequences, which can lead to sub-optimal choices for partial sequences. In this work, we present a novel reward model architecture that is trained, using a Bradley-Terry loss, to prefer the optimal expansion of a sequence with just a \\emph{single call} to the reward model at each step of the generation process. That is, a score for all possible candidate tokens is generated simultaneously, leading to efficient inference. We theoretically analyze various RGTG reward models and demonstrate that prior techniques prefer sub-optimal sequences compared to our method during inference. Empirically, our reward model leads to significantly faster inference than other RGTG methods. It requires fewer calls to the reward model and performs competitively compared to previous RGTG and offline RLHF methods.", "authors": ["Ahmad Rashid", "Ruotian Wu", "Rongqi Fan", "Hongliang Li", "Agustinus Kristiadi", "Pascal Poupart"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-06", "url": "https://arxiv.org/abs/2502.04517", "pdf_url": "https://arxiv.org/pdf/2502.04517v2", "arxiv_id": "2502.04517", "doi": "10.48550/arXiv.2502.04517", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2113}
{"id": "6200f20906672d86011fd07d8d65c6fef8d34f5d6f6164b0cdcbce53cb21863a", "sources": ["arxiv", "semantic_scholar"], "title": "PerPO: Perceptual Preference Optimization via Discriminative Rewarding", "abstract": "This paper presents Perceptual Preference Optimization (PerPO), a perception alignment method aimed at addressing the visual discrimination challenges in generative pre-trained multimodal large language models (MLLMs). To align MLLMs with human visual perception process, PerPO employs discriminative rewarding to gather diverse negative samples, followed by listwise preference optimization to rank them.By utilizing the reward as a quantitative margin for ranking, our method effectively bridges generative preference optimization and discriminative empirical risk minimization. PerPO significantly enhances MLLMs' visual discrimination capabilities while maintaining their generative strengths, mitigates image-unconditional reward hacking, and ensures consistent performance across visual tasks. This work marks a crucial step towards more perceptually aligned and versatile MLLMs. We also hope that PerPO will encourage the community to rethink MLLM alignment strategies.", "authors": ["Zining Zhu", "Liang Zhao", "Kangheng Lin", "Jinze Yang", "En Yu", "Chenglong Liu", "Haoran Wei", "Jianjian Sun", "Zheng Ge", "Xiangyu Zhang"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-05", "url": "https://arxiv.org/abs/2502.04371", "pdf_url": "https://arxiv.org/pdf/2502.04371v1", "arxiv_id": "2502.04371", "doi": "10.48550/arXiv.2502.04371", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "dbffc4cc0679aff4e76f98f36da75ee33cd60574d68c9c638e9090c70d21cdd9", "sources": ["arxiv", "semantic_scholar"], "title": "Reviving The Classics: Active Reward Modeling in Large Language Model Alignment", "abstract": "Building neural reward models from human preferences is a pivotal component in reinforcement learning from human feedback (RLHF) and large language model alignment research. Given the scarcity and high cost of human annotation, how to select the most informative pairs to annotate is an essential yet challenging open problem. In this work, we highlight the insight that an ideal comparison dataset for reward modeling should balance exploration of the representation space and make informative comparisons between pairs with moderate reward differences. Technically, challenges arise in quantifying the two objectives and efficiently prioritizing the comparisons to be annotated. To address this, we propose the Fisher information-based selection strategies, adapt theories from the classical experimental design literature, and apply them to the final linear layer of the deep neural network-based reward modeling tasks. Empirically, our method demonstrates remarkable performance, high computational efficiency, and stability compared to other selection methods from deep learning and classical statistical literature across multiple open-source LLMs and datasets. Further ablation studies reveal that incorporating cross-prompt comparisons in active reward modeling significantly enhances labeling efficiency, shedding light on the potential for improved annotation strategies in RLHF.", "authors": ["Yunyi Shen", "Hao Sun", "Jean-François Ton"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-04", "url": "https://arxiv.org/abs/2502.04354", "pdf_url": "https://arxiv.org/pdf/2502.04354v1", "arxiv_id": "2502.04354", "doi": "10.48550/arXiv.2502.04354", "citation_count": 9, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "df7959677c3be321f2cf85732cfe887e1322f01bd9dca314d2bf44d5a023ea90", "sources": ["arxiv", "semantic_scholar"], "title": "Process Reinforcement through Implicit Rewards", "abstract": "Dense process rewards have proven a more effective alternative to the sparse outcome-level rewards in the inference-time scaling of large language models (LLMs), particularly in tasks requiring complex multi-step reasoning. While dense rewards also offer an appealing choice for the reinforcement learning (RL) of LLMs since their fine-grained rewards have the potential to address some inherent issues of outcome rewards, such as training efficiency and credit assignment, this potential remains largely unrealized. This can be primarily attributed to the challenges of training process reward models (PRMs) online, where collecting high-quality process labels is prohibitively expensive, making them particularly vulnerable to reward hacking. To address these challenges, we propose PRIME (Process Reinforcement through IMplicit rEwards), which enables online PRM updates using only policy rollouts and outcome labels through implict process rewards. PRIME combines well with various advantage functions and forgoes the dedicated reward model training phrase that existing approaches require, substantially reducing the development overhead. We demonstrate PRIME's effectiveness on competitional math and coding. Starting from Qwen2.5-Math-7B-Base, PRIME achieves a 15.1% average improvement across several key reasoning benchmarks over the SFT model. Notably, our resulting model, Eurus-2-7B-PRIME, surpasses Qwen2.5-Math-7B-Instruct on seven reasoning benchmarks with 10% of its training data.", "authors": ["Ganqu Cui", "Lifan Yuan", "Zefan Wang", "Hanbin Wang", "Yuchen Zhang", "Jiacheng Chen", "Wendi Li", "Bingxiang He", "Yuchen Fan", "Tianyu Yu", "Qixin Xu", "Weize Chen", "Jiarui Yuan", "Huayu Chen", "Kaiyan Zhang", "Xingtai Lv", "Shuo Wang", "Yuan Yao", "Xu Han", "Hao Peng", "Yu Cheng", "Zhiyuan Liu", "Maosong Sun", "Bowen Zhou", "Ning Ding"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-02-03", "url": "https://arxiv.org/abs/2502.01456", "pdf_url": "https://arxiv.org/pdf/2502.01456v2", "arxiv_id": "2502.01456", "doi": "10.48550/arXiv.2502.01456", "citation_count": 365, "influential_citation_count": 54, "has_code": true, "code_url": "https://github.com/PRIME-RL/PRIME", "venue": "arXiv.org", "quality_score": 0.8702}
{"id": "7a04fc67afec50ccb7684918cade472d03a8154d5a39dc377c541e95a7ceb1df", "sources": ["arxiv", "semantic_scholar"], "title": "Avoiding $\\mathbf{exp(R_{max})}$ scaling in RLHF through Preference-based Exploration", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has emerged as a pivotal technique for large language model (LLM) alignment. This paper studies the setting of online RLHF and focus on improving sample efficiency. All existing algorithms in online RLHF, whether doing passive exploration or active exploration, suffer from a sample complexity that scales exponentially with the scale of the reward function. This fundamental limitation hinders their effectiveness in scenarios with heavily skewed preferences, e.g. questions with a unique correct solution. To address this, we introduce Self-Exploring Preference-Incentive Online Preference Optimization (SE-POPO), an online RLHF algorithm that for the first time achieves a sample complexity that scales polynomially with the reward scale, answering an open problem raised by Xie et al. (2024).. Theoretically, we demonstrate that the sample complexity of SE-POPO dominates that of existing exploration algorithms. Empirically, our systematic evaluation confirms that SE-POPO is more sample-efficient than both exploratory and non-exploratory baselines, in two primary application scenarios of RLHF as well as on public benchmarks, marking a significant step forward in RLHF algorithm design. The code is available at https://github.com/MYC000801/SE-POPO.", "authors": ["Mingyu Chen", "Yiding Chen", "Wen Sun", "Xuezhou Zhang"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-02-02", "url": "https://arxiv.org/abs/2502.00666", "pdf_url": "https://arxiv.org/pdf/2502.00666v3", "arxiv_id": "2502.00666", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/MYC000801/SE-POPO", "venue": null, "quality_score": 0.0753}
{"id": "576d881ab392727b176bddd178fbe0ec86579e2b33b8432787d42a18ed955d79", "sources": ["arxiv", "semantic_scholar"], "title": "The Energy Loss Phenomenon in RLHF: A New Perspective on Mitigating Reward Hacking", "abstract": "This work identifies the Energy Loss Phenomenon in Reinforcement Learning from Human Feedback (RLHF) and its connection to reward hacking. Specifically, energy loss in the final layer of a Large Language Model (LLM) gradually increases during the RL process, with an excessive increase in energy loss characterizing reward hacking. Beyond empirical analysis, we further provide a theoretical foundation by proving that, under mild conditions, the increased energy loss reduces the upper bound of contextual relevance in LLMs, which is a critical aspect of reward hacking as the reduced contextual relevance typically indicates overfitting to reward model-favored patterns in RL. To address this issue, we propose an Energy loss-aware PPO algorithm (EPPO) which penalizes the increase in energy loss in the LLM's final layer during reward calculation to prevent excessive energy loss, thereby mitigating reward hacking. We theoretically show that EPPO can be conceptually interpreted as an entropy-regularized RL algorithm, which provides deeper insights into its effectiveness. Extensive experiments across various LLMs and tasks demonstrate the commonality of the energy loss phenomenon, as well as the effectiveness of EPPO in mitigating reward hacking and improving RLHF performance.", "authors": ["Yuchun Miao", "Sen Zhang", "Liang Ding", "Yuqi Zhang", "Lefei Zhang", "Dacheng Tao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-31", "url": "https://arxiv.org/abs/2501.19358", "pdf_url": "https://arxiv.org/pdf/2501.19358v3", "arxiv_id": "2501.19358", "doi": "10.48550/arXiv.2501.19358", "citation_count": 15, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.301}
{"id": "e208453afaa7a3330f06de5f8bee81bce14da78249c079aa3d9e06167d09ec5e", "sources": ["arxiv", "semantic_scholar"], "title": "From Sparse to Dense: Toddler-inspired Reward Transition in Goal-Oriented Reinforcement Learning", "abstract": "Reinforcement learning (RL) agents often face challenges in balancing exploration and exploitation, particularly in environments where sparse or dense rewards bias learning. Biological systems, such as human toddlers, naturally navigate this balance by transitioning from free exploration with sparse rewards to goal-directed behavior guided by increasingly dense rewards. Inspired by this natural progression, we investigate the Toddler-Inspired Reward Transition in goal-oriented RL tasks. Our study focuses on transitioning from sparse to potential-based dense (S2D) rewards while preserving optimal strategies. Through experiments on dynamic robotic arm manipulation and egocentric 3D navigation tasks, we demonstrate that effective S2D reward transitions significantly enhance learning performance and sample efficiency. Additionally, using a Cross-Density Visualizer, we show that S2D transitions smooth the policy loss landscape, resulting in wider minima that improve generalization in RL models. In addition, we reinterpret Tolman's maze experiments, underscoring the critical role of early free exploratory learning in the context of S2D rewards.", "authors": ["Junseok Park", "Hyeonseo Yang", "Min Whoo Lee", "Won-Seok Choi", "Minsu Lee", "Byoung-Tak Zhang"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-29", "url": "https://arxiv.org/abs/2501.17842", "pdf_url": "https://arxiv.org/pdf/2501.17842v1", "arxiv_id": "2501.17842", "doi": "10.1109/TCDS.2025.3642038", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Cognitive and Developmental Systems", "quality_score": 0.0753}
{"id": "150abbe58489a0a745ffb9336125cbb14aaf9e14da2b0ee067e72d7f5d1ada23", "sources": ["arxiv", "semantic_scholar"], "title": "CRPO: Confidence-Reward Driven Preference Optimization for Machine Translation", "abstract": "Large language models (LLMs) have shown great potential in natural language processing tasks, but their application to machine translation (MT) remains challenging due to pretraining on English-centric data and the complexity of reinforcement learning from human feedback (RLHF). Direct Preference Optimization (DPO) has emerged as a simpler and more efficient alternative, but its performance depends heavily on the quality of preference data. To address this, we propose Confidence-Reward driven Preference Optimization (CRPO), a novel method that combines reward scores with model confidence to improve data selection for fine-tuning. CRPO selects challenging sentence pairs where the model is uncertain or underperforms, leading to more effective learning. While primarily designed for LLMs, CRPO also generalizes to encoder-decoder models like NLLB, demonstrating its versatility. Empirical results show that CRPO outperforms existing methods such as RS-DPO, RSO and MBR score in both translation accuracy and data efficiency.", "authors": ["Guofeng Cui", "Pichao Wang", "Yang Liu", "Zemian Ke", "Zhu Liu", "Vimal Bhat"], "categories": ["cs.CL", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-23", "url": "https://arxiv.org/abs/2501.13927", "pdf_url": "https://arxiv.org/pdf/2501.13927v1", "arxiv_id": "2501.13927", "doi": "10.48550/arXiv.2501.13927", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.1505}
{"id": "bef845d8d55a24c4ffde408599c320e7610e093817e0ebe0a1fab7df9a6e9853", "sources": ["arxiv", "semantic_scholar"], "title": "OpenGenAlign: A Preference Dataset and Benchmark for Trustworthy Reward Modeling in Open-Ended, Long-Context Generation", "abstract": "Reward Modeling is critical in evaluating and improving the generation of Large Language Models (LLMs). While numerous recent works have shown its feasibility in improving safety, helpfulness, reasoning, and instruction-following ability, its capability and generalization to open-ended long-context generation is still rarely explored. In this paper, we introduce OpenGenAlign, a framework and a high-quality dataset designed to develop reward models to evaluate and improve hallucination-free, comprehensive, reliable, and efficient open-ended long-context generation. We define four key metrics to assess generation quality and develop an automated pipeline to evaluate the outputs of multiple LLMs across long-context QA, Data-to-Text, and Summarization scenarios using o3, ending up with 33K high-quality preference data with a human agreement rate of 81\\%. Experimental results first demonstrate that existing reward models perform suboptimally on the held-out benchmark. And Our trained reward model achieves superior performance in the benchmark and effectively improves the generation quality of the policy models using Reinforcement Learning (RL). Additionally, OpenGenAlign could be used for effective guided generation in existing datasets. Furthermore, we demonstrate that the OpenGenAlign could be integrated with reward data from other domains to achieve better performance.", "authors": ["Hanning Zhang", "Juntong Song", "Juno Zhu", "Yuanhao Wu", "Tong Zhang", "Cheng Niu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-22", "url": "https://arxiv.org/abs/2501.13264", "pdf_url": "https://arxiv.org/pdf/2501.13264v3", "arxiv_id": "2501.13264", "doi": null, "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "3753fd3a63eba5f9356285cb0a7713344f300953b900449752f08c9e6eddf51e", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Reward Hacking: Causal Rewards for Large Language Model Alignment", "abstract": "Recent advances in large language models (LLMs) have demonstrated significant progress in performing complex tasks. While Reinforcement Learning from Human Feedback (RLHF) has been effective in aligning LLMs with human preferences, it is susceptible to spurious correlations in reward modeling. Consequently, it often introduces biases-such as length bias, sycophancy, conceptual bias, and discrimination-that hinder the model's ability to capture true causal relationships. To address this, we propose a novel causal reward modeling approach that integrates causality to mitigate these spurious correlations. Our method enforces counterfactual invariance, ensuring reward predictions remain consistent when irrelevant variables are altered. Through experiments on both synthetic and real-world datasets, we show that our approach mitigates various types of spurious correlations effectively, resulting in more reliable and fair alignment of LLMs with human preferences. As a drop-in enhancement to the existing RLHF workflow, our causal reward modeling provides a practical way to improve the trustworthiness and fairness of LLM finetuning.", "authors": ["Chaoqi Wang", "Zhuokai Zhao", "Yibo Jiang", "Zhaorun Chen", "Chen Zhu", "Yuxin Chen", "Jiayi Liu", "Lizhu Zhang", "Xiangjun Fan", "Hao Ma", "Sinong Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-16", "url": "https://arxiv.org/abs/2501.09620", "pdf_url": "https://arxiv.org/pdf/2501.09620v2", "arxiv_id": "2501.09620", "doi": "10.48550/arXiv.2501.09620", "citation_count": 37, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3949}
{"id": "fe67cd3c3abf15452a6ba0df282f8e1d2845b41c61209eec2f223b7cf7ceb40f", "sources": ["arxiv", "semantic_scholar"], "title": "Influencing Humans to Conform to Preference Models for RLHF", "abstract": "Designing a reinforcement learning from human feedback (RLHF) algorithm to approximate a human's unobservable reward function requires assuming, implicitly or explicitly, a model of human preferences. A preference model that poorly describes how humans generate preferences risks learning a poor approximation of the human's reward function. In this paper, we conduct three human studies to asses whether one can influence the expression of real human preferences to more closely conform to a desired preference model. Importantly, our approach does not seek to alter the human's unobserved reward function. Rather, we change how humans use this reward function to generate preferences, such that they better match whatever preference model is assumed by a particular RLHF algorithm. We introduce three interventions: showing humans the quantities that underlie a preference model, which is normally unobservable information derived from the reward function; training people to follow a specific preference model; and modifying the preference elicitation question. All intervention types show significant effects, providing practical tools to improve preference data quality and the resultant alignment of the learned reward functions. Overall we establish a novel research direction in model alignment: designing interfaces and training interventions to increase human conformance with the modeling assumptions of the algorithm that will learn from their input.", "authors": ["Stephane Hatgis-Kessell", "W. Bradley Knox", "Serena Booth", "Peter Stone"], "categories": ["cs.LG", "cs.AI", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-11", "url": "https://arxiv.org/abs/2501.06416", "pdf_url": "https://arxiv.org/pdf/2501.06416v3", "arxiv_id": "2501.06416", "doi": "10.48550/arXiv.2501.06416", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "77c68aa081aa4be934bfcc46be2538cbacd26587c225a6b3f035a376e98c5808", "sources": ["arxiv", "semantic_scholar"], "title": "Robustness in the Face of Partial Identifiability in Reward Learning", "abstract": "In Reward Learning (ReL), we are given feedback on an unknown target reward, and the goal is to use this information to recover it in order to carry out some downstream application, e.g., planning. When the feedback is not informative enough, the target reward is only partially identifiable, i.e., there exists a set of rewards, called the feasible set, that are equally plausible candidates for the target reward. In these cases, the ReL algorithm might recover a reward function different from the target reward, possibly leading to a failure in the application. In this paper, we introduce a general ReL framework that permits to quantify the drop in \"performance\" suffered in the considered application because of identifiability issues. Building on this, we propose a robust approach to address the identifiability problem in a principled way, by maximizing the \"performance\" with respect to the worst-case reward in the feasible set. We then develop Rob-ReL, a ReL algorithm that applies this robust approach to the subset of ReL problems aimed at assessing a preference between two policies, and we provide theoretical guarantees on sample and iteration complexity for Rob-ReL. We conclude with a proof-of-concept experiment to illustrate the considered setting.", "authors": ["Filippo Lazzati", "Alberto Maria Metelli"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2025-01-10", "url": "https://arxiv.org/abs/2501.06376", "pdf_url": "https://arxiv.org/pdf/2501.06376v2", "arxiv_id": "2501.06376", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0109}
{"id": "ad48472add361c1f0a13a577d2648ebf77cd0d6af80d5ad5f1f56a4540bca399", "sources": ["arxiv", "semantic_scholar"], "title": "Constraints as Rewards: Reinforcement Learning for Robots without Reward Functions", "abstract": "Reinforcement learning has become an essential algorithm for generating complex robotic behaviors. However, to learn such behaviors, it is necessary to design a reward function that describes the task, which often consists of multiple objectives that needs to be balanced. This tuning process is known as reward engineering and typically involves extensive trial-and-error. In this paper, to avoid this trial-and-error process, we propose the concept of Constraints as Rewards (CaR). CaR formulates the task objective using multiple constraint functions instead of a reward function and solves a reinforcement learning problem with constraints using the Lagrangian-method. By adopting this approach, different objectives are automatically balanced, because Lagrange multipliers serves as the weights among the objectives. In addition, we will demonstrate that constraints, expressed as inequalities, provide an intuitive interpretation of the optimization target designed for the task. We apply the proposed method to the standing-up motion generation task of a six-wheeled-telescopic-legged robot and demonstrate that the proposed method successfully acquires the target behavior, even though it is challenging to learn with manually designed reward functions.", "authors": ["Yu Ishihara", "Noriaki Takasugi", "Kotaro Kawakami", "Masaya Kinoshita", "Kazumi Aoyama"], "categories": ["cs.RO", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-08", "url": "https://arxiv.org/abs/2501.04228", "pdf_url": "https://arxiv.org/pdf/2501.04228v2", "arxiv_id": "2501.04228", "doi": "10.48550/arXiv.2501.04228", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "8c17a57d84befdc1acb0ee6e867d09ec7efee03ea7bf7a947701eb3bff7eada2", "sources": ["arxiv", "semantic_scholar"], "title": "Segmenting Text and Learning Their Rewards for Improved RLHF in Language Model", "abstract": "Reinforcement learning from human feedback (RLHF) has been widely adopted to align language models (LMs) with human preference. Prior RLHF works typically take a bandit formulation, which, though intuitive, ignores the sequential nature of LM generation and can suffer from the sparse reward issue. While recent works propose dense token-level RLHF, treating each token as an action may be oversubtle to proper reward assignment. In this paper, we seek to get the best of both by training and utilizing a segment-level reward model, which assigns a reward to each semantically complete text segment that spans over a short sequence of tokens. For reward learning, our method allows dynamic text segmentation and compatibility with standard sequence-preference datasets. For effective RL-based LM training against segment reward, we generalize the classical scalar bandit reward normalizers into location-aware normalizer functions and interpolate the segment reward for further densification. With these designs, our method performs competitively on three popular RLHF benchmarks for LM policy: AlpacaEval 2.0, Arena-Hard, and MT-Bench. Ablation studies are conducted to further demonstrate our method.", "authors": ["Yueqin Yin", "Shentao Yang", "Yujia Xie", "Ziyi Yang", "Yuting Sun", "Hany Awadalla", "Weizhu Chen", "Mingyuan Zhou"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-06", "url": "https://arxiv.org/abs/2501.02790", "pdf_url": "https://arxiv.org/pdf/2501.02790v1", "arxiv_id": "2501.02790", "doi": "10.48550/arXiv.2501.02790", "citation_count": 13, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2865}
{"id": "077650cd66fcdc083ecf99a15cd6723b37be0b64732d18eff2376f45ae2141f7", "sources": ["arxiv", "semantic_scholar"], "title": "Inversely Learning Transferable Rewards via Abstracted States", "abstract": "Inverse reinforcement learning (IRL) has progressed significantly toward accurately learning the underlying rewards in both discrete and continuous domains from behavior data. The next advance is to learn {\\em intrinsic} preferences in ways that produce useful behavior in settings or tasks which are different but aligned with the observed ones. In the context of robotic applications, this helps integrate robots into processing lines involving new tasks (with shared intrinsic preferences) without programming from scratch. We introduce a method to inversely learn an abstract reward function from behavior trajectories in two or more differing instances of a domain. The abstract reward function is then used to learn task behavior in another separate instance of the domain. This step offers evidence of its transferability and validates its correctness. We evaluate the method on trajectories in tasks from multiple domains in OpenAI's Gym testbed and AssistiveGym and show that the learned abstract reward functions can successfully learn task behaviors in instances of the respective domains, which have not been seen previously.", "authors": ["Yikang Gui", "Prashant Doshi"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-03", "url": "https://arxiv.org/abs/2501.01669", "pdf_url": "https://arxiv.org/pdf/2501.01669v4", "arxiv_id": "2501.01669", "doi": "10.48550/arXiv.2501.01669", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "e525fe5b4267cbd8493300432e8a59771f737218fdb942b484cba8bb4d0ae5ce", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamic Scaling of Unit Tests for Code Reward Modeling", "abstract": "Current large language models (LLMs) often struggle to produce accurate responses on the first attempt for complex reasoning tasks like code generation. Prior research tackles this challenge by generating multiple candidate solutions and validating them with LLM-generated unit tests. The execution results of unit tests serve as reward signals to identify correct solutions. As LLMs always confidently make mistakes, these unit tests are not reliable, thereby diminishing the quality of reward signals. Motivated by the observation that scaling the number of solutions improves LLM performance, we explore the impact of scaling unit tests to enhance reward signal quality. Our pioneer experiment reveals a positive correlation between the number of unit tests and reward signal quality, with greater benefits observed in more challenging problems. Based on these insights, we propose CodeRM-8B, a lightweight yet effective unit test generator that enables efficient and high-quality unit test scaling. Additionally, we implement a dynamic scaling mechanism that adapts the number of unit tests based on problem difficulty, further improving efficiency. Experimental results show that our approach significantly improves performance across various models on three benchmarks (e.g., with gains of 18.43% for Llama3-8B and 3.42% for GPT-4o-mini on HumanEval Plus).", "authors": ["Zeyao Ma", "Xiaokang Zhang", "Jing Zhang", "Jifan Yu", "Sijia Luo", "Jie Tang"], "categories": ["cs.CL", "cs.SE"], "fields_of_study": ["Computer Science"], "published_date": "2025-01-02", "url": "https://arxiv.org/abs/2501.01054", "pdf_url": "https://arxiv.org/pdf/2501.01054v1", "arxiv_id": "2501.01054", "doi": "10.48550/arXiv.2501.01054", "citation_count": 25, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3537}
{"id": "5519887027ab05cd5c3c61f839947cdbcc2bfed921181c38dfbee5dc262e0cbe", "sources": ["arxiv", "semantic_scholar"], "title": "Entropy-Regularized Process Reward Model", "abstract": "Large language models (LLMs) have shown promise in performing complex multi-step reasoning, yet they continue to struggle with mathematical reasoning, often making systematic errors. A promising solution is reinforcement learning (RL) guided by reward models, particularly those focusing on process rewards, which score each intermediate step rather than solely evaluating the final outcome. This approach is more effective at guiding policy models towards correct reasoning trajectories. In this work, we propose an entropy-regularized process reward model (ER-PRM) that integrates KL-regularized Markov Decision Processes (MDP) to balance policy optimization with the need to prevent the policy from shifting too far from its initial distribution. We derive a novel reward construction method based on the theoretical results. Our theoretical analysis shows that we could derive the optimal reward model from the initial policy sampling. Our empirical experiments on the MATH and GSM8K benchmarks demonstrate that ER-PRM consistently outperforms existing process reward models, achieving 1% improvement on GSM8K and 2-3% improvement on MATH under best-of-N evaluation, and more than 1% improvement under RLHF. These results highlight the efficacy of entropy-regularization in enhancing LLMs' reasoning capabilities.", "authors": ["Hanning Zhang", "Pengcheng Wang", "Shizhe Diao", "Yong Lin", "Rui Pan", "Hanze Dong", "Dylan Zhang", "Pavlo Molchanov", "Tong Zhang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-15", "url": "https://arxiv.org/abs/2412.11006", "pdf_url": "https://arxiv.org/pdf/2412.11006v2", "arxiv_id": "2412.11006", "doi": "10.48550/arXiv.2412.11006", "citation_count": 25, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3537}
{"id": "7e664d56cda5f5b1e8c609cea812712765d35e8359f430f3038b442060cd98a4", "sources": ["arxiv", "semantic_scholar"], "title": "Sail into the Headwind: Alignment via Robust Rewards and Dynamic Labels against Reward Hacking", "abstract": "Aligning AI systems with human preferences typically suffers from the infamous reward hacking problem, where optimization of an imperfect reward model leads to undesired behaviors. In this paper, we investigate reward hacking in offline preference optimization, which aims to improve an initial model using a preference dataset. We identify two types of reward hacking stemming from statistical fluctuations in the dataset: Type I Reward Hacking due to subpar choices appearing more favorable, and Type II Reward Hacking due to decent choices appearing less favorable. We prove that many (mainstream or theoretical) preference optimization methods suffer from both types of reward hacking. To mitigate Type I Reward Hacking, we propose POWER, a new preference optimization method that combines Guiasu's weighted entropy with a robust reward maximization objective. POWER enjoys finite-sample guarantees under general function approximation, competing with the best covered policy in the data. To mitigate Type II Reward Hacking, we analyze the learning dynamics of preference optimization and develop a novel technique that dynamically updates preference labels toward certain \"stationary labels\", resulting in diminishing gradients for untrustworthy samples. Empirically, POWER with dynamic labels (POWER-DL) consistently outperforms state-of-the-art methods on alignment benchmarks, achieving improvements of up to 13.0 points on AlpacaEval 2.0 and 11.5 points on Arena-Hard over DPO, while also improving or maintaining performance on downstream tasks such as mathematical reasoning. Strong theoretical guarantees and empirical results demonstrate the promise of POWER-DL in mitigating reward hacking.", "authors": ["Paria Rashidinejad", "Yuandong Tian"], "categories": ["cs.LG", "cs.AI", "math.OC", "math.ST", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-12-12", "url": "https://arxiv.org/abs/2412.09544", "pdf_url": "https://arxiv.org/pdf/2412.09544v1", "arxiv_id": "2412.09544", "doi": "10.48550/arXiv.2412.09544", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2785}
{"id": "c0bb9e34631a7e1e58531da88fa3d883bd4c800126860eac111cc741a90f8f81", "sources": ["arxiv", "semantic_scholar"], "title": "T-REG: Preference Optimization with Token-Level Reward Regularization", "abstract": "Reinforcement learning from human feedback (RLHF) has been crucial in aligning large language models (LLMs) with human values. Traditionally, RLHF involves generating responses to a query and using a reward model to assign a reward to the entire response. However, this approach faces challenges due to its reliance on a single, sparse reward, which makes it challenging for the model to identify which parts of the sequence contribute most significantly to the final reward. Recent methods have attempted to address this limitation by introducing token-level rewards. However, these methods often rely on either a trained credit assignment model or AI annotators, raising concerns about the quality and reliability of the rewards. In this paper, we propose token-level reward regularization (T-REG), a novel approach that leverages both sequence-level and token-level rewards for preference optimization. Harnessing the self-refinement capabilities of LLMs, our method uses contrastive prompting to enable LLMs to self-generate token-level rewards. These self-generated rewards then act as reward regularization, guiding the model to more effectively distribute sequence-level rewards across tokens. This facilitates better token-level credit assignment and enhances alignment performance. Experiments on the instruction following benchmarks, including Alpaca Eval 2 and Arena-Hard, show that our method consistently outperforms baseline methods by up to 3.8% and 4.4%, respectively. We will release the code and models at https://github.com/wzhouad/T-REG.", "authors": ["Wenxuan Zhou", "Shujian Zhang", "Lingxiao Zhao", "Tao Meng"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-12-03", "url": "https://arxiv.org/abs/2412.02685", "pdf_url": "https://arxiv.org/pdf/2412.02685v1", "arxiv_id": "2412.02685", "doi": "10.48550/arXiv.2412.02685", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/wzhouad/T-REG", "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "2c1357194c5e573a3f79fef011a91e596a3e4ede021f37f474a1e088fc239e5b", "sources": ["arxiv", "semantic_scholar"], "title": "Self-Generated Critiques Boost Reward Modeling for Language Models", "abstract": "Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy.", "authors": ["Yue Yu", "Zhengxing Chen", "Aston Zhang", "Liang Tan", "Chenguang Zhu", "Richard Yuanzhe Pang", "Yundi Qian", "Xuewei Wang", "Suchin Gururangan", "Chao Zhang", "Melanie Kambadur", "Dhruv Mahajan", "Rui Hou"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-25", "url": "https://arxiv.org/abs/2411.16646", "pdf_url": "https://arxiv.org/pdf/2411.16646v3", "arxiv_id": "2411.16646", "doi": "10.48550/arXiv.2411.16646", "citation_count": 71, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.4643}
{"id": "efe0c5429481db04d670e5f778a7f2b6efa81c163fb97d77d3343071f7989c2a", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Fine-Tuning Two-Step Diffusion Models via Learning Differentiable Latent-Space Surrogate Reward", "abstract": "Recent research has shown that fine-tuning diffusion models (DMs) with arbitrary rewards, including non-differentiable ones, is feasible with reinforcement learning (RL) techniques, enabling flexible model alignment. However, applying existing RL methods to step-distilled DMs is challenging for ultra-fast ($\\le2$-step) image generation. Our analysis suggests several limitations of policy-based RL methods such as PPO or DPO toward this goal. Based on the insights, we propose fine-tuning DMs with learned differentiable surrogate rewards. Our method, named LaSRO, learns surrogate reward models in the latent space of SDXL to convert arbitrary rewards into differentiable ones for effective reward gradient guidance. LaSRO leverages pre-trained latent DMs for reward modeling and tailors reward optimization for $\\le2$-step image generation with efficient off-policy exploration. LaSRO is effective and stable for improving ultra-fast image generation with different reward objectives, outperforming popular RL methods including DDPO and Diffusion-DPO. We further show LaSRO's connection to value-based RL, providing theoretical insights. See our webpage \\href{https://sites.google.com/view/lasro}{here}.", "authors": ["Zhiwei Jia", "Yuesong Nan", "Huixi Zhao", "Gengdai Liu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-11-22", "url": "https://arxiv.org/abs/2411.15247", "pdf_url": "https://arxiv.org/pdf/2411.15247v3", "arxiv_id": "2411.15247", "doi": "10.1109/CVPR52734.2025.01205", "citation_count": 17, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.3138}
{"id": "1465189b937e1d294c7d94a6a4a304179bc95c7eb55a679f7622383a6d2ac2bd", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Modeling with Weak Supervision for Language Models", "abstract": "Recent advancements in large language models (LLMs) have led to their increased application across various tasks, with reinforcement learning from human feedback (RLHF) being a crucial part of their training to align responses with user intentions. In the RLHF process, a reward model is trained using responses preferences determined by human labelers or AI systems, which then refines the LLM through reinforcement learning. This work introduces weak supervision as a strategy to extend RLHF datasets and enhance reward model performance. Weak supervision employs noisy or imprecise data labeling, reducing reliance on expensive manually labeled data. By analyzing RLHF datasets to identify heuristics that correlate with response preference, we wrote simple labeling functions and then calibrated a label model to weakly annotate unlabeled data. Our evaluation show that while weak supervision significantly benefits smaller datasets by improving reward model performance, its effectiveness decreases with larger, originally labeled datasets. Additionally, using an LLM to generate and then weakly label responses offers a promising method for extending preference data.", "authors": ["Ben Hauptvogel", "Malte Ostendorff", "Georg Rehm", "Sebastian Möller"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-28", "url": "https://arxiv.org/abs/2410.20869", "pdf_url": "https://arxiv.org/pdf/2410.20869v1", "arxiv_id": "2410.20869", "doi": "10.1109/ACDSA65407.2025.11166626", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "b3d76883701b8570fe6f2f3a67cc907bfedc9bea22738487a8bd9097b1708cb0", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Simple Sum of Delayed Rewards: Non-Markovian Reward Modeling for Reinforcement Learning", "abstract": "Reinforcement Learning (RL) empowers agents to acquire various skills by learning from reward signals. Unfortunately, designing high-quality instance-level rewards often demands significant effort. An emerging alternative, RL with delayed reward, focuses on learning from rewards presented periodically, which can be obtained from human evaluators assessing the agent's performance over sequences of behaviors. However, traditional methods in this domain assume the existence of underlying Markovian rewards and that the observed delayed reward is simply the sum of instance-level rewards, both of which often do not align well with real-world scenarios. In this paper, we introduce the problem of RL from Composite Delayed Reward (RLCoDe), which generalizes traditional RL from delayed rewards by eliminating the strong assumption. We suggest that the delayed reward may arise from a more complex structure reflecting the overall contribution of the sequence. To address this problem, we present a framework for modeling composite delayed rewards, using a weighted sum of non-Markovian components to capture the different contributions of individual steps. Building on this framework, we propose Composite Delayed Reward Transformer (CoDeTr), which incorporates a specialized in-sequence attention mechanism to effectively model these contributions. We conduct experiments on challenging locomotion tasks where the agent receives delayed rewards computed from composite functions of observable step rewards. The experimental results indicate that CoDeTr consistently outperforms baseline methods across evaluated metrics. Additionally, we demonstrate that it effectively identifies the most significant time steps within the sequence and accurately predicts rewards that closely reflect the environment feedback.", "authors": ["Yuting Tang", "Xin-Qiang Cai", "Jing-Cheng Pang", "Qiyu Wu", "Yao-Xiang Ding", "Masashi Sugiyama"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-26", "url": "https://arxiv.org/abs/2410.20176", "pdf_url": "https://arxiv.org/pdf/2410.20176v1", "arxiv_id": "2410.20176", "doi": "10.48550/arXiv.2410.20176", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "943cb7d54a6b12a48d4caf2de506344228b4334704629091e908caa19623306b", "sources": ["arxiv", "semantic_scholar"], "title": "Skywork-Reward: Bag of Tricks for Reward Modeling in LLMs", "abstract": "In this report, we introduce a collection of methods to enhance reward modeling for LLMs, focusing specifically on data-centric techniques. We propose effective data selection and filtering strategies for curating high-quality open-source preference datasets, culminating in the Skywork-Reward data collection, which contains only 80K preference pairs -- significantly smaller than existing datasets. Using this curated dataset, we developed the Skywork-Reward model series -- Skywork-Reward-Gemma-27B and Skywork-Reward-Llama-3.1-8B -- with the former currently holding the top position on the RewardBench leaderboard. Notably, our techniques and datasets have directly enhanced the performance of many top-ranked models on RewardBench, highlighting the practical impact of our contributions in real-world preference learning applications.", "authors": ["Chris Yuhao Liu", "Liang Zeng", "Jiacai Liu", "Rui Yan", "Jujie He", "Chaojie Wang", "Shuicheng Yan", "Yang Liu", "Yahui Zhou"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-24", "url": "https://arxiv.org/abs/2410.18451", "pdf_url": "https://arxiv.org/pdf/2410.18451v1", "arxiv_id": "2410.18451", "doi": "10.48550/arXiv.2410.18451", "citation_count": 305, "influential_citation_count": 33, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.7657}
{"id": "4430b109d5b9558073d70a23d224d1b02f479dd85cd4cc1999f24494837de309", "sources": ["arxiv", "semantic_scholar"], "title": "Optimal Design for Reward Modeling in RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has become a popular approach to align language models (LMs) with human preferences. This method involves collecting a large dataset of human pairwise preferences across various text generations and using it to infer (implicitly or explicitly) a reward model. Numerous methods have been proposed to learn the reward model and align a LM with it. However, the costly process of collecting human preferences has received little attention and could benefit from theoretical insights. This paper addresses this issue and aims to formalize the reward training model in RLHF. We frame the selection of an effective dataset as a simple regret minimization task, using a linear contextual dueling bandit method. Given the potentially large number of arms, this approach is more coherent than the best-arm identification setting. We then propose an offline framework for solving this problem. Under appropriate assumptions - linearity of the reward model in the embedding space, and boundedness of the reward parameter - we derive bounds on the simple regret. Finally, we provide a lower bound that matches our upper bound up to constant and logarithmic terms. To our knowledge, this is the first theoretical contribution in this area to provide an offline approach as well as worst-case guarantees.", "authors": ["Antoine Scheid", "Etienne Boursier", "Alain Durmus", "Michael I. Jordan", "Pierre Ménard", "Eric Moulines", "Michal Valko"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-10-22", "url": "https://arxiv.org/abs/2410.17055", "pdf_url": "https://arxiv.org/pdf/2410.17055v2", "arxiv_id": "2410.17055", "doi": "10.48550/arXiv.2410.17055", "citation_count": 23, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3451}
{"id": "7b2194c4187c14c3596e95e5ed47975692c824c4e4d2ec10bca85d60f32d067a", "sources": ["arxiv", "semantic_scholar"], "title": "How to Evaluate Reward Models for RLHF", "abstract": "We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback). The gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance. However, this process is prohibitively expensive. To address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks. These proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics across 12 domains. To investigate which reward model metrics are most correlated to gold-standard RLHF outcomes, we launch an end-to-end RLHF experiment on a large-scale crowdsourced human preference platform to view real reward model downstream performance as ground truth. Ultimately, we compile our data and findings into Preference Proxy Evaluations (PPE), the first reward model benchmark explicitly linked to post-RLHF real-world human preference performance, which we open-source for public use and further development. Our code and evaluations can be found at https://github.com/lmarena/PPE .", "authors": ["Evan Frick", "Tianle Li", "Connor Chen", "Wei-Lin Chiang", "Anastasios N. Angelopoulos", "Jiantao Jiao", "Banghua Zhu", "Joseph E. Gonzalez", "Ion Stoica"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-18", "url": "https://arxiv.org/abs/2410.14872", "pdf_url": "https://arxiv.org/pdf/2410.14872v2", "arxiv_id": "2410.14872", "doi": "10.48550/arXiv.2410.14872", "citation_count": 75, "influential_citation_count": 21, "has_code": true, "code_url": "https://github.com/lmarena/PPE", "venue": "International Conference on Learning Representations", "quality_score": 0.6712}
{"id": "f286521b8bdfaa75a762ecd313a89b99708a4667a39a517685f14f0d51ba8d4c", "sources": ["arxiv", "semantic_scholar"], "title": "A Large Language Model-Driven Reward Design Framework via Dynamic Feedback for Reinforcement Learning", "abstract": "Large Language Models (LLMs) have shown significant potential in designing reward functions for Reinforcement Learning (RL) tasks. However, obtaining high-quality reward code often involves human intervention, numerous LLM queries, or repetitive RL training. To address these issues, we propose CARD, a LLM-driven Reward Design framework that iteratively generates and improves reward function code. Specifically, CARD includes a Coder that generates and verifies the code, while a Evaluator provides dynamic feedback to guide the Coder in improving the code, eliminating the need for human feedback. In addition to process feedback and trajectory feedback, we introduce Trajectory Preference Evaluation (TPE), which evaluates the current reward function based on trajectory preferences. If the code fails the TPE, the Evaluator provides preference feedback, avoiding RL training at every iteration and making the reward function better aligned with the task objective. Empirical results on Meta-World and ManiSkill2 demonstrate that our method achieves an effective balance between task performance and token efficiency, outperforming or matching the baselines across all tasks. On 10 out of 12 tasks, CARD shows better or comparable performance to policies trained with expert-designed rewards, and our method even surpasses the oracle on 3 tasks.", "authors": ["Shengjie Sun", "Runze Liu", "Jiafei Lyu", "Jing-Wen Yang", "Liangpeng Zhang", "Xiu Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-18", "url": "https://arxiv.org/abs/2410.14660", "pdf_url": "https://arxiv.org/pdf/2410.14660v1", "arxiv_id": "2410.14660", "doi": "10.48550/arXiv.2410.14660", "citation_count": 36, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Knowledge-Based Systems", "quality_score": 0.3921}
{"id": "37ea869a2c4166035f8c5253a02e9fe5664643a1e027e8a92ae84b3a1a92350e", "sources": ["arxiv", "semantic_scholar"], "title": "CREAM: Consistency Regularized Self-Rewarding Language Models", "abstract": "Recent self-rewarding large language models (LLM) have successfully applied LLM-as-a-Judge to iteratively improve the alignment performance without the need of human annotations for preference data. These methods commonly utilize the same LLM to act as both the policy model (which generates responses) and the reward model (which scores and ranks those responses). The ranked responses are then used as preference pairs to train the LLM via direct alignment technologies (e.g. DPO). However, it is noteworthy that throughout this process, there is no guarantee of accuracy in the rewarding and ranking, which is critical for ensuring accurate rewards and high-quality preference data. Empirical results from relatively small LLMs (e.g., 7B parameters) also indicate that improvements from self-rewarding may diminish after several iterations in certain situations, which we hypothesize is due to accumulated bias in the reward system. This bias can lead to unreliable preference data for training the LLM. To address this issue, we first formulate and analyze the generalized iterative preference fine-tuning framework for self-rewarding language model. We then introduce the regularization to this generalized framework to mitigate the overconfident preference labeling in the self-rewarding process. Based on this theoretical insight, we propose a Consistency Regularized sElf-rewarding lAnguage Model (CREAM) that leverages the consistency of rewards across different iterations to regularize the self-rewarding training, helping the model to learn from more reliable preference data. With this explicit regularization, our empirical results demonstrate the superiority of CREAM in improving both reward consistency and alignment performance. The code is publicly available at https://github.com/Raibows/CREAM.", "authors": ["Zhaoyang Wang", "Weilei He", "Zhiyuan Liang", "Xuchao Zhang", "Chetan Bansal", "Ying Wei", "Weitong Zhang", "Huaxiu Yao"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-16", "url": "https://arxiv.org/abs/2410.12735", "pdf_url": "https://arxiv.org/pdf/2410.12735v5", "arxiv_id": "2410.12735", "doi": "10.48550/arXiv.2410.12735", "citation_count": 36, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/Raibows/CREAM", "venue": "International Conference on Learning Representations", "quality_score": 0.3921}
{"id": "2143ebac6a54872154548ab0b9870cf95a675c5ba7fcb1852bd5f2e78b20598d", "sources": ["arxiv", "semantic_scholar"], "title": "AlphaDPO: Adaptive Reward Margin for Direct Preference Optimization", "abstract": "Aligning large language models (LLMs) with human values and intentions is crucial for their utility, honesty, and safety. Reinforcement learning from human feedback (RLHF) is a popular approach to achieve this alignment, but it faces challenges in computational efficiency and training stability. Recent methods like Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO) have proposed offline alternatives to RLHF, simplifying the process by reparameterizing the reward function. However, DPO depends on a potentially suboptimal reference model, and SimPO's assumption of a fixed target reward margin may lead to suboptimal decisions in diverse data settings. In this work, we propose $α$-DPO, an adaptive preference optimization algorithm designed to address these limitations by introducing a dynamic reward margin. Specifically, $α$-DPO employs an adaptive preference distribution, balancing the policy model and the reference model to achieve personalized reward margins. We provide theoretical guarantees for $α$-DPO, demonstrating its effectiveness as a surrogate optimization objective and its ability to balance alignment and diversity through KL divergence control. Empirical evaluations on AlpacaEval 2 and Arena-Hard show that $α$-DPO consistently outperforms DPO and SimPO across various model settings, establishing it as a robust approach for fine-tuning LLMs. Our method achieves significant improvements in win rates, highlighting its potential as a powerful tool for LLM alignment. The code is available at https://github.com/junkangwu/alpha-DPO", "authors": ["Junkang Wu", "Xue Wang", "Zhengyi Yang", "Jiancan Wu", "Jinyang Gao", "Bolin Ding", "Xiang Wang", "Xiangnan He"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-14", "url": "https://arxiv.org/abs/2410.10148", "pdf_url": "https://arxiv.org/pdf/2410.10148v4", "arxiv_id": "2410.10148", "doi": null, "citation_count": 22, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/junkangwu/alpha-DPO", "venue": "International Conference on Machine Learning", "quality_score": 0.3404}
{"id": "27f2e2b90347828aabb00a2e9b2a805623660f2f86f296bcbc89ee6047c67b4c", "sources": ["arxiv", "semantic_scholar"], "title": "Burning RED: Unlocking Subtask-Driven Reinforcement Learning and Risk-Awareness in Average-Reward Markov Decision Processes", "abstract": "Average-reward Markov decision processes (MDPs) provide a foundational framework for sequential decision-making under uncertainty. However, average-reward MDPs have remained largely unexplored in reinforcement learning (RL) settings, with the majority of RL-based efforts having been allocated to discounted MDPs. In this work, we study a unique structural property of average-reward MDPs and utilize it to introduce Reward-Extended Differential (or RED) reinforcement learning: a novel RL framework that can be used to effectively and efficiently solve various learning objectives, or subtasks, simultaneously in the average-reward setting. We introduce a family of RED learning algorithms for prediction and control, including proven-convergent algorithms for the tabular case. We then showcase the power of these algorithms by demonstrating how they can be used to learn a policy that optimizes, for the first time, the well-known conditional value-at-risk (CVaR) risk measure in a fully-online manner, without the use of an explicit bi-level optimization scheme or an augmented state-space.", "authors": ["Juan Sebastian Rojas", "Chi-Guhn Lee"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-14", "url": "https://arxiv.org/abs/2410.10578", "pdf_url": "https://arxiv.org/pdf/2410.10578v11", "arxiv_id": "2410.10578", "doi": "10.48550/arXiv.2410.10578", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "efe4288a378033aca8da9ad5b39e48213137aaba0cac8973e018a5997359a23b", "sources": ["arxiv", "semantic_scholar"], "title": "Taming Overconfidence in LLMs: Reward Calibration in RLHF", "abstract": "Language model calibration refers to the alignment between the confidence of the model and the actual performance of its responses. While previous studies point out the overconfidence phenomenon in Large Language Models (LLMs) and show that LLMs trained with Reinforcement Learning from Human Feedback (RLHF) are overconfident with a more sharpened output probability, in this study, we reveal that RLHF tends to lead models to express verbalized overconfidence in their own responses. We investigate the underlying cause of this overconfidence and demonstrate that reward models used for Proximal Policy Optimization (PPO) exhibit inherent biases towards high-confidence scores regardless of the actual quality of responses. Building upon this insight, we propose two PPO variants: PPO-M: PPO with Calibrated Reward Modeling and PPO-C: PPO with Calibrated Reward Calculation. PPO-M integrates explicit confidence scores in reward model training, which calibrates reward models to better capture the alignment between response quality and verbalized confidence. PPO-C adjusts the reward score during PPO based on the difference between the current reward and the exponential average of past rewards. Both PPO-M and PPO-C can be seamlessly integrated into the current PPO pipeline and do not require additional golden labels. We evaluate our methods on both Llama3-8B and Mistral-7B across six diverse datasets including multiple-choice and open-ended generation. Experimental results demonstrate that both of our methods can reduce calibration error and maintain performance comparable to standard PPO. We further show that they could preserve model capabilities in open-ended conversational settings.", "authors": ["Jixuan Leng", "Chengsong Huang", "Banghua Zhu", "Jiaxin Huang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-13", "url": "https://arxiv.org/abs/2410.09724", "pdf_url": "https://arxiv.org/pdf/2410.09724v2", "arxiv_id": "2410.09724", "doi": "10.48550/arXiv.2410.09724", "citation_count": 89, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.5}
{"id": "8fddda6eef8d02e9572ec6cd62da8d67dd586321cf2812a0a505285958bbef20", "sources": ["arxiv", "semantic_scholar"], "title": "Simultaneous Reward Distillation and Preference Learning: Get You a Language Model Who Can Do Both", "abstract": "Traditional RLHF-based LLM alignment methods explicitly maximize the expected rewards from a separate reward model. More recent supervised alignment methods like Direct Preference Optimization (DPO) circumvent this phase to avoid problems including model drift and reward overfitting. Although popular due to its simplicity, DPO and similar direct alignment methods which rely heavily on the Bradley-Terry-based pairwise preference formulation can still lead to degenerate policies when challenged by non-deterministic or noisy preference labels, for example human scoring of two candidate outputs with low confidence. This paper introduces DRDO (Direct Reward Distillation and policy-Optimization), which simultaneously models rewards and preferences to avoid such degeneracy. DRDO directly mimics rewards assigned by an oracle while learning human preferences with a novel preference likelihood formulation. Results on the Ultrafeedback and TL;DR datasets demonstrate that DRDO-trained policies surpass methods such as DPO and e-DPO in terms of expected rewards and are more robust, on average, to noisy preference signals as well as out-of-distribution (OOD) settings.", "authors": ["Abhijnan Nath", "Changsoo Jung", "Ethan Seefried", "Nikhil Krishnaswamy"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-11", "url": "https://arxiv.org/abs/2410.08458", "pdf_url": "https://arxiv.org/pdf/2410.08458v2", "arxiv_id": "2410.08458", "doi": "10.48550/arXiv.2410.08458", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "87c31d6d06bc7e1f9238cc9e572b1d1d87a8ae8598e38810d7102cf9408fb1b9", "sources": ["arxiv", "semantic_scholar"], "title": "Reward-Augmented Data Enhances Direct Preference Alignment of LLMs", "abstract": "Preference alignment in Large Language Models (LLMs) has significantly improved their ability to adhere to human instructions and intentions. However, existing direct alignment algorithms primarily focus on relative preferences and often overlook the qualitative aspects of responses, despite having access to preference data that includes reward scores from judge models during AI feedback. Striving to maximize the implicit reward gap between the chosen and the slightly inferior rejected responses can cause overfitting and unnecessary unlearning of the high-quality rejected responses. The unawareness of the reward scores also drives the LLM to indiscriminately favor the low-quality chosen responses and fail to generalize to optimal responses that are sparse in data. To overcome these shortcomings, our study introduces reward-conditioned LLM policies that discern and learn from the entire spectrum of response quality within the dataset, helping extrapolate to more optimal regions. We propose an effective yet simple data relabeling method that conditions the preference pairs on quality scores to construct a reward-augmented dataset. The experiments across various benchmarks and diverse models demonstrate that our approach consistently boosts DPO by a considerable margin. Through comprehensive ablation studies, we demonstrate that our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere data expansion. Our code is available at https://github.com/shenao-zhang/reward-augmented-preference.", "authors": ["Shenao Zhang", "Zhihan Liu", "Boyi Liu", "Yufeng Zhang", "Yingxiang Yang", "Yongfei Liu", "Liyu Chen", "Tao Sun", "Zhaoran Wang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.08067", "pdf_url": "https://arxiv.org/pdf/2410.08067v6", "arxiv_id": "2410.08067", "doi": "10.48550/arXiv.2410.08067", "citation_count": 10, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/shenao-zhang/reward-augmented-preference", "venue": "International Conference on Machine Learning", "quality_score": 0.2603}
{"id": "a0e8050ef1e753c11b17c1ca56149143b558b0fc2b98ed356953a4b79eec5b43", "sources": ["arxiv", "semantic_scholar"], "title": "GenARM: Reward Guided Generation with Autoregressive Reward Model for Test-time Alignment", "abstract": "Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model--a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining. Our project page is available at: https://genarm.github.io.", "authors": ["Yuancheng Xu", "Udari Madhushani Sehwag", "Alec Koppel", "Sicheng Zhu", "Bang An", "Furong Huang", "Sumitra Ganesh"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-10", "url": "https://arxiv.org/abs/2410.08193", "pdf_url": "https://arxiv.org/pdf/2410.08193v5", "arxiv_id": "2410.08193", "doi": "10.48550/arXiv.2410.08193", "citation_count": 49, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4771}
{"id": "7ebe594b678b124783f3c1b499fdd42ad22f680ea7dc2cdeda5d5a4bd86e91a2", "sources": ["arxiv", "semantic_scholar"], "title": "The Accuracy Paradox in RLHF: When Better Reward Models Don't Yield Better Language Models", "abstract": "Reinforcement Learning from Human Feedback significantly enhances Natural Language Processing by aligning language models with human expectations. A critical factor in this alignment is the strength of reward models used during training. This study explores whether stronger reward models invariably lead to better language models. In this paper, through experiments on relevance, factuality, and completeness tasks using the QA-FEEDBACK dataset and reward models based on Longformer, we uncover a surprising paradox: language models trained with moderately accurate reward models outperform those guided by highly accurate ones. This challenges the widely held belief that stronger reward models always lead to better language models, and opens up new avenues for future research into the key factors driving model performance and how to choose the most suitable reward models. Code and additional details are available at https://github.com/EIT-NLP/AccuracyParadox-RLHF.", "authors": ["Yanjun Chen", "Dawei Zhu", "Yirong Sun", "Xinghao Chen", "Wei Zhang", "Xiaoyu Shen"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-09", "url": "https://arxiv.org/abs/2410.06554", "pdf_url": "https://arxiv.org/pdf/2410.06554v2", "arxiv_id": "2410.06554", "doi": "10.48550/arXiv.2410.06554", "citation_count": 15, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/EIT-NLP/AccuracyParadox-RLHF", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.301}
{"id": "763b67fdae837dcfa4a75566e7b910a3b7689d1745fdc0fc86044cf43f3fd89e", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Learning From Preference With Ties", "abstract": "Reward learning plays a pivotal role in Reinforcement Learning from Human Feedback (RLHF), ensuring the alignment of language models. The Bradley-Terry (BT) model stands as the prevalent choice for capturing human preferences from datasets containing pairs of chosen and rejected responses. In preference modeling, the focus is not on absolute values but rather on the reward difference between chosen and rejected responses, referred to as preference strength. Thus, precise evaluation of preference strength holds paramount importance in preference modeling. However, an easily overlooked factor significantly affecting preference strength measurement is that human attitudes towards two responses may not solely indicate a preference for one over the other and ties are also a common occurrence. To address this, we propose the adoption of the generalized Bradley-Terry model -- the Bradley-Terry model with ties (BTT) -- to accommodate tied preferences, thus leveraging additional information. We prove that even with the access to the true distributions of prompt and response, disregarding ties can lead to a notable bias in preference strength measurement. Comprehensive experiments further validate the advantages of incorporating ties in preference modeling. Notably, fine-tuning with BTT significantly outperforms fine-tuning with BT on synthetic preference datasets with ties, labeled by state-of-the-art open-source LLMs.", "authors": ["Jinsong Liu", "Dongdong Ge", "Ruihao Zhu"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-05", "url": "https://arxiv.org/abs/2410.05328", "pdf_url": "https://arxiv.org/pdf/2410.05328v1", "arxiv_id": "2410.05328", "doi": "10.48550/arXiv.2410.05328", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "fb4e028da8f56385b92ca788f686f2b9851b83696491042cb5f7c23c8bcf9807", "sources": ["arxiv", "semantic_scholar"], "title": "From Reward Shaping to Q-Shaping: Achieving Unbiased Learning with LLM-Guided Knowledge", "abstract": "Q-shaping is an extension of Q-value initialization and serves as an alternative to reward shaping for incorporating domain knowledge to accelerate agent training, thereby improving sample efficiency by directly shaping Q-values. This approach is both general and robust across diverse tasks, allowing for immediate impact assessment while guaranteeing optimality. We evaluated Q-shaping across 20 different environments using a large language model (LLM) as the heuristic provider. The results demonstrate that Q-shaping significantly enhances sample efficiency, achieving a \\textbf{16.87\\%} improvement over the best baseline in each environment and a \\textbf{253.80\\%} improvement compared to LLM-based reward shaping methods. These findings establish Q-shaping as a superior and unbiased alternative to conventional reward shaping in reinforcement learning.", "authors": ["Xiefeng Wu"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-02", "url": "https://arxiv.org/abs/2410.01458", "pdf_url": "https://arxiv.org/pdf/2410.01458v1", "arxiv_id": "2410.01458", "doi": "10.48550/arXiv.2410.01458", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "d76ba88b123f013738c576466e2717cc29270e3f475e6ce27d2c4178f224b45e", "sources": ["arxiv", "semantic_scholar"], "title": "Generative Reward Models", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has greatly improved the performance of modern Large Language Models (LLMs). The RLHF process is resource-intensive and technically challenging, generally requiring a large collection of human preference labels over model-generated outputs. Reinforcement Learning from AI Feedback (RLAIF) addresses this data collection challenge by leveraging synthetic preferences generated by an LLM. However, recent work has shown that synthetic preferences labels may not align well with human preference judgments. To address this, we propose a hybrid approach that unifies RLHF and RLAIF methodologies. We introduce GenRM, an iterative algorithm that trains an LLM on self-generated reasoning traces, leading to synthetic preference labels matching human preference judgments. Empirically, we show that zero-shot LLM-based judgments under-perform compared to Bradley-Terry reward models on in-distribution tasks (between 9-36%). In contrast, GenRM achieves in-distribution accuracy comparable to Bradley-Terry models, while significantly outperforming them on out-of-distribution tasks (between 10-45%). Moreover, GenRM surpasses the performance of using LLMs as judges on both in-distribution (by 9-31%) and out-of-distribution tasks (by 2- 6%). Our results show that combining the strengths of RLHF and RLAIF offers a promising approach for improving the quality of synthetic preference labels.", "authors": ["Dakota Mahan", "Duy Van Phung", "Rafael Rafailov", "Chase Blagden", "Nathan Lile", "Louis Castricato", "Jan-Philipp Fränken", "Chelsea Finn", "Alon Albalak"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-02", "url": "https://arxiv.org/abs/2410.12832", "pdf_url": "https://arxiv.org/pdf/2410.12832v1", "arxiv_id": "2410.12832", "doi": "10.48550/arXiv.2410.12832", "citation_count": 112, "influential_citation_count": 10, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5207}
{"id": "1d5deb5b1e96dafc009817b578b44aa6fcdbdbf052d99a8790339a1c8eb7b43a", "sources": ["arxiv", "semantic_scholar"], "title": "Evaluating Robustness of Reward Models for Mathematical Reasoning", "abstract": "Reward models are key in reinforcement learning from human feedback (RLHF) systems, aligning the model behavior with human preferences. Particularly in the math domain, there have been plenty of studies using reward models to align policies for improving reasoning capabilities. Recently, as the importance of reward models has been emphasized, RewardBench is proposed to understand their behavior. However, we figure out that the math subset of RewardBench has different representations between chosen and rejected completions, and relies on a single comparison, which may lead to unreliable results as it only see an isolated case. Therefore, it fails to accurately present the robustness of reward models, leading to a misunderstanding of its performance and potentially resulting in reward hacking. In this work, we introduce a new design for reliable evaluation of reward models, and to validate this, we construct RewardMATH, a benchmark that effectively represents the robustness of reward models in mathematical reasoning tasks. We demonstrate that the scores on RewardMATH strongly correlate with the results of optimized policy and effectively estimate reward overoptimization, whereas the existing benchmark shows almost no correlation. The results underscore the potential of our design to enhance the reliability of evaluation, and represent the robustness of reward model. We make our code and data publicly available.", "authors": ["Sunghwan Kim", "Dongjin Kang", "Taeyoon Kwon", "Hyungjoo Chae", "Jungsoo Won", "Dongha Lee", "Jinyoung Yeo"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-02", "url": "https://arxiv.org/abs/2410.01729", "pdf_url": "https://arxiv.org/pdf/2410.01729v1", "arxiv_id": "2410.01729", "doi": "10.48550/arXiv.2410.01729", "citation_count": 17, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3138}
{"id": "d57856bde639c17b697db94cec73467a63400f97793c7326e9b69ac5ead1d312", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Scalar Reward Model: Learning Generative Judge from Preference Data", "abstract": "Learning from preference feedback is a common practice for aligning large language models~(LLMs) with human value. Conventionally, preference data is learned and encoded into a scalar reward model that connects a value head with an LLM to produce a scalar score as preference or reward. However, scalar models lack interpretability and are known to be susceptible to biases in datasets. This paper investigates leveraging the generation capability of LLMs to address both limitations in one shot. Specifically, we prompt the pre-trained LLM to generate positive and negative judgments, both supported with rationales in natural language form. The self-generated contrastive judgment pairs are used to train the generative judge with Direct Preference Optimization (DPO). This proposal of training the generative Judge using self-generated Contrastive judgments (Con-J) ensures natural interpretability due to the generated rationales together with the judgments, as well as high robustness against bias without the need for an additional reward head. Experimental results show that the performance of Con-J is comparable to the scalar reward model trained on the same collection of preference data, and demonstrate its superior interpretability and robustness in encoding human preferences.", "authors": ["Ziyi Ye", "Xiangsheng Li", "Qiuchi Li", "Qingyao Ai", "Yujia Zhou", "Wei Shen", "Dong Yan", "Yiqun Liu"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-10-01", "url": "https://arxiv.org/abs/2410.03742", "pdf_url": "https://arxiv.org/pdf/2410.03742v2", "arxiv_id": "2410.03742", "doi": "10.48550/arXiv.2410.03742", "citation_count": 35, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3891}
{"id": "f4b18fddf6f779532635c86788e728500b0402ac3f8d108aecd9eff9e4e379b2", "sources": ["arxiv", "semantic_scholar"], "title": "Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference", "abstract": "Reward inference (learning a reward model from human preferences) is a critical intermediate step in the Reinforcement Learning from Human Feedback (RLHF) pipeline for fine-tuning Large Language Models (LLMs). In practice, RLHF faces fundamental challenges such as distribution shift, reward model overfitting, and problem misspecification. An alternative approach is direct policy optimization without reward inference, such as Direct Preference Optimization (DPO), which provides a much simpler pipeline and has shown empirical success in LLM applications. However, DPO utilizes the closed-form expression between the optimal policy and the reward function, which is only suitable under the bandit setting or deterministic MDPs. This paper develops two RLHF algorithms without reward inference for general RL problems beyond bandits and deterministic MDPs, and general preference models beyond the Bradley-Terry model. The key idea is to estimate the local value function difference from human preferences and then approximate the policy gradient with a zeroth-order gradient approximator. For both algorithms, we establish polynomial convergence rates in terms of the number of policy gradient iterations, the number of trajectory samples, and human preference queries per iteration. Numerical experiments in stochastic environments validate the performance of our proposed algorithms, outperforming popular RLHF baselines such as DPO and PPO. Our paper shows there exist provably efficient methods to solve general RLHF problems without reward inference.", "authors": ["Qining Zhang", "Lei Ying"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-09-25", "url": "https://arxiv.org/abs/2409.17401", "pdf_url": "https://arxiv.org/pdf/2409.17401v2", "arxiv_id": "2409.17401", "doi": "10.48550/arXiv.2409.17401", "citation_count": 13, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2865}
{"id": "246f4d103ddc49779c85bcb051a89dac31ec5eca5a4145a14894310682d40e6f", "sources": ["arxiv", "semantic_scholar"], "title": "Aligning Language Models Using Follow-up Likelihood as Reward Signal", "abstract": "In natural human-to-human conversations, participants often receive feedback signals from one another based on their follow-up reactions. These reactions can include verbal responses, facial expressions, changes in emotional state, and other non-verbal cues. Similarly, in human-machine interactions, the machine can leverage the user's follow-up utterances as feedback signals to assess whether it has appropriately addressed the user's request. Therefore, we propose using the likelihood of follow-up utterances as rewards to differentiate preferred responses from less favored ones, without relying on human or commercial LLM-based preference annotations. Our proposed reward mechanism, ``Follow-up Likelihood as Reward\" (FLR), matches the performance of strong reward models trained on large-scale human or GPT-4 annotated data on 8 pairwise-preference and 4 rating-based benchmarks. Building upon the FLR mechanism, we propose to automatically mine preference data from the online generations of a base policy model. The preference data are subsequently used to boost the helpfulness of the base model through direct alignment from preference (DAP) methods, such as direct preference optimization (DPO). Lastly, we demonstrate that fine-tuning the language model that provides follow-up likelihood with natural language feedback significantly enhances FLR's performance on reward modeling benchmarks and effectiveness in aligning the base policy model's helpfulness.", "authors": ["Chen Zhang", "Dading Chong", "Feng Jiang", "Chengguang Tang", "Anningzhe Gao", "Guohua Tang", "Haizhou Li"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-20", "url": "https://arxiv.org/abs/2409.13948", "pdf_url": "https://arxiv.org/pdf/2409.13948v3", "arxiv_id": "2409.13948", "doi": "10.48550/arXiv.2409.13948", "citation_count": 7, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/e0397123/FLR", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2258}
{"id": "d71c4eb6a201058e28ccd44bb13823458ce6591f783500e1d6cf6db66965ca7a", "sources": ["arxiv", "semantic_scholar"], "title": "OMG-RL:Offline Model-based Guided Reward Learning for Heparin Treatment", "abstract": "Accurate medication dosing holds an important position in the overall patient therapeutic process. Therefore, much research has been conducted to develop optimal administration strategy based on Reinforcement learning (RL). However, Relying solely on a few explicitly defined reward functions makes it difficult to learn a treatment strategy that encompasses the diverse characteristics of various patients. Moreover, the multitude of drugs utilized in clinical practice makes it infeasible to construct a dedicated reward function for each medication. Here, we tried to develop a reward network that captures clinicians' therapeutic intentions, departing from explicit rewards, and to derive an optimal heparin dosing policy. In this study, we introduce Offline Model-based Guided Reward Learning (OMG-RL), which performs offline inverse RL (IRL). Through OMG-RL, we learn a parameterized reward function that captures the expert's intentions from limited data, thereby enhancing the agent's policy. We validate the proposed approach on the heparin dosing task. We show that OMG-RL policy is positively reinforced not only in terms of the learned reward network but also in activated partial thromboplastin time (aPTT), a key indicator for monitoring the effects of heparin. This means that the OMG-RL policy adequately reflects clinician's intentions. This approach can be widely utilized not only for the heparin dosing problem but also for RL-based medication dosing tasks in general.", "authors": ["Yooseok Lim", "Sujee Lee"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-20", "url": "https://arxiv.org/abs/2409.13299", "pdf_url": "https://arxiv.org/pdf/2409.13299v2", "arxiv_id": "2409.13299", "doi": "10.1016/j.bea.2025.100198", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Biomedical Engineering Advances", "quality_score": 0.0753}
{"id": "95239746dd142834d64645a9c1bd2f430e0057aa9db35bd9b6853c6d04dae177", "sources": ["arxiv", "semantic_scholar"], "title": "Reward-Robust RLHF in LLMs", "abstract": "As Large Language Models (LLMs) continue to progress toward more advanced forms of intelligence, Reinforcement Learning from Human Feedback (RLHF) is increasingly seen as a key pathway toward achieving Artificial General Intelligence (AGI). However, the reliance on reward-model-based (RM-based) alignment methods introduces significant challenges due to the inherent instability and imperfections of Reward Models (RMs), which can lead to critical issues such as reward hacking and misalignment with human intentions. In this paper, we introduce a reward-robust RLHF framework aimed at addressing these fundamental challenges, paving the way for more reliable and resilient learning in LLMs. Our approach introduces a novel optimization objective that carefully balances performance and robustness by incorporating Bayesian Reward Model Ensembles (BRME) to model the uncertainty set of reward functions. This allows the framework to integrate both nominal performance and minimum reward signals, ensuring more stable learning even with imperfect RMs. Empirical results demonstrate that our framework consistently outperforms baselines across diverse benchmarks, showing improved accuracy and long-term stability. We also provide a theoretical analysis, demonstrating that reward-robust RLHF approaches the stability of constant reward settings, which proves to be acceptable even in a stochastic-case analysis. Together, these contributions highlight the framework potential to enhance both the performance and stability of LLM alignment.", "authors": ["Yuzi Yan", "Xingzhou Lou", "Jialian Li", "Yiping Zhang", "Jian Xie", "Chao Yu", "Yu Wang", "Dong Yan", "Yuan Shen"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-18", "url": "https://arxiv.org/abs/2409.15360", "pdf_url": "https://arxiv.org/pdf/2409.15360v3", "arxiv_id": "2409.15360", "doi": "10.48550/arXiv.2409.15360", "citation_count": 30, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3728}
{"id": "31ab2053a76371be3f6c2f24732b54264d281e6066bf4d7e7d36ce85a1d16d95", "sources": ["arxiv", "semantic_scholar"], "title": "Quantile Regression for Distributional Reward Models in RLHF", "abstract": "Reinforcement learning from human feedback (RLHF) has become a key method for aligning large language models (LLMs) with human preferences through the use of reward models. However, traditional reward models typically generate point estimates, which oversimplify the diversity and complexity of human values and preferences. In this paper, we introduce Quantile Reward Models (QRMs), a novel approach to reward modeling that learns a distribution over rewards instead of a single scalar value. Our method uses quantile regression to estimate a full, potentially multimodal distribution over preferences, providing a more powerful and nuanced representation of preferences. This distributional approach can better capture the diversity of human values, addresses label noise, and accommodates conflicting preferences by modeling them as distinct modes in the distribution. Our experimental results show that QRM outperforms comparable traditional point-estimate models on RewardBench. Furthermore, we demonstrate that the additional information provided by the distributional estimates can be utilized in downstream applications, such as risk-aware reinforcement learning, resulting in LLM policies that generate fewer extremely negative responses. Our code and model are released at https://github.com/Nicolinho/QRM.", "authors": ["Nicolai Dorka"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-16", "url": "https://arxiv.org/abs/2409.10164", "pdf_url": "https://arxiv.org/pdf/2409.10164v1", "arxiv_id": "2409.10164", "doi": "10.48550/arXiv.2409.10164", "citation_count": 60, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/Nicolinho/QRM", "venue": "arXiv.org", "quality_score": 0.4463}
{"id": "323bf7937b345c8c9ac3793670e227e764f062866dd5eb37790dda2257cdc8f1", "sources": ["arxiv", "semantic_scholar"], "title": "Provably Efficient Infinite-Horizon Average-Reward Reinforcement Learning with Linear Function Approximation", "abstract": "This paper proposes a computationally tractable algorithm for learning infinite-horizon average-reward linear Markov decision processes (MDPs) and linear mixture MDPs under the Bellman optimality condition. While guaranteeing computational efficiency, our algorithm for linear MDPs achieves the best-known regret upper bound of $\\widetilde{\\mathcal{O}}(d^{3/2}\\mathrm{sp}(v^*)\\sqrt{T})$ over $T$ time steps where $\\mathrm{sp}(v^*)$ is the span of the optimal bias function $v^*$ and $d$ is the dimension of the feature mapping. For linear mixture MDPs, our algorithm attains a regret bound of $\\widetilde{\\mathcal{O}}(d\\cdot\\mathrm{sp}(v^*)\\sqrt{T})$. The algorithm applies novel techniques to control the covering number of the value function class and the span of optimistic estimators of the value function, which is of independent interest.", "authors": ["Woojin Chae", "Dabeen Lee"], "categories": ["cs.LG", "cs.DS", "math.OC"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-09-16", "url": "https://arxiv.org/abs/2409.10772", "pdf_url": "https://arxiv.org/pdf/2409.10772v2", "arxiv_id": "2409.10772", "doi": "10.48550/arXiv.2409.10772", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "6ee923ec4ab5df96d80873a36b4d44e8a4660ba5729562cd3056d79c7df86c34", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Causally Invariant Reward Functions from Diverse Demonstrations", "abstract": "Inverse reinforcement learning methods aim to retrieve the reward function of a Markov decision process based on a dataset of expert demonstrations. The commonplace scarcity and heterogeneous sources of such demonstrations can lead to the absorption of spurious correlations in the data by the learned reward function. Consequently, this adaptation often exhibits behavioural overfitting to the expert data set when a policy is trained on the obtained reward function under distribution shift of the environment dynamics. In this work, we explore a novel regularization approach for inverse reinforcement learning methods based on the causal invariance principle with the goal of improved reward function generalization. By applying this regularization to both exact and approximate formulations of the learning task, we demonstrate superior policy performance when trained using the recovered reward functions in a transfer setting", "authors": ["Ivan Ovinnikov", "Eugene Bykovets", "Joachim M. Buhmann"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-12", "url": "https://arxiv.org/abs/2409.08012", "pdf_url": "https://arxiv.org/pdf/2409.08012v1", "arxiv_id": "2409.08012", "doi": "10.48550/arXiv.2409.08012", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "4a8c9fa032154cc22a408afb2dffea86152c9a7f49a070ad08f162c65a6d5332", "sources": ["arxiv", "semantic_scholar"], "title": "Policy Filtration for RLHF to Mitigate Noise in Reward Models", "abstract": "While direct policy optimization methods exist, pioneering LLMs are fine-tuned with reinforcement learning from human feedback (RLHF) to generate better responses under the supervision of a reward model learned from preference data. One major challenge of RLHF is the inaccuracy of the intermediate reward model, especially in the tasks that requires complex reasoning for the reward model to score a response. We find that the reliability of the reward model varies across responses assigned with different rewards. This motivates us to filter the samples whose rewards may be unreliable to improve the signal-to-noise ratio during policy learning, resulting in Policy Filtration for Proximal Policy Optimization (PF-PPO). To choose a proper policy filtering strategy, we use the coefficient of determination (R2) between the rewards and actual scores on filtered samples as the metrics to help us find promising strategies since it measures how well the rewards filtered by PF-PPO indicate real performance. We provide extensive experiments to validate the effectiveness of PF-PPO in code generation and math reasoning tasks. In code generation, PF-PPO achieves the state-of-the-art performance of 7-billion-parameter models on HumanEval (+7.9%), MBPP (+0.7%), and LeetCode Contest (+10.0%) which is a more challenging benchmark created by us. In math reasoning, PF-PPO yields performance increase using different reward models and benchmarks (Ape210K and CMATH). Code is available on https://github.com/DtYXs/verl/tree/pf-ppo.", "authors": ["Chuheng Zhang", "Wei Shen", "Li Zhao", "Xuyun Zhang", "Xiaolong Xu", "Wanchun Dou", "Jiang Bian"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-11", "url": "https://arxiv.org/abs/2409.06957", "pdf_url": "https://arxiv.org/pdf/2409.06957v5", "arxiv_id": "2409.06957", "doi": null, "citation_count": 17, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/DtYXs/verl/tree/pf-ppo", "venue": "International Conference on Machine Learning", "quality_score": 0.3138}
{"id": "041bfa6e5d8ab9f9c2931c26453d787b3f8f77f6b7f3b91d21aa2a37f5a86f73", "sources": ["arxiv", "semantic_scholar"], "title": "Semi-Supervised Reward Modeling via Iterative Self-Training", "abstract": "Reward models (RM) capture the values and preferences of humans and play a central role in Reinforcement Learning with Human Feedback (RLHF) to align pretrained large language models (LLMs). Traditionally, training these models relies on extensive human-annotated preference data, which poses significant challenges in terms of scalability and cost. To overcome these limitations, we propose Semi-Supervised Reward Modeling (SSRM), an approach that enhances RM training using unlabeled data. Given an unlabeled dataset, SSRM involves three key iterative steps: pseudo-labeling unlabeled examples, selecting high-confidence examples through a confidence threshold, and supervised finetuning on the refined dataset. Across extensive experiments on various model configurations, we demonstrate that SSRM significantly improves reward models without incurring additional labeling costs. Notably, SSRM can achieve performance comparable to models trained entirely on labeled data of equivalent volumes. Overall, SSRM substantially reduces the dependency on large volumes of human-annotated data, thereby decreasing the overall cost and time involved in training effective reward models.", "authors": ["Yifei He", "Haoxiang Wang", "Ziyan Jiang", "Alexandros Papangelis", "Han Zhao"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-10", "url": "https://arxiv.org/abs/2409.06903", "pdf_url": "https://arxiv.org/pdf/2409.06903v1", "arxiv_id": "2409.06903", "doi": "10.48550/arXiv.2409.06903", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2603}
{"id": "7680263db76435f58211d60342981a6731d606edd41202695ec4a1d27c0c1483", "sources": ["arxiv", "semantic_scholar"], "title": "On the Limited Generalization Capability of the Implicit Reward Model Induced by Direct Preference Optimization", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is an effective approach for aligning language models to human preferences. Central to RLHF is learning a reward function for scoring human preferences. Two main approaches for learning a reward model are 1) training an EXplicit Reward Model (EXRM) as in RLHF, and 2) using an implicit reward learned from preference data through methods such as Direct Preference Optimization (DPO). Prior work has shown that the implicit reward model of DPO (denoted as DPORM) can approximate an EXRM in the limit. DPORM's effectiveness directly implies the optimality of the learned policy, and also has practical implication for LLM alignment methods including iterative DPO. However, it is unclear how well DPORM empirically matches the performance of EXRM. This work studies the accuracy at distinguishing preferred and rejected answers for both DPORM and EXRM. Our findings indicate that even though DPORM fits the training dataset comparably, it generalizes less effectively than EXRM, especially when the validation datasets contain distribution shifts. Across five out-of-distribution settings, DPORM has a mean drop in accuracy of 3% and a maximum drop of 7%. These findings highlight that DPORM has limited generalization ability and substantiates the integration of an explicit reward model in iterative DPO approaches.", "authors": ["Yong Lin", "Skyler Seto", "Maartje ter Hoeve", "Katherine Metcalf", "Barry-John Theobald", "Xuan Wang", "Yizhe Zhang", "Chen Huang", "Tong Zhang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-05", "url": "https://arxiv.org/abs/2409.03650", "pdf_url": "https://arxiv.org/pdf/2409.03650v2", "arxiv_id": "2409.03650", "doi": "10.48550/arXiv.2409.03650", "citation_count": 28, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.3656}
{"id": "1915083cfac9324fe919dc0d2936cdd5fd8c6da5a8dc510063c3d7b01b47fc06", "sources": ["arxiv", "semantic_scholar"], "title": "ELO-Rated Sequence Rewards: Advancing Reinforcement Learning Models", "abstract": "Reinforcement Learning (RL) heavily relies on the careful design of the reward function. However, accurately assigning rewards to each state-action pair in Long-Term Reinforcement Learning (LTRL) tasks remains a significant challenge. As a result, RL agents are often trained under expert guidance. Inspired by the ordinal utility theory in economics, we propose a novel reward estimation algorithm: ELO-Rating based Reinforcement Learning (ERRL). This approach features two key contributions. First, it uses expert preferences over trajectories rather than cardinal rewards (utilities) to compute the ELO rating of each trajectory as its reward. Second, a new reward redistribution algorithm is introduced to alleviate training instability in the absence of a fixed anchor reward. In long-term scenarios (up to 5000 steps), where traditional RL algorithms struggle, our method outperforms several state-of-the-art baselines. Additionally, we conduct a comprehensive analysis of how expert preferences influence the results.", "authors": ["Qi Ju", "Falin Hei", "Zhemei Fang", "Yunfeng Luo"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-09-05", "url": "https://arxiv.org/abs/2409.03301", "pdf_url": "https://arxiv.org/pdf/2409.03301v2", "arxiv_id": "2409.03301", "doi": "10.1109/DDCLS61622.2024.10606606", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "929593d5ae8ebeacf03e512cf7888d4a5b7443658b63634a6f620ff7ea42c044", "sources": ["arxiv", "semantic_scholar"], "title": "Robust Q-Learning under Corrupted Rewards", "abstract": "Recently, there has been a surge of interest in analyzing the non-asymptotic behavior of model-free reinforcement learning algorithms. However, the performance of such algorithms in non-ideal environments, such as in the presence of corrupted rewards, is poorly understood. Motivated by this gap, we investigate the robustness of the celebrated Q-learning algorithm to a strong-contamination attack model, where an adversary can arbitrarily perturb a small fraction of the observed rewards. We start by proving that such an attack can cause the vanilla Q-learning algorithm to incur arbitrarily large errors. We then develop a novel robust synchronous Q-learning algorithm that uses historical reward data to construct robust empirical Bellman operators at each time step. Finally, we prove a finite-time convergence rate for our algorithm that matches known state-of-the-art bounds (in the absence of attacks) up to a small inevitable $O(\\varepsilon)$ error term that scales with the adversarial corruption fraction $\\varepsilon$. Notably, our results continue to hold even when the true reward distributions have infinite support, provided they admit bounded second moments.", "authors": ["Sreejeet Maity", "Aritra Mitra"], "categories": ["cs.LG", "eess.SY", "math.OC", "stat.ML"], "fields_of_study": ["Computer Science", "Engineering", "Mathematics"], "published_date": "2024-09-05", "url": "https://arxiv.org/abs/2409.03237", "pdf_url": "https://arxiv.org/pdf/2409.03237v1", "arxiv_id": "2409.03237", "doi": "10.1109/CDC56724.2024.10885945", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Conference on Decision and Control", "quality_score": 0.0753}
{"id": "eb3c03601592d761817c1b6bc59a866fea6eda0e049ba948b29deb51ec93980d", "sources": ["arxiv", "semantic_scholar"], "title": "Building Math Agents with Multi-Turn Iterative Preference Learning", "abstract": "Recent studies have shown that large language models' (LLMs) mathematical problem-solving capabilities can be enhanced by integrating external tools, such as code interpreters, and employing multi-turn Chain-of-Thought (CoT) reasoning. While current methods focus on synthetic data generation and Supervised Fine-Tuning (SFT), this paper studies the complementary direct preference learning approach to further improve model performance. However, existing direct preference learning algorithms are originally designed for the single-turn chat task, and do not fully address the complexities of multi-turn reasoning and external tool integration required for tool-integrated mathematical reasoning tasks. To fill in this gap, we introduce a multi-turn direct preference learning framework, tailored for this context, that leverages feedback from code interpreters and optimizes trajectory-level preferences. This framework includes multi-turn DPO and multi-turn KTO as specific implementations. The effectiveness of our framework is validated through training of various language models using an augmented prompt set from the GSM8K and MATH datasets. Our results demonstrate substantial improvements: a supervised fine-tuned Gemma-1.1-it-7B model's performance increased from 77.5% to 83.9% on GSM8K and from 46.1% to 51.2% on MATH. Similarly, a Gemma-2-it-9B model improved from 84.1% to 86.3% on GSM8K and from 51.0% to 54.5% on MATH.", "authors": ["Wei Xiong", "Chengshuai Shi", "Jiaming Shen", "Aviv Rosenberg", "Zhen Qin", "Daniele Calandriello", "Misha Khalman", "Rishabh Joshi", "Bilal Piot", "Mohammad Saleh", "Chi Jin", "Tong Zhang", "Tianqi Liu"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-09-04", "url": "https://arxiv.org/abs/2409.02392", "pdf_url": "https://arxiv.org/pdf/2409.02392v2", "arxiv_id": "2409.02392", "doi": "10.48550/arXiv.2409.02392", "citation_count": 65, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4549}
{"id": "1df60deee3ae3db78d1364be992cf50763a7a0ff37d37d391251a3a45990ee4d", "sources": ["arxiv", "semantic_scholar"], "title": "Sequence to Sequence Reward Modeling: Improving RLHF by Language Feedback", "abstract": "Aligning the behavior of Large language models (LLMs) with human intentions and values remains a critical challenge. Reinforcement learning from human feedback (RLHF) aligns LLMs by training a reward model (RM) on human preferences and fine-tuning the LLMs to maximize RM feedback. Despite its effectiveness and popularity, RLHF is prone to biased local optimization. It means RM fails to provide feedback that accurately aligns with human preference, causing LLMs to explore unexpected generalizations, and failing to achieve alignment objectives. To mitigate this issue, we propose a novel \\textit{sequence-to-sequence (seq2seq) reward modeling} method. Its key insight is that learning from language feedback rather than scalar feedback improves RLHF without additional annotations. We replaced the reward modeling target from binary maximum likelihood estimation (MLE) with sequence MLE. This method enables richer and fine-grained language feedback without additional annotations, models, or training stages. Our experiments demonstrated its effectiveness, specifically, reducing the refusal-to-response paradigm in single-turn safety dialogues and the long-response bias in text summarization tasks. We provide further analysis that seq2seq RM improves RLHF performance across 2B and 7B LLMs on 3 NLP tasks, achieving an average win rate of 76.9\\%. We further show that seq2seq RM can still improve the performance of RLHF under out-of-distribution prompts.", "authors": ["Jiayi Zhou", "Jiaming Ji", "Juntao Dai", "Dong Li", "Yaodong Yang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-30", "url": "https://arxiv.org/abs/2409.00162", "pdf_url": "https://arxiv.org/pdf/2409.00162v2", "arxiv_id": "2409.00162", "doi": "10.48550/arXiv.2409.00162", "citation_count": 20, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3306}
{"id": "6074ed21faefea569948baa2fab5fe0f322fcebb08ee00ae2f40bf3c904f4b4d", "sources": ["arxiv", "semantic_scholar"], "title": "Critique-out-Loud Reward Models", "abstract": "Traditionally, reward models used for reinforcement learning from human feedback (RLHF) are trained to directly predict preference scores without leveraging the generation capabilities of the underlying large language model (LLM). This limits the capabilities of reward models as they must reason implicitly about the quality of a response, i.e., preference modeling must be performed in a single forward pass through the model. To enable reward models to reason explicitly about the quality of a response, we introduce Critique-out-Loud (CLoud) reward models. CLoud reward models operate by first generating a natural language critique of the assistant's response that is then used to predict a scalar reward for the quality of the response. We demonstrate the success of CLoud reward models for both Llama-3-8B and 70B base models: compared to classic reward models CLoud reward models improve pairwise preference classification accuracy on RewardBench by 4.65 and 5.84 percentage points for the 8B and 70B base models respectively. Furthermore, CLoud reward models lead to a Pareto improvement for win rate on ArenaHard when used as the scoring model for Best-of-N. Finally, we explore how to exploit the dynamic inference compute capabilities of CLoud reward models by performing self-consistency decoding for reward prediction.", "authors": ["Zachary Ankner", "Mansheej Paul", "Brandon Cui", "Jonathan D. Chang", "Prithviraj Ammanabrolu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-21", "url": "https://arxiv.org/abs/2408.11791", "pdf_url": "https://arxiv.org/pdf/2408.11791v1", "arxiv_id": "2408.11791", "doi": null, "citation_count": 89, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4886}
{"id": "66df04436868d67d520458bc56da553c1bb8da5d2e32edcefa01a57ee95478fa", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Difference Optimization For Sample Reweighting In Offline RLHF", "abstract": "With the rapid advances in Large Language Models (LLMs), aligning LLMs with human preferences become increasingly important. Although Reinforcement Learning with Human Feedback (RLHF) proves effective, it is complicated and highly resource-intensive. As such, offline RLHF has been introduced as an alternative solution, which directly optimizes LLMs with ranking losses on a fixed preference dataset. Current offline RLHF only captures the \"ordinal relationship\" between responses, overlooking the crucial aspect of how much one is preferred over the others. To address this issue, we propose a simple yet effective solution called Reward Difference Optimization, shorted as RDO. Specifically, we introduce reward difference coefficients to reweigh sample pairs in offline RLHF. We then develop a difference model which captures rich interactions between a pair of responses for predicting these difference coefficients. Experiments with 7B LLMs on the HH and TL;DR datasets substantiate the effectiveness of our method in both automatic metrics and human evaluation, thereby highlighting its potential for aligning LLMs with human intent and values", "authors": ["Shiqi Wang", "Zhengze Zhang", "Rui Zhao", "Fei Tan", "Cam Tu Nguyen"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-18", "url": "https://arxiv.org/abs/2408.09385", "pdf_url": "https://arxiv.org/pdf/2408.09385v2", "arxiv_id": "2408.09385", "doi": "10.18653/v1/2024.findings-emnlp.115", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2603}
{"id": "e7521bb546072b3ee4876620727d80124649409f4be057259d0e6da1e8adf79b", "sources": ["arxiv", "semantic_scholar"], "title": "Listwise Reward Estimation for Offline Preference-based Reinforcement Learning", "abstract": "In Reinforcement Learning (RL), designing precise reward functions remains to be a challenge, particularly when aligning with human intent. Preference-based RL (PbRL) was introduced to address this problem by learning reward models from human feedback. However, existing PbRL methods have limitations as they often overlook the second-order preference that indicates the relative strength of preference. In this paper, we propose Listwise Reward Estimation (LiRE), a novel approach for offline PbRL that leverages second-order preference information by constructing a Ranked List of Trajectories (RLT), which can be efficiently built by using the same ternary feedback type as traditional methods. To validate the effectiveness of LiRE, we propose a new offline PbRL dataset that objectively reflects the effect of the estimated rewards. Our extensive experiments on the dataset demonstrate the superiority of LiRE, i.e., outperforming state-of-the-art baselines even with modest feedback budgets and enjoying robustness with respect to the number of feedbacks and feedback noise. Our code is available at https://github.com/chwoong/LiRE", "authors": ["Heewoong Choi", "Sangwon Jung", "Hongjoon Ahn", "Taesup Moon"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-08", "url": "https://arxiv.org/abs/2408.04190", "pdf_url": "https://arxiv.org/pdf/2408.04190v1", "arxiv_id": "2408.04190", "doi": "10.48550/arXiv.2408.04190", "citation_count": 18, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/chwoong/LiRE", "venue": "International Conference on Machine Learning", "quality_score": 0.3891}
{"id": "db58c02f61ff3019d1334cebeb6e373fb1d20660d46e02057ac1157d1f06c9ad", "sources": ["arxiv", "semantic_scholar"], "title": "RVI-SAC: Average Reward Off-Policy Deep Reinforcement Learning", "abstract": "In this paper, we propose an off-policy deep reinforcement learning (DRL) method utilizing the average reward criterion. While most existing DRL methods employ the discounted reward criterion, this can potentially lead to a discrepancy between the training objective and performance metrics in continuing tasks, making the average reward criterion a recommended alternative. We introduce RVI-SAC, an extension of the state-of-the-art off-policy DRL method, Soft Actor-Critic (SAC), to the average reward criterion. Our proposal consists of (1) Critic updates based on RVI Q-learning, (2) Actor updates introduced by the average reward soft policy improvement theorem, and (3) automatic adjustment of Reset Cost enabling the average reward reinforcement learning to be applied to tasks with termination. We apply our method to the Gymnasium's Mujoco tasks, a subset of locomotion tasks, and demonstrate that RVI-SAC shows competitive performance compared to existing methods.", "authors": ["Yukinari Hisaki", "Isao Ono"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-08-04", "url": "https://arxiv.org/abs/2408.01972", "pdf_url": "https://arxiv.org/pdf/2408.01972v1", "arxiv_id": "2408.01972", "doi": "10.48550/arXiv.2408.01972", "citation_count": 6, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/yhisaki/average-reward-drl", "venue": "International Conference on Machine Learning", "quality_score": 0.2113}
{"id": "9852bac370f5c28cb8e03b2686ab7c7207b88f06154d0a368be0fad044ddc8b5", "sources": ["arxiv", "semantic_scholar"], "title": "Exploring and Addressing Reward Confusion in Offline Preference Learning", "abstract": "Spurious correlations in a reward model's training data can prevent Reinforcement Learning from Human Feedback (RLHF) from identifying the desired goal and induce unwanted behaviors. This paper shows that offline RLHF is susceptible to reward confusion, especially in the presence of spurious correlations in offline data. We create a benchmark to study this problem and propose a method that can significantly reduce reward confusion by leveraging transitivity of preferences while building a global preference chain with active learning.", "authors": ["Xin Chen", "Sam Toyer", "Florian Shkurti"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-22", "url": "https://arxiv.org/abs/2407.16025", "pdf_url": "https://arxiv.org/pdf/2407.16025v2", "arxiv_id": "2407.16025", "doi": "10.48550/arXiv.2407.16025", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "105b357a72ad1219114368e8201e54e37573afdaf3ae305d485120196317c160", "sources": ["arxiv", "semantic_scholar"], "title": "Boosting Reward Model with Preference-Conditional Multi-Aspect Synthetic Data Generation", "abstract": "Reward models (RMs) are crucial for aligning large language models (LLMs) with human preferences. They are trained using preference datasets where each example consists of one input prompt, two responses, and a preference label. As curating a high-quality human labeled preference dataset is both time-consuming and expensive, people often rely on existing powerful LLMs for preference label generation. This can potentially introduce noise and impede RM training. In this work, we present RMBoost, a novel synthetic preference data generation paradigm to boost reward model quality. Unlike traditional methods, which generate two responses before obtaining the preference label, RMBoost first generates one response and selects a preference label, followed by generating the second more (or less) preferred response conditioned on the pre-selected preference label and the first response. This approach offers two main advantages. First, RMBoost reduces labeling noise since preference pairs are constructed intentionally. Second, RMBoost facilitates the creation of more diverse responses by incorporating various quality aspects (e.g., helpfulness, relevance, completeness) into the prompts. We conduct extensive experiments across three diverse datasets and demonstrate that RMBoost outperforms other synthetic preference data generation techniques and significantly boosts the performance of four distinct reward models.", "authors": ["Jiaming Shen", "Ran Xu", "Yennie Jun", "Zhen Qin", "Tianqi Liu", "Carl Yang", "Yi Liang", "Simon Baumgartner", "Michael Bendersky"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-22", "url": "https://arxiv.org/abs/2407.16008", "pdf_url": "https://arxiv.org/pdf/2407.16008v2", "arxiv_id": "2407.16008", "doi": "10.48550/arXiv.2407.16008", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "60f3d81e4701f3995ea2b1b8ba7ccf17e84a55647852a5ddd7fd0e10d7559c0d", "sources": ["arxiv", "semantic_scholar"], "title": "Comprehensive Overview of Reward Engineering and Shaping in Advancing Reinforcement Learning Applications", "abstract": "The aim of Reinforcement Learning (RL) in real-world applications is to create systems capable of making autonomous decisions by learning from their environment through trial and error. This paper emphasizes the importance of reward engineering and reward shaping in enhancing the efficiency and effectiveness of reinforcement learning algorithms. Reward engineering involves designing reward functions that accurately reflect the desired outcomes, while reward shaping provides additional feedback to guide the learning process, accelerating convergence to optimal policies. Despite significant advancements in reinforcement learning, several limitations persist. One key challenge is the sparse and delayed nature of rewards in many real-world scenarios, which can hinder learning progress. Additionally, the complexity of accurately modeling real-world environments and the computational demands of reinforcement learning algorithms remain substantial obstacles. On the other hand, recent advancements in deep learning and neural networks have significantly improved the capability of reinforcement learning systems to handle high-dimensional state and action spaces, enabling their application to complex tasks such as robotics, autonomous driving, and game playing. This paper provides a comprehensive review of the current state of reinforcement learning, focusing on the methodologies and techniques used in reward engineering and reward shaping. It critically analyzes the limitations and recent advancements in the field, offering insights into future research directions and potential applications in various domains.", "authors": ["Sinan Ibrahim", "Mostafa Mostafa", "Ali Jnadi", "Hadi Salloum", "Pavel Osinenko"], "categories": ["cs.LG", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2024-07-22", "url": "https://arxiv.org/abs/2408.10215", "pdf_url": "https://arxiv.org/pdf/2408.10215v2", "arxiv_id": "2408.10215", "doi": "10.1109/ACCESS.2024.3504735", "citation_count": 97, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Access", "quality_score": 0.4978}
{"id": "86e6d7d889df62948fe5f0b2f7a803a79d97909590cb8d811ea33fc0c4c19476", "sources": ["arxiv", "semantic_scholar"], "title": "Catastrophic Goodhart: regularizing RLHF with KL divergence does not mitigate heavy-tailed reward misspecification", "abstract": "When applying reinforcement learning from human feedback (RLHF), the reward is learned from data and, therefore, always has some error. It is common to mitigate this by regularizing the policy with KL divergence from a base model, with the hope that balancing reward with regularization will achieve desirable outcomes despite this reward misspecification. We show that when the reward function has light-tailed error, optimal policies under less restrictive KL penalties achieve arbitrarily high utility. However, if error is heavy-tailed, some policies obtain arbitrarily high reward despite achieving no more utility than the base model--a phenomenon we call catastrophic Goodhart. We adapt a discrete optimization method to measure the tails of reward models, finding that they are consistent with light-tailed error. However, the pervasiveness of heavy-tailed distributions in many real-world applications indicates that future sources of RL reward could have heavy-tailed error, increasing the likelihood of reward hacking even with KL regularization.", "authors": ["Thomas Kwa", "Drake Thomas", "Adrià Garriga-Alonso"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-19", "url": "https://arxiv.org/abs/2407.14503", "pdf_url": "https://arxiv.org/pdf/2407.14503v2", "arxiv_id": "2407.14503", "doi": "10.48550/arXiv.2407.14503", "citation_count": 13, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.2865}
{"id": "0b033f08b2ef131b424f996844d4cd9c8def9d1cab471624055f1e127b048769", "sources": ["arxiv", "semantic_scholar"], "title": "Data-Centric Human Preference with Rationales for Direct Preference Alignment", "abstract": "Aligning language models with human preferences through reinforcement learning from human feedback is crucial for their safe and effective deployment. The human preference is typically represented through comparison where one response is chosen over another for a given prompt. However, standard preference datasets often lack explicit information on why a particular choice was made, presenting an ambiguity that can hinder efficient learning and robust alignment, especially given the high cost of acquiring extensive human annotations. While many studies focus on algorithmic improvements, this work adopts a data-centric perspective, exploring how to enhance learning from existing preference data. We propose augmenting standard preference pairs with rationales that explain the reasoning behind the human preference. Specifically, we introduce a simple and principled framework that leverages machine-generated rationales to enrich preference data for preference optimization algorithms. Our comprehensive analysis demonstrates that incorporating rationales improves learning efficiency. Extensive experiments reveal some advantages: rationale-augmented learning accelerates convergence and can achieve higher final model performance. Furthermore, this approach is versatile and compatible with various direct preference optimization algorithms. Our findings showcase the potential of thoughtful data design in preference learning, demonstrating that enriching existing datasets with explanatory rationales can help unlock improvements in model alignment and annotation efficiency.", "authors": ["Hoang Anh Just", "Ming Jin", "Anit Sahu", "Huy Phan", "Ruoxi Jia"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-19", "url": "https://arxiv.org/abs/2407.14477", "pdf_url": "https://arxiv.org/pdf/2407.14477v4", "arxiv_id": "2407.14477", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "c85a529df48be1782a9138e8f979d8847034656bf8479c50bf09c37c5ea4ac30", "sources": ["arxiv", "semantic_scholar"], "title": "Offline Reinforcement Learning with Imputed Rewards", "abstract": "Offline Reinforcement Learning (ORL) offers a robust solution to training agents in applications where interactions with the environment must be strictly limited due to cost, safety, or lack of accurate simulation environments. Despite its potential to facilitate deployment of artificial agents in the real world, Offline Reinforcement Learning typically requires very many demonstrations annotated with ground-truth rewards. Consequently, state-of-the-art ORL algorithms can be difficult or impossible to apply in data-scarce scenarios. In this paper we propose a simple but effective Reward Model that can estimate the reward signal from a very limited sample of environment transitions annotated with rewards. Once the reward signal is modeled, we use the Reward Model to impute rewards for a large sample of reward-free transitions, thus enabling the application of ORL techniques. We demonstrate the potential of our approach on several D4RL continuous locomotion tasks. Our results show that, using only 1\\% of reward-labeled transitions from the original datasets, our learned reward model is able to impute rewards for the remaining 99\\% of the transitions, from which performant agents can be learned using Offline Reinforcement Learning.", "authors": ["Carlo Romeo", "Andrew D. Bagdanov"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-15", "url": "https://arxiv.org/abs/2407.10839", "pdf_url": "https://arxiv.org/pdf/2407.10839v1", "arxiv_id": "2407.10839", "doi": "10.48550/arXiv.2407.10839", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "9320769eac425eaddb8dfa8d7b4a3e2a901f5726efde20de35b23bad4d90158f", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Federated RLHF with Aggregated Client Preference for LLMs", "abstract": "Reinforcement learning with human feedback (RLHF) fine-tunes a pretrained large language model (LLM) using user preference data, enabling it to generate content aligned with human preferences. However, due to privacy concerns, users may be reluctant to share sensitive preference data. To address this, we propose utilizing Federated Learning (FL) techniques, allowing large-scale preference collection from diverse real-world users without requiring them to transmit data to a central server. Our federated RLHF methods (i.e., FedBis and FedBiscuit) encode each client's preferences into binary selectors and aggregate them to capture common preferences. In particular, FedBiscuit overcomes key challenges, such as preference heterogeneity and reward hacking, through innovative solutions like grouping clients with similar preferences to reduce heterogeneity and using multiple binary selectors to enhance LLM output quality. To evaluate the performance of the proposed methods, we establish the first federated RLHF benchmark with a heterogeneous human preference dataset. Experimental results show that by integrating the LLM with aggregated client preferences, FedBis and FedBiscuit significantly enhance the professionalism and readability of the generated content.", "authors": ["Feijie Wu", "Xiaoze Liu", "Haoyu Wang", "Xingchen Wang", "Lu Su", "Jing Gao"], "categories": ["cs.CL", "cs.DC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-03", "url": "https://arxiv.org/abs/2407.03038", "pdf_url": "https://arxiv.org/pdf/2407.03038v3", "arxiv_id": "2407.03038", "doi": null, "citation_count": 17, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3138}
{"id": "6331e56046be56d47b2990fd87a7bfa25d99c9abed2354260fbbb13bbf8fb0f1", "sources": ["arxiv", "semantic_scholar"], "title": "Cost-Effective Proxy Reward Model Construction with On-Policy and Active Learning", "abstract": "Reinforcement learning with human feedback (RLHF), as a widely adopted approach in current large language model pipelines, is \\textit{bottlenecked by the size of human preference data}. While traditional methods rely on offline preference dataset constructions, recent approaches have shifted towards online settings, where a learner uses a small amount of labeled seed data and a large pool of unlabeled prompts to iteratively construct new preference data through self-generated responses and high-quality reward/preference feedback. However, most current online algorithms still focus on preference labeling during policy model updating with given feedback oracles, which incurs significant expert query costs. \\textit{We are the first to explore cost-effective proxy reward oracles construction strategies for further labeling preferences or rewards with extremely limited labeled data and expert query budgets}. Our approach introduces two key innovations: (1) on-policy query to avoid OOD and imbalance issues in seed data, and (2) active learning to select the most informative data for preference queries. Using these methods, we train a evaluation model with minimal expert-labeled data, which then effectively labels nine times more preference pairs for further RLHF training. For instance, our model using Direct Preference Optimization (DPO) gains around over 1% average improvement on AlpacaEval2, MMLU-5shot and MMLU-0shot, with only 1.7K query cost. Our methodology is orthogonal to other direct expert query-based strategies and therefore might be integrated with them to further reduce query costs.", "authors": ["Yifang Chen", "Shuohang Wang", "Ziyi Yang", "Hiteshi Sharma", "Nikos Karampatziakis", "Donghan Yu", "Kevin Jamieson", "Simon Shaolei Du", "Yelong Shen"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-02", "url": "https://arxiv.org/abs/2407.02119", "pdf_url": "https://arxiv.org/pdf/2407.02119v2", "arxiv_id": "2407.02119", "doi": "10.48550/arXiv.2407.02119", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "2ee0380638a4076792e1678207e80a012ded386328cac8ae1c0157399a3f817a", "sources": ["arxiv", "semantic_scholar"], "title": "DogeRM: Equipping Reward Models with Domain Knowledge through Model Merging", "abstract": "Reinforcement learning from human feedback (RLHF) is a popular strategy for aligning large language models (LLMs) with desired behaviors. Reward modeling is a crucial step in RLHF. However, collecting paired preference data for training reward models is often costly and time-consuming, especially for domain-specific preferences requiring expert annotation. To address this challenge, we propose the \\textbf{Do}main knowled\\textbf{ge} merged \\textbf{R}eward \\textbf{M}odel (DogeRM), a novel framework that integrates domain-specific knowledge into a general reward model by model merging. The experiments demonstrate that DogeRM enhances performance across different benchmarks and provide a detailed analysis showcasing the effects of model merging, showing the great potential of facilitating model alignment.", "authors": ["Tzu-Han Lin", "Chen-An Li", "Hung-yi Lee", "Yun-Nung Chen"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-07-01", "url": "https://arxiv.org/abs/2407.01470", "pdf_url": "https://arxiv.org/pdf/2407.01470v2", "arxiv_id": "2407.01470", "doi": "10.48550/arXiv.2407.01470", "citation_count": 7, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/MiuLab/DogeRM", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.2258}
{"id": "d5cd102a5e297613170a1ecd828595a3062acaee14f3c3bd5a50aa9ee07edb51", "sources": ["arxiv", "semantic_scholar"], "title": "Revisiting Sparse Rewards for Goal-Reaching Reinforcement Learning", "abstract": "Many real-world robot learning problems, such as pick-and-place or arriving at a destination, can be seen as a problem of reaching a goal state as soon as possible. These problems, when formulated as episodic reinforcement learning tasks, can easily be specified to align well with our intended goal: -1 reward every time step with termination upon reaching the goal state, called minimum-time tasks. Despite this simplicity, such formulations are often overlooked in favor of dense rewards due to their perceived difficulty and lack of informativeness. Our studies contrast the two reward paradigms, revealing that the minimum-time task specification not only facilitates learning higher-quality policies but can also surpass dense-reward-based policies on their own performance metrics. Crucially, we also identify the goal-hit rate of the initial policy as a robust early indicator for learning success in such sparse feedback settings. Finally, using four distinct real-robotic platforms, we show that it is possible to learn pixel-based policies from scratch within two to three hours using constant negative rewards.", "authors": ["Gautham Vasan", "Yan Wang", "Fahim Shahriar", "James Bergstra", "Martin Jagersand", "A. Rupam Mahmood"], "categories": ["cs.RO", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-29", "url": "https://arxiv.org/abs/2407.00324", "pdf_url": "https://arxiv.org/pdf/2407.00324v2", "arxiv_id": "2407.00324", "doi": "10.48550/arXiv.2407.00324", "citation_count": 17, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3138}
{"id": "79566d7c03f3b9b0db7eb085d4bf4d4df34f87e689a51b575af9f7bc56d6dee9", "sources": ["arxiv", "semantic_scholar"], "title": "The Perils of Optimizing Learned Reward Functions: Low Training Error Does Not Guarantee Low Regret", "abstract": "In reinforcement learning, specifying reward functions that capture the intended task can be very challenging. Reward learning aims to address this issue by learning the reward function. However, a learned reward model may have a low error on the data distribution, and yet subsequently produce a policy with large regret. We say that such a reward model has an error-regret mismatch. The main source of an error-regret mismatch is the distributional shift that commonly occurs during policy optimization. In this paper, we mathematically show that a sufficiently low expected test error of the reward model guarantees low worst-case regret, but that for any fixed expected test error, there exist realistic data distributions that allow for error-regret mismatch to occur. We then show that similar problems persist even when using policy regularization techniques, commonly employed in methods such as RLHF. We hope our results stimulate the theoretical and empirical study of improved methods to learn reward models, and better ways to measure their quality reliably.", "authors": ["Lukas Fluri", "Leon Lang", "Alessandro Abate", "Patrick Forré", "David Krueger", "Joar Skalse"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-22", "url": "https://arxiv.org/abs/2406.15753", "pdf_url": "https://arxiv.org/pdf/2406.15753v3", "arxiv_id": "2406.15753", "doi": "10.48550/arXiv.2406.15753", "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "2909befe743a4a6a12dbef8669538a5fe987016bce3778aa608487cbb7873f12", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts", "abstract": "Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. The RLHF process typically starts by training a reward model (RM) using human preference data. Conventional RMs are trained on pairwise responses to the same user request, with relative ratings indicating which response humans prefer. The trained RM serves as a proxy for human preferences. However, due to the black-box nature of RMs, their outputs lack interpretability, as humans cannot intuitively understand why an RM thinks a response is good or not. As RMs act as human preference proxies, we believe they should be human-interpretable to ensure that their internal decision processes are consistent with human preferences and to prevent reward hacking in LLM alignment. To build RMs with interpretable preferences, we propose a two-stage approach: i) train an Absolute-Rating Multi-Objective Reward Model (ArmoRM) with multi-dimensional absolute-rating data, each dimension corresponding to a human-interpretable objective (e.g., honesty, verbosity, safety); ii) employ a Mixture-of-Experts (MoE) strategy with a gating network that automatically selects the most suitable reward objectives based on the context. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow MLP on top of the ArmoRM. Our trained model, ArmoRM-Llama3-8B, obtains state-of-the-art performance on RewardBench, a benchmark evaluating RMs for language modeling. Notably, the performance of our model surpasses the LLM-as-a-judge method with GPT-4 judges by a margin, and approaches the performance of the much larger Nemotron-4 340B reward model.", "authors": ["Haoxiang Wang", "Wei Xiong", "Tengyang Xie", "Han Zhao", "Tong Zhang"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-18", "url": "https://arxiv.org/abs/2406.12845", "pdf_url": "https://arxiv.org/pdf/2406.12845v1", "arxiv_id": "2406.12845", "doi": "10.48550/arXiv.2406.12845", "citation_count": 387, "influential_citation_count": 37, "has_code": true, "code_url": "https://github.com/RLHFlow/RLHF-Reward-Modeling/", "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.7899}
{"id": "c568da5d93308b366f63b010c4a5c91f647d60dc29bdfc50d3010f3b2ee583f7", "sources": ["arxiv", "semantic_scholar"], "title": "Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models", "abstract": "In reinforcement learning, specification gaming occurs when AI systems learn undesired behaviors that are highly rewarded due to misspecified training goals. Specification gaming can range from simple behaviors like sycophancy to sophisticated and pernicious behaviors like reward-tampering, where a model directly modifies its own reward mechanism. However, these more pernicious behaviors may be too complex to be discovered via exploration. In this paper, we study whether Large Language Model (LLM) assistants which find easily discovered forms of specification gaming will generalize to perform rarer and more blatant forms, up to and including reward-tampering. We construct a curriculum of increasingly sophisticated gameable environments and find that training on early-curriculum environments leads to more specification gaming on remaining environments. Strikingly, a small but non-negligible proportion of the time, LLM assistants trained on the full curriculum generalize zero-shot to directly rewriting their own reward function. Retraining an LLM not to game early-curriculum environments mitigates, but does not eliminate, reward-tampering in later environments. Moreover, adding harmlessness training to our gameable environments does not prevent reward-tampering. These results demonstrate that LLMs can generalize from common forms of specification gaming to more pernicious reward tampering and that such behavior may be nontrivial to remove.", "authors": ["Carson Denison", "Monte MacDiarmid", "Fazl Barez", "David Duvenaud", "Shauna Kravec", "Samuel Marks", "Nicholas Schiefer", "Ryan Soklaski", "Alex Tamkin", "Jared Kaplan", "Buck Shlegeris", "Samuel R. Bowman", "Ethan Perez", "Evan Hubinger"], "categories": ["cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-14", "url": "https://arxiv.org/abs/2406.10162", "pdf_url": "https://arxiv.org/pdf/2406.10162v3", "arxiv_id": "2406.10162", "doi": "10.48550/arXiv.2406.10162", "citation_count": 128, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5276}
{"id": "fa75d4bee19acbc74765ea888ef70f6e489005243d9a463cd5961778aeac475b", "sources": ["arxiv", "semantic_scholar"], "title": "Binary Reward Labeling: Bridging Offline Preference and Reward-Based Reinforcement Learning", "abstract": "Offline reinforcement learning has become one of the most practical RL settings. However, most existing works on offline RL focus on the standard setting with scalar reward feedback. It remains unknown how to universally transfer the existing rich understanding of offline RL from the reward-based to the preference-based setting. In this work, we propose a general framework to bridge this gap. Our key insight is transforming preference feedback to scalar rewards via binary reward labeling (BRL), and then any reward-based offline RL algorithms can be applied to the dataset with the reward labels. The information loss during the feedback signal transition is minimized with binary reward labeling in the practical learning scenarios. We theoretically show the connection between several recent PBRL techniques and our framework combined with specific offline RL algorithms. By combining reward labeling with different algorithms, our framework can lead to new and potentially more efficient offline PBRL algorithms. We empirically test our framework on preference datasets based on the standard D4RL benchmark. When combined with a variety of efficient reward-based offline RL algorithms, the learning result achieved under our framework is comparable to training the same algorithm on the dataset with actual rewards in many cases and better than the recent PBRL baselines in most cases.", "authors": ["Yinglun Xu", "David Zhu", "Rohan Gumaste", "Gagandeep Singh"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-14", "url": "https://arxiv.org/abs/2406.10445", "pdf_url": "https://arxiv.org/pdf/2406.10445v3", "arxiv_id": "2406.10445", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "1a8e97045aa09bd740a53771330c0c99111899886044eee6b2a9c64d8da7a8c3", "sources": ["arxiv", "semantic_scholar"], "title": "Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs", "abstract": "Reward models trained on human preference data have been proven to effectively align Large Language Models (LLMs) with human intent within the framework of reinforcement learning from human feedback (RLHF). However, current reward models have limited generalization capabilities to unseen prompts and responses, which can lead to an unexpected phenomenon known as reward over-optimization, resulting in a decline in actual performance due to excessive optimization of rewards. While previous research has advocated for constraining policy optimization, our study introduces a novel approach to enhance the reward model's generalization ability against distribution shifts by regularizing the hidden states. Specifically, we retain the base model's language model head and incorporate a suite of text-generation losses to preserve the hidden states' text-generation capabilities, while concurrently learning a reward head behind the same hidden states. Our experimental results demonstrate that the introduced regularization technique markedly improves the accuracy of learned reward models across a variety of out-of-distribution (OOD) tasks and effectively alleviates the over-optimization issue in RLHF, offering a more reliable and robust preference learning paradigm.", "authors": ["Rui Yang", "Ruomeng Ding", "Yong Lin", "Huan Zhang", "Tong Zhang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-14", "url": "https://arxiv.org/abs/2406.10216", "pdf_url": "https://arxiv.org/pdf/2406.10216v2", "arxiv_id": "2406.10216", "doi": "10.48550/arXiv.2406.10216", "citation_count": 140, "influential_citation_count": 14, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.588}
{"id": "8258cf7b8369ae91a9d094e379a4ca9b33293cc2d832268a0870a402f597b78b", "sources": ["arxiv", "semantic_scholar"], "title": "Bootstrapping Language Models with DPO Implicit Rewards", "abstract": "Human alignment in large language models (LLMs) is an active area of research. A recent groundbreaking work, direct preference optimization (DPO), has greatly simplified the process from past work in reinforcement learning from human feedback (RLHF) by bypassing the reward learning stage in RLHF. DPO, after training, provides an implicit reward model. In this work, we make a novel observation that this implicit reward model can by itself be used in a bootstrapping fashion to further align the LLM. Our approach is to use the rewards from a current LLM to construct a preference dataset, which is then used in subsequent DPO rounds. We incorporate two refinements to further improve our approach: 1) length-regularized reward shaping to make the preference dataset length-unbiased; 2) experience replay to enhance the quality of the preference dataset. Our approach, named self-alignment with DPO ImpliCit rEwards (DICE), shows great improvements in alignment. It achieves an increase of more than 8$\\\\%$ in lengthcontrolled win rate on AlpacaEval 2 for all the different base models that we tried, without relying on external feedback. Our code is available at https://github.com/sail-sg/dice.", "authors": ["Changyu Chen", "Zichen Liu", "Chao Du", "Tianyu Pang", "Qian Liu", "Arunesh Sinha", "Pradeep Varakantham", "Min Lin"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-14", "url": "https://arxiv.org/abs/2406.09760", "pdf_url": "https://arxiv.org/pdf/2406.09760v2", "arxiv_id": "2406.09760", "doi": "10.48550/arXiv.2406.09760", "citation_count": 55, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/sail-sg/dice", "venue": "International Conference on Learning Representations", "quality_score": 0.437}
{"id": "df087e33fb42318afd31c6cdef8cbafed2d9b95e8f1d32c0ccdeffc078647865", "sources": ["arxiv", "semantic_scholar"], "title": "Online Bandit Learning with Offline Preference Data for Improved RLHF", "abstract": "Reinforcement Learning with Human Feedback (RLHF) is at the core of fine-tuning methods for generative AI models for language and images. Such feedback is often sought as rank or preference feedback from human raters, as opposed to eliciting scores since the latter tends to be noisy. On the other hand, RL theory and algorithms predominantly assume that a reward feedback is available. In particular, approaches for online learning that can be helpful in adaptive data collection via active learning cannot incorporate offline preference data. In this paper, we adopt a finite-armed linear bandit model as a prototypical model of online learning. We consider an offline preference dataset to be available generated by an expert of unknown 'competence'. We propose warmPref-PS, a posterior sampling algorithm for online learning that can be warm-started with an offline dataset with noisy preference feedback. We show that by modeling the 'competence' of the expert that generated it, we are able to use such a dataset most effectively. We support our claims with novel theoretical analysis of its Bayesian regret, as well as, extensive empirical evaluation of an approximate loss function that optimizes for infinitely many arms, and performs substantially better than baselines.", "authors": ["Akhil Agnihotri", "Rahul Jain", "Deepak Ramachandran", "Zheng Wen"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-13", "url": "https://arxiv.org/abs/2406.09574", "pdf_url": "https://arxiv.org/pdf/2406.09574v4", "arxiv_id": "2406.09574", "doi": "10.48550/arXiv.2406.09574", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "199cd7c0b0c797b2f0a2a53345805286d6981dc89200522d2eafaf4a81d6a208", "sources": ["arxiv", "semantic_scholar"], "title": "It Takes Two: On the Seamlessness between Reward and Policy Model in RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) involves training policy models (PMs) and reward models (RMs) to align language models with human preferences. Instead of focusing solely on PMs and RMs independently, we propose to examine their interactions during fine-tuning, introducing the concept of seamlessness. Our study starts with observing the saturation phenomenon, where continual improvements in RM and PM do not translate into RLHF progress. Our analysis shows that RMs fail to assign proper scores to PM responses, resulting in a 35% mismatch rate with human preferences, highlighting a significant discrepancy between PM and RM. To measure seamlessness between PM and RM without human effort, we propose an automatic metric, SEAM. SEAM quantifies the discrepancies between PM and RM judgments induced by data samples. We validate the effectiveness of SEAM in data selection and model augmentation. Our experiments demonstrate that (1) using SEAM-filtered data for RL training improves RLHF performance by 4.5%, and (2) SEAM-guided model augmentation results in a 4% performance improvement over standard augmentation methods.", "authors": ["Taiming Lu", "Lingfeng Shen", "Xinyu Yang", "Weiting Tan", "Beidi Chen", "Huaxiu Yao"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-12", "url": "https://arxiv.org/abs/2406.07971", "pdf_url": "https://arxiv.org/pdf/2406.07971v2", "arxiv_id": "2406.07971", "doi": "10.48550/arXiv.2406.07971", "citation_count": 5, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "061a4bf5692061f8e46a6bdac3d9c795069665e61f9642de0d68279e5b82f058", "sources": ["arxiv", "semantic_scholar"], "title": "Reinforcement Learning from Human Feedback without Reward Inference: Model-Free Algorithm and Instance-Dependent Analysis", "abstract": "In this paper, we study reinforcement learning from human feedback (RLHF) under an episodic Markov decision process with a general trajectory-wise reward model. We developed a model-free RLHF best policy identification algorithm, called $\\mathsf{BSAD}$, without explicit reward model inference, which is a critical intermediate step in the contemporary RLHF paradigms for training large language models (LLM). The algorithm identifies the optimal policy directly from human preference information in a backward manner, employing a dueling bandit sub-routine that constantly duels actions to identify the superior one. $\\mathsf{BSAD}$ adopts a reward-free exploration and best-arm-identification-like adaptive stopping criteria to equalize the visitation among all states in the same decision step while moving to the previous step as soon as the optimal action is identifiable, leading to a provable, instance-dependent sample complexity $\\tilde{\\mathcal{O}}(c_{\\mathcal{M}}SA^3H^3M\\log\\frac{1}δ)$ which resembles the result in classic RL, where $c_{\\mathcal{M}}$ is the instance-dependent constant and $M$ is the batch size. Moreover, $\\mathsf{BSAD}$ can be transformed into an explore-then-commit algorithm with logarithmic regret and generalized to discounted MDPs using a frame-based approach. Our results show: (i) sample-complexity-wise, RLHF is not significantly harder than classic RL and (ii) end-to-end RLHF may deliver improved performance by avoiding pitfalls in reward inferring such as overfit and distribution shift.", "authors": ["Qining Zhang", "Honghao Wei", "Lei Ying"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-11", "url": "https://arxiv.org/abs/2406.07455", "pdf_url": "https://arxiv.org/pdf/2406.07455v2", "arxiv_id": "2406.07455", "doi": "10.48550/arXiv.2406.07455", "citation_count": 4, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1747}
{"id": "2acc90a5fd42c173cb9b5ac4e2533a78ae6c237d136f060f906d20b37b647c26", "sources": ["arxiv", "semantic_scholar"], "title": "Pretraining Decision Transformers with Reward Prediction for In-Context Multi-task Structured Bandit Learning", "abstract": "We study learning to learn for the multi-task structured bandit problem where the goal is to learn a near-optimal algorithm that minimizes cumulative regret. The tasks share a common structure and an algorithm should exploit the shared structure to minimize the cumulative regret for an unseen but related test task. We use a transformer as a decision-making algorithm to learn this shared structure from data collected by a demonstrator on a set of training task instances. Our objective is to devise a training procedure such that the transformer will learn to outperform the demonstrator's learning algorithm on unseen test task instances. Prior work on pretraining decision transformers either requires privileged information like access to optimal arms or cannot outperform the demonstrator. Going beyond these approaches, we introduce a pre-training approach that trains a transformer network to learn a near-optimal policy in-context. This approach leverages the shared structure across tasks, does not require access to optimal actions, and can outperform the demonstrator. We validate these claims over a wide variety of structured bandit problems to show that our proposed solution is general and can quickly identify expected rewards on unseen test tasks to support effective exploration.", "authors": ["Subhojyoti Mukherjee", "Josiah P. Hanna", "Qiaomin Xie", "Robert Nowak"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-07", "url": "https://arxiv.org/abs/2406.05064", "pdf_url": "https://arxiv.org/pdf/2406.05064v3", "arxiv_id": "2406.05064", "doi": "10.48550/arXiv.2406.05064", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "87b3307cd7739ba403aec1975fdf5368e0e45e0d95910a15bcb1cab20e6edfa5", "sources": ["arxiv", "semantic_scholar"], "title": "Prototypical Reward Network for Data-Efficient RLHF", "abstract": "The reward model for Reinforcement Learning from Human Feedback (RLHF) has proven effective in fine-tuning Large Language Models (LLMs). Notably, collecting human feedback for RLHF can be resource-intensive and lead to scalability issues for LLMs and complex tasks. Our proposed framework Proto-RM leverages prototypical networks to enhance reward models under limited human feedback. By enabling stable and reliable structural learning from fewer samples, Proto-RM significantly enhances LLMs' adaptability and accuracy in interpreting human preferences. Extensive experiments on various datasets demonstrate that Proto-RM significantly improves the performance of reward models and LLMs in human feedback tasks, achieving comparable and usually better results than traditional methods, while requiring significantly less data. in data-limited scenarios. This research offers a promising direction for enhancing the efficiency of reward models and optimizing the fine-tuning of language models under restricted feedback conditions.", "authors": ["Jinghan Zhang", "Xiting Wang", "Yiqiao Jin", "Changyu Chen", "Xinhao Zhang", "Kunpeng Liu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-06", "url": "https://arxiv.org/abs/2406.06606", "pdf_url": "https://arxiv.org/pdf/2406.06606v2", "arxiv_id": "2406.06606", "doi": "10.48550/arXiv.2406.06606", "citation_count": 33, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.3829}
{"id": "a2dd5157a43539724f09aa848e8424619583c3b3f5bdb287e554101f57ea5520", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has been crucial to the recent success of Large Language Models (LLMs), however, it is often a complex and brittle process. In the classical RLHF framework, a reward model is first trained to represent human preferences, which is in turn used by an online reinforcement learning (RL) algorithm to optimize the LLM. A prominent issue with such methods is reward over-optimization or reward hacking, where performance as measured by the learned proxy reward model increases, but true quality plateaus or even deteriorates. Direct Alignment Algorithms (DDAs) like Direct Preference Optimization have emerged as alternatives to the classical RLHF pipeline by circumventing the reward modeling phase. However, although DAAs do not use a separate proxy reward model, they still commonly deteriorate from over-optimization. While the so-called reward hacking phenomenon is not well-defined for DAAs, we still uncover similar trends: at higher KL budgets, DAA algorithms exhibit similar degradation patterns to their classic RLHF counterparts. In particular, we find that DAA methods deteriorate not only across a wide range of KL budgets but also often before even a single epoch of the dataset is completed. Through extensive empirical experimentation, this work formulates and formalizes the reward over-optimization or hacking problem for DAAs and explores its consequences across objectives, training regimes, and model scales.", "authors": ["Rafael Rafailov", "Yaswanth Chittepu", "Ryan Park", "Harshit Sikchi", "Joey Hejna", "Bradley Knox", "Chelsea Finn", "Scott Niekum"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-06-05", "url": "https://arxiv.org/abs/2406.02900", "pdf_url": "https://arxiv.org/pdf/2406.02900v2", "arxiv_id": "2406.02900", "doi": "10.48550/arXiv.2406.02900", "citation_count": 138, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5358}
{"id": "280d5bac51f0260f83277ac2c604e3e4672a98f82ef8fabb83d159a18631fbae", "sources": ["arxiv", "semantic_scholar"], "title": "Towards the Transferability of Rewards Recovered via Regularized Inverse Reinforcement Learning", "abstract": "Inverse reinforcement learning (IRL) aims to infer a reward from expert demonstrations, motivated by the idea that the reward, rather than the policy, is the most succinct and transferable description of a task [Ng et al., 2000]. However, the reward corresponding to an optimal policy is not unique, making it unclear if an IRL-learned reward is transferable to new transition laws in the sense that its optimal policy aligns with the optimal policy corresponding to the expert's true reward. Past work has addressed this problem only under the assumption of full access to the expert's policy, guaranteeing transferability when learning from two experts with the same reward but different transition laws that satisfy a specific rank condition [Rolland et al., 2022]. In this work, we show that the conditions developed under full access to the expert's policy cannot guarantee transferability in the more practical scenario where we have access only to demonstrations of the expert. Instead of a binary rank condition, we propose principal angles as a more refined measure of similarity and dissimilarity between transition laws. Based on this, we then establish two key results: 1) a sufficient condition for transferability to any transition laws when learning from at least two experts with sufficiently different transition laws, and 2) a sufficient condition for transferability to local changes in the transition law when learning from a single expert. Furthermore, we also provide a probably approximately correct (PAC) algorithm and an end-to-end analysis for learning transferable rewards from demonstrations of multiple experts.", "authors": ["Andreas Schlaginhaufen", "Maryam Kamgarpour"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-06-03", "url": "https://arxiv.org/abs/2406.01793", "pdf_url": "https://arxiv.org/pdf/2406.01793v2", "arxiv_id": "2406.01793", "doi": "10.48550/arXiv.2406.01793", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.2258}
{"id": "26b7e746c0d2dc57c0e09fa40642f65231053c8208c4070adec49dd6228abca5", "sources": ["arxiv", "semantic_scholar"], "title": "Provably Efficient Interactive-Grounded Learning with Personalized Reward", "abstract": "Interactive-Grounded Learning (IGL) [Xie et al., 2021] is a powerful framework in which a learner aims at maximizing unobservable rewards through interacting with an environment and observing reward-dependent feedback on the taken actions. To deal with personalized rewards that are ubiquitous in applications such as recommendation systems, Maghakian et al. [2022] study a version of IGL with context-dependent feedback, but their algorithm does not come with theoretical guarantees. In this work, we consider the same problem and provide the first provably efficient algorithms with sublinear regret under realizability. Our analysis reveals that the step-function estimator of prior work can deviate uncontrollably due to finite-sample effects. Our solution is a novel Lipschitz reward estimator which underestimates the true reward and enjoys favorable generalization performances. Building on this estimator, we propose two algorithms, one based on explore-then-exploit and the other based on inverse-gap weighting. We apply IGL to learning from image feedback and learning from text feedback, which are reward-free settings that arise in practice. Experimental results showcase the importance of using our Lipschitz reward estimator and the overall effectiveness of our algorithms.", "authors": ["Mengxiao Zhang", "Yuheng Zhang", "Haipeng Luo", "Paul Mineiro"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-31", "url": "https://arxiv.org/abs/2405.20677", "pdf_url": "https://arxiv.org/pdf/2405.20677v1", "arxiv_id": "2405.20677", "doi": "10.48550/arXiv.2405.20677", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.1505}
{"id": "d39f60b136ca642f8323e78ba8d03cf812b4268b675dda1dcc2aba94e23b75cf", "sources": ["arxiv", "semantic_scholar"], "title": "Group Robust Preference Optimization in Reward-free RLHF", "abstract": "Adapting large language models (LLMs) for specific tasks usually involves fine-tuning through reinforcement learning with human feedback (RLHF) on preference data. While these data often come from diverse labelers' groups (e.g., different demographics, ethnicities, company teams, etc.), traditional RLHF approaches adopt a \"one-size-fits-all\" approach, i.e., they indiscriminately assume and optimize a single preference model, thus not being robust to unique characteristics and needs of the various groups. To address this limitation, we propose a novel Group Robust Preference Optimization (GRPO) method to align LLMs to individual groups' preferences robustly. Our approach builds upon reward-free direct preference optimization methods, but unlike previous approaches, it seeks a robust policy which maximizes the worst-case group performance. To achieve this, GRPO adaptively and sequentially weights the importance of different groups, prioritizing groups with worse cumulative loss. We theoretically study the feasibility of GRPO and analyze its convergence for the log-linear policy class. By fine-tuning LLMs with GRPO using diverse group-based global opinion data, we significantly improved performance for the worst-performing groups, reduced loss imbalances across groups, and improved probability accuracies compared to non-robust baselines.", "authors": ["Shyam Sundhar Ramesh", "Yifan Hu", "Iason Chaimalas", "Viraj Mehta", "Pier Giuseppe Sessa", "Haitham Bou Ammar", "Ilija Bogunovic"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-30", "url": "https://arxiv.org/abs/2405.20304", "pdf_url": "https://arxiv.org/pdf/2405.20304v1", "arxiv_id": "2405.20304", "doi": "10.48550/arXiv.2405.20304", "citation_count": 108, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5094}
{"id": "0029623d8ab411a0c4d4b4a6219b952ae601f99e6f585ad609de030570f977a2", "sources": ["arxiv", "semantic_scholar"], "title": "Robust Preference Optimization through Reward Model Distillation", "abstract": "Language model (LM) post-training (or alignment) involves maximizing a reward function that is derived from preference annotations. Direct Preference Optimization (DPO) is a popular offline alignment method that trains a policy directly on preference data without the need to train a reward model or apply reinforcement learning. However, the empirical evidence suggests that DPO typically assigns implicit rewards that overfit, and trend towards infinite magnitude. This frequently leads to degenerate policies, sometimes causing even the probabilities of the preferred generations to go to zero. In this work, we analyze this phenomenon and use distillation to get a better proxy for the true preference distribution over generation pairs: we train the LM such that its induced implicit reward, i.e., the scaled log-likelihood ratio of the model to the reference model, matches an explicit reward model trained on the preference data. Moreover, to account for uncertainty in the reward model we are distilling from, we optimize against a family of reward models that, as a whole, is likely to include at least one reasonable proxy for the preference distribution. Our results show that distilling from such a family of reward models leads to improved robustness to distribution shift in preference annotations, while preserving the simple supervised nature of DPO.", "authors": ["Adam Fisch", "Jacob Eisenstein", "Vicky Zayats", "Alekh Agarwal", "Ahmad Beirami", "Chirag Nagpal", "Pete Shaw", "Jonathan Berant"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-29", "url": "https://arxiv.org/abs/2405.19316", "pdf_url": "https://arxiv.org/pdf/2405.19316v2", "arxiv_id": "2405.19316", "doi": "10.48550/arXiv.2405.19316", "citation_count": 71, "influential_citation_count": 14, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.588}
{"id": "e495dd8c9890adc2b0920f7a4fddfedc667616945c177c0f38dea178744aee27", "sources": ["arxiv", "semantic_scholar"], "title": "Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF", "abstract": "Reinforcement learning from human feedback (RLHF) has demonstrated great promise in aligning large language models (LLMs) with human preference. Depending on the availability of preference data, both online and offline RLHF are active areas of investigation. A key bottleneck is understanding how to incorporate uncertainty estimation in the reward function learned from the preference data for RLHF, regardless of how the preference data is collected. While the principles of optimism or pessimism under uncertainty are well-established in standard reinforcement learning (RL), a practically-implementable and theoretically-grounded form amenable to large language models is not yet available, as standard techniques for constructing confidence intervals become intractable under arbitrary policy parameterizations. In this paper, we introduce a unified approach to online and offline RLHF -- value-incentivized preference optimization (VPO) -- which regularizes the maximum-likelihood estimate of the reward function with the corresponding value function, modulated by a $\\textit{sign}$ to indicate whether the optimism or pessimism is chosen. VPO also directly optimizes the policy with implicit reward modeling, and therefore shares a simpler RLHF pipeline similar to direct preference optimization. Theoretical guarantees of VPO are provided for both online and offline settings, matching the rates of their standard RL counterparts. Moreover, experiments on text summarization and dialog verify the practicality and effectiveness of VPO.", "authors": ["Shicong Cen", "Jincheng Mei", "Katayoon Goshvadi", "Hanjun Dai", "Tong Yang", "Sherry Yang", "Dale Schuurmans", "Yuejie Chi", "Bo Dai"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-29", "url": "https://arxiv.org/abs/2405.19320", "pdf_url": "https://arxiv.org/pdf/2405.19320v4", "arxiv_id": "2405.19320", "doi": "10.48550/arXiv.2405.19320", "citation_count": 72, "influential_citation_count": 11, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.5396}
{"id": "efba60c17c4a2c6b704d41e2f9de0a5039563681d98a7d855c9a17fc700a2702", "sources": ["arxiv", "semantic_scholar"], "title": "On the Algorithmic Bias of Aligning Large Language Models with RLHF: Preference Collapse and Matching Regularization", "abstract": "Accurately aligning large language models (LLMs) with human preferences is crucial for informing fair, economically sound, and statistically efficient decision-making processes. However, we argue that the predominant approach for aligning LLMs with human preferences through a reward model -- reinforcement learning from human feedback (RLHF) -- suffers from an inherent algorithmic bias due to its Kullback--Leibler-based regularization in optimization. In extreme cases, this bias could lead to a phenomenon we term preference collapse, where minority preferences are virtually disregarded. To mitigate this algorithmic bias, we introduce preference matching (PM) RLHF, a novel approach that provably aligns LLMs with the preference distribution of the reward model under the Bradley--Terry--Luce/Plackett--Luce model. Central to our approach is a PM regularizer that takes the form of the negative logarithm of the LLM's policy probability distribution over responses, which helps the LLM balance response diversification and reward maximization. Notably, we obtain this regularizer by solving an ordinary differential equation that is necessary for the PM property. For practical implementation, we introduce a conditional variant of PM RLHF that is tailored to natural language generation. Finally, we empirically validate the effectiveness of conditional PM RLHF through experiments on the OPT and Llama-family models, demonstrating a 29% to 41% improvement in alignment with human preferences, as measured by a certain metric, compared to standard RLHF.", "authors": ["Jiancong Xiao", "Ziniu Li", "Xingyu Xie", "Emily Getzen", "Cong Fang", "Qi Long", "Weijie J. Su"], "categories": ["stat.ML", "cs.LG", "stat.ME"], "fields_of_study": ["Mathematics", "Computer Science", "Medicine"], "published_date": "2024-05-26", "url": "https://arxiv.org/abs/2405.16455", "pdf_url": "https://arxiv.org/pdf/2405.16455v2", "arxiv_id": "2405.16455", "doi": "10.1080/01621459.2025.2555067", "citation_count": 76, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Journal of the American Statistical Association", "quality_score": 0.4716}
{"id": "9de144568eb6ef7af877f9690fbdc9f236799424ae7f1ce637024457109fb896", "sources": ["arxiv", "semantic_scholar"], "title": "Pragmatic Feature Preferences: Learning Reward-Relevant Preferences from Human Input", "abstract": "Humans use social context to specify preferences over behaviors, i.e. their reward functions. Yet, algorithms for inferring reward models from preference data do not take this social learning view into account. Inspired by pragmatic human communication, we study how to extract fine-grained data regarding why an example is preferred that is useful for learning more accurate reward models. We propose to enrich binary preference queries to ask both (1) which features of a given example are preferable in addition to (2) comparisons between examples themselves. We derive an approach for learning from these feature-level preferences, both for cases where users specify which features are reward-relevant, and when users do not. We evaluate our approach on linear bandit settings in both vision- and language-based domains. Results support the efficiency of our approach in quickly converging to accurate rewards with fewer comparisons vs. example-only labels. Finally, we validate the real-world applicability with a behavioral experiment on a mushroom foraging task. Our findings suggest that incorporating pragmatic feature preferences is a promising approach for more efficient user-aligned reward learning.", "authors": ["Andi Peng", "Yuying Sun", "Tianmin Shu", "David Abel"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14769", "pdf_url": "https://arxiv.org/pdf/2405.14769v1", "arxiv_id": "2405.14769", "doi": "10.48550/arXiv.2405.14769", "citation_count": 7, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2258}
{"id": "9dad6963c64685edc3ee8a26936d8e1a9192c5733cf78ffee3b68defd4fe03a2", "sources": ["arxiv", "semantic_scholar"], "title": "SimPO: Simple Preference Optimization with a Reference-Free Reward", "abstract": "Direct Preference Optimization (DPO) is a widely used offline preference optimization algorithm that reparameterizes reward functions in reinforcement learning from human feedback (RLHF) to enhance simplicity and training stability. In this work, we propose SimPO, a simpler yet more effective approach. The effectiveness of SimPO is attributed to a key design: using the average log probability of a sequence as the implicit reward. This reward formulation better aligns with model generation and eliminates the need for a reference model, making it more compute and memory efficient. Additionally, we introduce a target reward margin to the Bradley-Terry objective to encourage a larger margin between the winning and losing responses, further improving the algorithm's performance. We compare SimPO to DPO and its latest variants across various state-of-the-art training setups, including both base and instruction-tuned models such as Mistral, Llama 3, and Gemma 2. We evaluate on extensive chat-based evaluation benchmarks, including AlpacaEval 2, MT-Bench, and Arena-Hard. Our results demonstrate that SimPO consistently and significantly outperforms existing approaches without substantially increasing response length. Specifically, SimPO outperforms DPO by up to 6.4 points on AlpacaEval 2 and by up to 7.5 points on Arena-Hard. Our top-performing model, built on Gemma-2-9B-it, achieves a 72.4% length-controlled win rate on AlpacaEval 2, a 59.1% win rate on Arena-Hard, and ranks 1st on Chatbot Arena among <10B models with real user votes.", "authors": ["Yu Meng", "Mengzhou Xia", "Danqi Chen"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-23", "url": "https://arxiv.org/abs/2405.14734", "pdf_url": "https://arxiv.org/pdf/2405.14734v3", "arxiv_id": "2405.14734", "doi": "10.48550/arXiv.2405.14734", "citation_count": 1034, "influential_citation_count": 211, "has_code": true, "code_url": "https://github.com/princeton-nlp/SimPO", "venue": "Neural Information Processing Systems", "quality_score": 1.0}
{"id": "e7eab5d02ade3a25fbd3128423f53f945f1f9813d40c584f43c7cc92286e1d34", "sources": ["arxiv", "semantic_scholar"], "title": "LIRE: listwise reward enhancement for preference alignment", "abstract": "Recently, tremendous strides have been made to align the generation of Large Language Models (LLMs) with human values to mitigate toxic or unhelpful content. Leveraging Reinforcement Learning from Human Feedback (RLHF) proves effective and is widely adopted by researchers. However, implementing RLHF is complex, and its sensitivity to hyperparameters renders achieving stable performance and scalability challenging. Furthermore, prevailing approaches to preference alignment primarily concentrate on pairwise comparisons, with limited exploration into multi-response scenarios, thereby overlooking the potential richness within the candidate pool. For the above reasons, we propose a new approach: Listwise Reward Enhancement for Preference Alignment (LIRE), a gradient-based reward optimization approach that incorporates the offline rewards of multiple responses into a streamlined listwise framework, thus eliminating the need for online sampling during training. LIRE is straightforward to implement, requiring minimal parameter tuning, and seamlessly aligns with the pairwise paradigm while naturally extending to multi-response scenarios. Moreover, we introduce a self-enhancement algorithm aimed at iteratively refining the reward during training. Our experiments demonstrate that LIRE consistently outperforms existing methods across several benchmarks on dialogue and summarization tasks, with good transferability to out-of-distribution data, assessed using proxy reward models and human annotators.", "authors": ["Mingye Zhu", "Yi Liu", "Lei Zhang", "Junbo Guo", "Zhendong Mao"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-22", "url": "https://arxiv.org/abs/2405.13516", "pdf_url": "https://arxiv.org/pdf/2405.13516v2", "arxiv_id": "2405.13516", "doi": "10.48550/arXiv.2405.13516", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.25}
{"id": "2a5ba021ed9dda0e756925bfce66689efd25db990d94007ac97524642006e5e3", "sources": ["arxiv", "semantic_scholar"], "title": "SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling", "abstract": "Human preference alignment is critical in building powerful and reliable large language models (LLMs). However, current methods either ignore the multi-dimensionality of human preferences (e.g. helpfulness and harmlessness) or struggle with the complexity of managing multiple reward models. To address these issues, we propose Sequential Preference Optimization (SPO), a method that sequentially fine-tunes LLMs to align with multiple dimensions of human preferences. SPO avoids explicit reward modeling, directly optimizing the models to align with nuanced human preferences. We theoretically derive closed-form optimal SPO policy and loss function. Gradient analysis is conducted to show how SPO manages to fine-tune the LLMs while maintaining alignment on previously optimized dimensions. Empirical results on LLMs of different size and multiple evaluation datasets demonstrate that SPO successfully aligns LLMs across multiple dimensions of human preferences and significantly outperforms the baselines.", "authors": ["Xingzhou Lou", "Junge Zhang", "Jian Xie", "Lifeng Liu", "Dong Yan", "Kaiqi Huang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-05-21", "url": "https://arxiv.org/abs/2405.12739", "pdf_url": "https://arxiv.org/pdf/2405.12739v2", "arxiv_id": "2405.12739", "doi": "10.48550/arXiv.2405.12739", "citation_count": 22, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3404}
{"id": "a2f5e3e36eeb1ddaad4f8bd87036b7791196ea9ffdcfe158b9a7be7e4f5b4150", "sources": ["arxiv", "semantic_scholar"], "title": "A Unified Linear Programming Framework for Offline Reward Learning from Human Demonstrations and Feedback", "abstract": "Inverse Reinforcement Learning (IRL) and Reinforcement Learning from Human Feedback (RLHF) are pivotal methodologies in reward learning, which involve inferring and shaping the underlying reward function of sequential decision-making problems based on observed human demonstrations and feedback. Most prior work in reward learning has relied on prior knowledge or assumptions about decision or preference models, potentially leading to robustness issues. In response, this paper introduces a novel linear programming (LP) framework tailored for offline reward learning. Utilizing pre-collected trajectories without online exploration, this framework estimates a feasible reward set from the primal-dual optimality conditions of a suitably designed LP, and offers an optimality guarantee with provable sample efficiency. Our LP framework also enables aligning the reward functions with human feedback, such as pairwise trajectory comparison data, while maintaining computational tractability and sample efficiency. We demonstrate that our framework potentially achieves better performance compared to the conventional maximum likelihood estimation (MLE) approach through analytical examples and numerical experiments.", "authors": ["Kihyun Kim", "Jiawei Zhang", "Asuman Ozdaglar", "Pablo A. Parrilo"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-20", "url": "https://arxiv.org/abs/2405.12421", "pdf_url": "https://arxiv.org/pdf/2405.12421v3", "arxiv_id": "2405.12421", "doi": "10.48550/arXiv.2405.12421", "citation_count": 3, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.1505}
{"id": "6138ce6591fbd5e2db5e1e4a44f40b3adf4b22fd7f9561e07204d582574ef425", "sources": ["arxiv", "semantic_scholar"], "title": "RLHF Workflow: From Reward Modeling to Online RLHF", "abstract": "We present the workflow of Online Iterative Reinforcement Learning from Human Feedback (RLHF) in this technical report, which is widely reported to outperform its offline counterpart by a large margin in the recent large language model (LLM) literature. However, existing open-source RLHF projects are still largely confined to the offline learning setting. In this technical report, we aim to fill in this gap and provide a detailed recipe that is easy to reproduce for online iterative RLHF. In particular, since online human feedback is usually infeasible for open-source communities with limited resources, we start by constructing preference models using a diverse set of open-source datasets and use the constructed proxy preference model to approximate human feedback. Then, we discuss the theoretical insights and algorithmic principles behind online iterative RLHF, followed by a detailed practical implementation. Our trained LLM achieves impressive performance on LLM chatbot benchmarks, including AlpacaEval-2, Arena-Hard, and MT-Bench, as well as other academic benchmarks such as HumanEval and TruthfulQA. We have shown that supervised fine-tuning (SFT) and iterative RLHF can obtain state-of-the-art performance with fully open-source datasets. Further, we have made our models, curated datasets, and comprehensive step-by-step code guidebooks publicly available. Please refer to https://github.com/RLHFlow/RLHF-Reward-Modeling and https://github.com/RLHFlow/Online-RLHF for more detailed information.", "authors": ["Hanze Dong", "Wei Xiong", "Bo Pang", "Haoxiang Wang", "Han Zhao", "Yingbo Zhou", "Nan Jiang", "Doyen Sahoo", "Caiming Xiong", "Tong Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-05-13", "url": "https://arxiv.org/abs/2405.07863", "pdf_url": "https://arxiv.org/pdf/2405.07863v3", "arxiv_id": "2405.07863", "doi": "10.48550/arXiv.2405.07863", "citation_count": 259, "influential_citation_count": 45, "has_code": true, "code_url": "https://github.com/RLHFlow/RLHF-Reward-Modeling", "venue": null, "quality_score": 0.8314}
{"id": "f36e4c9ac0992510e78fb56ddc025e054bffe6866bbae648b460f2f6185881f5", "sources": ["arxiv", "semantic_scholar"], "title": "Numeric Reward Machines", "abstract": "Reward machines inform reinforcement learning agents about the reward structure of the environment and often drastically speed up the learning process. However, reward machines only accept Boolean features such as robot-reached-gold. Consequently, many inherently numeric tasks cannot profit from the guidance offered by reward machines. To address this gap, we aim to extend reward machines with numeric features such as distance-to-gold. For this, we present two types of reward machines: numeric-Boolean and numeric. In a numeric-Boolean reward machine, distance-to-gold is emulated by two Boolean features distance-to-gold-decreased and robot-reached-gold. In a numeric reward machine, distance-to-gold is used directly alongside the Boolean feature robot-reached-gold. We compare our new approaches to a baseline reward machine in the Craft domain, where the numeric feature is the agent-to-target distance. We use cross-product Q-learning, Q-learning with counter-factual experiences, and the options framework for learning. Our experimental results show that our new approaches significantly outperform the baseline approach. Extending reward machines with numeric features opens up new possibilities of using reward machines in inherently numeric tasks.", "authors": ["Kristina Levina", "Nikolaos Pappas", "Athanasios Karapantelakis", "Aneta Vulgarakis Feljan", "Jendrik Seipp"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-30", "url": "https://arxiv.org/abs/2404.19370", "pdf_url": "https://arxiv.org/pdf/2404.19370v1", "arxiv_id": "2404.19370", "doi": "10.48550/arXiv.2404.19370", "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "851b8c512b7a6c0443f852851e275dd774f1c8ac10f3e03037169360bde160bd", "sources": ["arxiv", "semantic_scholar"], "title": "RLHF from Heterogeneous Feedback via Personalization and Preference Aggregation", "abstract": "Reinforcement learning from human feedback (RLHF) has been an effective technique for aligning AI systems with human values, with remarkable successes in fine-tuning large-language models recently. Most existing RLHF paradigms make the underlying assumption that human preferences are relatively homogeneous, and can be encoded by a single reward model. In this paper, we focus on addressing the issues due to the inherent heterogeneity in human preferences, as well as their potential strategic behavior in providing feedback. Specifically, we propose two frameworks to address heterogeneous human feedback in principled ways: personalization-based one and aggregation-based one. For the former, we propose two approaches based on representation learning and clustering, respectively, for learning multiple reward models that trades off the bias (due to preference heterogeneity) and variance (due to the use of fewer data for learning each model by personalization). We then establish sample complexity guarantees for both approaches. For the latter, we aim to adhere to the single-model framework, as already deployed in the current RLHF paradigm, by carefully aggregating diverse and truthful preferences from humans. We propose two approaches based on reward and preference aggregation, respectively: the former utilizes both utilitarianism and Leximin approaches to aggregate individual reward models, with sample complexity guarantees; the latter directly aggregates the human feedback in the form of probabilistic opinions. Under the probabilistic-opinion-feedback model, we also develop an approach to handle strategic human labelers who may bias and manipulate the aggregated preferences with untruthful feedback. Based on the ideas in mechanism design, our approach ensures truthful preference reporting, with the induced aggregation rule maximizing social welfare functions.", "authors": ["Chanwoo Park", "Mingyang Liu", "Dingwen Kong", "Kaiqing Zhang", "Asuman Ozdaglar"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-30", "url": "https://arxiv.org/abs/2405.00254", "pdf_url": "https://arxiv.org/pdf/2405.00254v2", "arxiv_id": "2405.00254", "doi": "10.48550/arXiv.2405.00254", "citation_count": 76, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4716}
{"id": "f2034151cb507917c85736129458443f8debda5021b2e78760071a1ea1f91906", "sources": ["arxiv", "semantic_scholar"], "title": "Hindsight PRIORs for Reward Learning from Human Preferences", "abstract": "Preference based Reinforcement Learning (PbRL) removes the need to hand specify a reward function by learning a reward from preference feedback over policy behaviors. Current approaches to PbRL do not address the credit assignment problem inherent in determining which parts of a behavior most contributed to a preference, which result in data intensive approaches and subpar reward functions. We address such limitations by introducing a credit assignment strategy (Hindsight PRIOR) that uses a world model to approximate state importance within a trajectory and then guides rewards to be proportional to state importance through an auxiliary predicted return redistribution objective. Incorporating state importance into reward learning improves the speed of policy learning, overall policy performance, and reward recovery on both locomotion and manipulation tasks. For example, Hindsight PRIOR recovers on average significantly (p<0.05) more reward on MetaWorld (20%) and DMC (15%). The performance gains and our ablations demonstrate the benefits even a simple credit assignment strategy can have on reward learning and that state importance in forward dynamics prediction is a strong proxy for a state's contribution to a preference decision. Code repository can be found at https://github.com/apple/ml-rlhf-hindsight-prior.", "authors": ["Mudit Verma", "Katherine Metcalf"], "categories": ["cs.LG", "cs.AI", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-12", "url": "https://arxiv.org/abs/2404.08828", "pdf_url": "https://arxiv.org/pdf/2404.08828v1", "arxiv_id": "2404.08828", "doi": "10.48550/arXiv.2404.08828", "citation_count": 12, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/apple/ml-rlhf-hindsight-prior", "venue": "International Conference on Learning Representations", "quality_score": 0.2785}
{"id": "65986eb4af7ba1240a22f2bd61c0505c9d3a6d180c3297a18510bcbe6983ccae", "sources": ["arxiv", "semantic_scholar"], "title": "Best-of-Venom: Attacking RLHF by Injecting Poisoned Preference Data", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is a popular method for aligning Language Models (LM) with human values and preferences. RLHF requires a large number of preference pairs as training data, which are often used in both the Supervised Fine-Tuning and Reward Model training and therefore publicly available datasets are commonly used. In this work, we study to what extent a malicious actor can manipulate the LMs generations by poisoning the preferences, i.e., injecting poisonous preference pairs into these datasets and the RLHF training process. We propose strategies to build poisonous preference pairs and test their performance by poisoning two widely used preference datasets. Our results show that preference poisoning is highly effective: injecting a small amount of poisonous data (1-5\\% of the original dataset), we can effectively manipulate the LM to generate a target entity in a target sentiment (positive or negative). The findings from our experiments also shed light on strategies to defend against the preference poisoning attack.", "authors": ["Tim Baumgärtner", "Yang Gao", "Dana Alon", "Donald Metzler"], "categories": ["cs.CL", "cs.AI", "cs.CR", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-08", "url": "https://arxiv.org/abs/2404.05530", "pdf_url": "https://arxiv.org/pdf/2404.05530v2", "arxiv_id": "2404.05530", "doi": "10.48550/arXiv.2404.05530", "citation_count": 41, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4058}
{"id": "d19fab21c2fdf9a7145176122c4c71c3b4882b3b17ae194e813335e1b3421ad5", "sources": ["arxiv", "semantic_scholar"], "title": "Towards Understanding the Influence of Reward Margin on Preference Model Performance", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is a widely used framework for the training of language models. However, the process of using RLHF to develop a language model that is well-aligned presents challenges, especially when it comes to optimizing the reward model. Our research has found that existing reward models, when trained using the traditional ranking objective based on human preference data, often struggle to effectively distinguish between responses that are more or less favorable in real-world scenarios. To bridge this gap, our study introduces a novel method to estimate the preference differences without the need for detailed, exhaustive labels from human annotators. Our experimental results provide empirical evidence that incorporating margin values into the training process significantly improves the effectiveness of reward models. This comparative analysis not only demonstrates the superiority of our approach in terms of reward prediction accuracy but also highlights its effectiveness in practical applications.", "authors": ["Bowen Qin", "Duanyu Feng", "Xi Yang"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-04-07", "url": "https://arxiv.org/abs/2404.04932", "pdf_url": "https://arxiv.org/pdf/2404.04932v1", "arxiv_id": "2404.04932", "doi": "10.48550/arXiv.2404.04932", "citation_count": 11, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2698}
{"id": "22ba8ac251ef1cccc0afe6d786a0991cbe4e2b26bf6808868986b8533079f815", "sources": ["arxiv", "semantic_scholar"], "title": "Fine-Tuning Language Models with Reward Learning on Policy", "abstract": "Reinforcement learning from human feedback (RLHF) has emerged as an effective approach to aligning large language models (LLMs) to human preferences. RLHF contains three steps, i.e., human preference collecting, reward learning, and policy optimization, which are usually performed serially. Despite its popularity, however, (fixed) reward models may suffer from inaccurate off-distribution, since policy optimization continuously shifts LLMs' data distribution. Repeatedly collecting new preference data from the latest LLMs may alleviate this issue, which unfortunately makes the resulting system more complicated and difficult to optimize. In this paper, we propose reward learning on policy (RLP), an unsupervised framework that refines a reward model using policy samples to keep it on-distribution. Specifically, an unsupervised multi-view learning method is introduced to learn robust representations of policy samples. Meanwhile, a synthetic preference generation approach is developed to simulate high-quality preference data with policy outputs. Extensive experiments on three benchmark datasets show that RLP consistently outperforms the state-of-the-art. Our code is available at \\url{https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/rlp}.", "authors": ["Hao Lang", "Fei Huang", "Yongbin Li"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-28", "url": "https://arxiv.org/abs/2403.19279", "pdf_url": "https://arxiv.org/pdf/2403.19279v1", "arxiv_id": "2403.19279", "doi": "10.48550/arXiv.2403.19279", "citation_count": 15, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/rlp}", "venue": "North American Chapter of the Association for Computational Linguistics", "quality_score": 0.301}
{"id": "af4dae082ffe0398f54b011e9f4f8a7657e98e65519dedc234e444ce723559ec", "sources": ["arxiv", "semantic_scholar"], "title": "RewardBench: Evaluating Reward Models for Language Modeling", "abstract": "Reward models (RMs) are at the crux of successfully using RLHF to align pretrained models to human preferences, yet there has been relatively little study that focuses on evaluation of those models. Evaluating reward models presents an opportunity to understand the opaque technologies used for alignment of language models and which values are embedded in them. Resources for reward model training and understanding are sparse in the nascent open-source community around them. To enhance scientific understanding of reward models, we present RewardBench, a benchmark dataset and code-base for evaluation. The RewardBench dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety, to benchmark how reward models perform on challenging, structured and out-of-distribution queries. We create specific comparison datasets for RMs that have subtle, but verifiable reasons (e.g. bugs, incorrect facts) why one answer should be preferred to another. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods, such as the direct MLE training of classifiers and the implicit reward modeling of Direct Preference Optimization (DPO). We present many findings on propensity for refusals, reasoning limitations, and instruction following shortcomings of various reward models towards a better understanding of the RLHF process.", "authors": ["Nathan Lambert", "Valentina Pyatkin", "Jacob Morrison", "LJ Miranda", "Bill Yuchen Lin", "Khyathi Chandu", "Nouha Dziri", "Sachin Kumar", "Tom Zick", "Yejin Choi", "Noah A. Smith", "Hannaneh Hajishirzi"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-20", "url": "https://arxiv.org/abs/2403.13787", "pdf_url": "https://arxiv.org/pdf/2403.13787v2", "arxiv_id": "2403.13787", "doi": "10.48550/arXiv.2403.13787", "citation_count": 430, "influential_citation_count": 66, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.913}
{"id": "2a25b0ceb94ef48cc0e883f57963552916af6ae943ea3471ef0eeb3a9cf00b98", "sources": ["arxiv", "semantic_scholar"], "title": "The Value of Reward Lookahead in Reinforcement Learning", "abstract": "In reinforcement learning (RL), agents sequentially interact with changing environments while aiming to maximize the obtained rewards. Usually, rewards are observed only after acting, and so the goal is to maximize the expected cumulative reward. Yet, in many practical settings, reward information is observed in advance -- prices are observed before performing transactions; nearby traffic information is partially known; and goals are oftentimes given to agents prior to the interaction. In this work, we aim to quantifiably analyze the value of such future reward information through the lens of competitive analysis. In particular, we measure the ratio between the value of standard RL agents and that of agents with partial future-reward lookahead. We characterize the worst-case reward distribution and derive exact ratios for the worst-case reward expectations. Surprisingly, the resulting ratios relate to known quantities in offline RL and reward-free exploration. We further provide tight bounds for the ratio given the worst-case dynamics. Our results cover the full spectrum between observing the immediate rewards before acting to observing all the rewards before the interaction starts.", "authors": ["Nadav Merlis", "Dorian Baudry", "Vianney Perchet"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-03-18", "url": "https://arxiv.org/abs/2403.11637", "pdf_url": "https://arxiv.org/pdf/2403.11637v2", "arxiv_id": "2403.11637", "doi": "10.48550/arXiv.2403.11637", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.1945}
{"id": "da4435940eea50578b0e7094f02429978fd08ba2c02f40306307c1069b6bd7a8", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Reinforcement Learning from Human Feedback Using Contrastive Rewards", "abstract": "Reinforcement learning from human feedback (RLHF) is the mainstream paradigm used to align large language models (LLMs) with human preferences. Yet existing RLHF heavily relies on accurate and informative reward models, which are vulnerable and sensitive to noise from various sources, e.g. human labeling errors, making the pipeline fragile. In this work, we improve the effectiveness of the reward model by introducing a penalty term on the reward, named as \\textit{contrastive rewards}. %Contrastive rewards Our approach involves two steps: (1) an offline sampling step to obtain responses to prompts that serve as baseline calculation and (2) a contrastive reward calculated using the baseline responses and used in the Proximal Policy Optimization (PPO) step. We show that contrastive rewards enable the LLM to penalize reward uncertainty, improve robustness, encourage improvement over baselines, calibrate according to task difficulty, and reduce variance in PPO. We show empirically contrastive rewards can improve RLHF substantially, evaluated by both GPTs and humans, and our method consistently outperforms strong baselines.", "authors": ["Wei Shen", "Xiaoying Zhang", "Yuanshun Yao", "Rui Zheng", "Hongyi Guo", "Yang Liu"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-12", "url": "https://arxiv.org/abs/2403.07708", "pdf_url": "https://arxiv.org/pdf/2403.07708v2", "arxiv_id": "2403.07708", "doi": "10.48550/arXiv.2403.07708", "citation_count": 28, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3656}
{"id": "2400b981bea8af6c77d1732b3d10c92dbb0a906ac368e35271d4e59961129191", "sources": ["arxiv", "semantic_scholar"], "title": "A Generalized Acquisition Function for Preference-based Reward Learning", "abstract": "Preference-based reward learning is a popular technique for teaching robots and autonomous systems how a human user wants them to perform a task. Previous works have shown that actively synthesizing preference queries to maximize information gain about the reward function parameters improves data efficiency. The information gain criterion focuses on precisely identifying all parameters of the reward function. This can potentially be wasteful as many parameters may result in the same reward, and many rewards may result in the same behavior in the downstream tasks. Instead, we show that it is possible to optimize for learning the reward function up to a behavioral equivalence class, such as inducing the same ranking over behaviors, distribution over choices, or other related definitions of what makes two rewards similar. We introduce a tractable framework that can capture such definitions of similarity. Our experiments in a synthetic environment, an assistive robotics environment with domain transfer, and a natural language processing problem with real datasets demonstrate the superior performance of our querying method over the state-of-the-art information gain method.", "authors": ["Evan Ellis", "Gaurav R. Ghosal", "Stuart J. Russell", "Anca Dragan", "Erdem Bıyık"], "categories": ["cs.RO", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-09", "url": "https://arxiv.org/abs/2403.06003", "pdf_url": "https://arxiv.org/pdf/2403.06003v1", "arxiv_id": "2403.06003", "doi": "10.1109/ICRA57147.2024.10611472", "citation_count": 9, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Robotics and Automation", "quality_score": 0.25}
{"id": "a7ef3ec2eb6d65a780d02a5419bb0b5f655a62b0a0e32ad29f80211190ed7495", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Model Learning vs. Direct Policy Optimization: A Comparative Analysis of Learning from Human Preferences", "abstract": "In this paper, we take a step towards a deeper understanding of learning from human preferences by systematically comparing the paradigm of reinforcement learning from human feedback (RLHF) with the recently proposed paradigm of direct preference optimization (DPO). We focus our attention on the class of loglinear policy parametrization and linear reward functions. In order to compare the two paradigms, we first derive minimax statistical bounds on the suboptimality gap induced by both RLHF and DPO, assuming access to an oracle that exactly solves the optimization problems. We provide a detailed discussion on the relative comparison between the two paradigms, simultaneously taking into account the sample size, policy and reward class dimensions, and the regularization temperature. Moreover, we extend our analysis to the approximate optimization setting and derive exponentially decaying convergence rates for both RLHF and DPO. Next, we analyze the setting where the ground-truth reward is not realizable and find that, while RLHF incurs a constant additional error, DPO retains its asymptotically decaying gap by just tuning the temperature accordingly. Finally, we extend our comparison to the Markov decision process setting, where we generalize our results with exact optimization. To the best of our knowledge, we are the first to provide such a comparative analysis for RLHF and DPO.", "authors": ["Andi Nika", "Debmalya Mandal", "Parameswaran Kamalaruban", "Georgios Tzannetos", "Goran Radanović", "Adish Singla"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-03-04", "url": "https://arxiv.org/abs/2403.01857", "pdf_url": "https://arxiv.org/pdf/2403.01857v2", "arxiv_id": "2403.01857", "doi": "10.48550/arXiv.2403.01857", "citation_count": 22, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3404}
{"id": "fb3768f37c93486888a234f975a36cb7492f7e3feb0410c6b07074a0c9af6187", "sources": ["arxiv", "semantic_scholar"], "title": "Arithmetic Control of LLMs for Diverse User Preferences: Directional Preference Alignment with Multi-Objective Rewards", "abstract": "Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).", "authors": ["Haoxiang Wang", "Yong Lin", "Wei Xiong", "Rui Yang", "Shizhe Diao", "Shuang Qiu", "Han Zhao", "Tong Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-02-28", "url": "https://arxiv.org/abs/2402.18571", "pdf_url": "https://arxiv.org/pdf/2402.18571v3", "arxiv_id": "2402.18571", "doi": "10.48550/arXiv.2402.18571", "citation_count": 156, "influential_citation_count": 10, "has_code": true, "code_url": "https://github.com/Haoxiang-Wang/directional-preference-alignment", "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.549}
{"id": "660469780d947364845ff6ade89ddadf4535b61fd4313b835bcceb20aae1fc8f", "sources": ["arxiv", "semantic_scholar"], "title": "Sample-Efficient Preference-based Reinforcement Learning with Dynamics Aware Rewards", "abstract": "Preference-based reinforcement learning (PbRL) aligns a robot behavior with human preferences via a reward function learned from binary feedback over agent behaviors. We show that dynamics-aware reward functions improve the sample efficiency of PbRL by an order of magnitude. In our experiments we iterate between: (1) learning a dynamics-aware state-action representation (z^{sa}) via a self-supervised temporal consistency task, and (2) bootstrapping the preference-based reward function from (z^{sa}), which results in faster policy learning and better final policy performance. For example, on quadruped-walk, walker-walk, and cheetah-run, with 50 preference labels we achieve the same performance as existing approaches with 500 preference labels, and we recover 83\\% and 66\\% of ground truth reward policy performance versus only 38\\% and 21\\%. The performance gains demonstrate the benefits of explicitly learning a dynamics-aware reward model. Repo: \\texttt{https://github.com/apple/ml-reed}.", "authors": ["Katherine Metcalf", "Miguel Sarabia", "Natalie Mackraz", "Barry-John Theobald"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-28", "url": "https://arxiv.org/abs/2402.17975", "pdf_url": "https://arxiv.org/pdf/2402.17975v1", "arxiv_id": "2402.17975", "doi": "10.48550/arXiv.2402.17975", "citation_count": 11, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/apple/ml-reed}", "venue": "Conference on Robot Learning", "quality_score": 0.2698}
{"id": "470bb3525cfd25f7ea134b22022e1cf32776dac078deeb23bd5dc3c94ed06281", "sources": ["arxiv", "semantic_scholar"], "title": "Batch Active Learning of Reward Functions from Human Preferences", "abstract": "Data generation and labeling are often expensive in robot learning. Preference-based learning is a concept that enables reliable labeling by querying users with preference questions. Active querying methods are commonly employed in preference-based learning to generate more informative data at the expense of parallelization and computation time. In this paper, we develop a set of novel algorithms, batch active preference-based learning methods, that enable efficient learning of reward functions using as few data samples as possible while still having short query generation times and also retaining parallelizability. We introduce a method based on determinantal point processes (DPP) for active batch generation and several heuristic-based alternatives. Finally, we present our experimental results for a variety of robotics tasks in simulation. Our results suggest that our batch active learning algorithm requires only a few queries that are computed in a short amount of time. We showcase one of our algorithms in a study to learn human users' preferences.", "authors": ["Erdem Bıyık", "Nima Anari", "Dorsa Sadigh"], "categories": ["cs.LG", "cs.AI", "cs.RO", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-02-24", "url": "https://arxiv.org/abs/2402.15757", "pdf_url": "https://arxiv.org/pdf/2402.15757v1", "arxiv_id": "2402.15757", "doi": "10.1145/3649885", "citation_count": 16, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3076}
{"id": "98f7772b2986eea4fe555836b7e0fa2fe9790ca06f0c8896d758da976a860aec", "sources": ["arxiv", "semantic_scholar"], "title": "Leveraging Domain Knowledge for Efficient Reward Modelling in RLHF: A Case-Study in E-Commerce Opinion Summarization", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has become a dominating strategy in aligning Language Models (LMs) with human values/goals. The key to the strategy is learning a reward model ($\\varphi$), which can reflect the latent reward model of humans. While this strategy has proven effective, the training methodology requires a lot of human preference annotation (usually in the order of tens of thousands) to train $\\varphi$. Such a large-scale annotation is justifiable when it's a one-time effort, and the reward model is universally applicable. However, human goals are subjective and depend on the task, requiring task-specific preference annotations, which can be impractical to fulfill. To address this challenge, we propose a novel approach to infuse domain knowledge into $\\varphi$, which reduces the amount of preference annotation required ($21\\times$), omits Alignment Tax, and provides some interpretability. We validate our approach in E-Commerce Opinion Summarization, with a significant reduction in dataset size (to just $940$ samples) while advancing the SOTA ($\\sim4$ point ROUGE-L improvement, $68\\%$ of times preferred by humans over SOTA). Our contributions include a novel Reward Modeling technique and two new datasets: PromptOpinSumm (supervised data for Opinion Summarization) and OpinPref (a gold-standard human preference dataset). The proposed methodology opens up avenues for efficient RLHF, making it more adaptable to applications with varying human values. We release the artifacts (Code: github.com/efficient-rlhf. PromptOpinSumm: hf.co/prompt-opin-summ. OpinPref: hf.co/opin-pref) for usage under MIT License.", "authors": ["Swaroop Nath", "Tejpalsingh Siledar", "Sankara Sri Raghava Ravindra Muddu", "Rupasai Rangaraju", "Harshad Khadilkar", "Pushpak Bhattacharyya", "Suman Banerjee", "Amey Patil", "Sudhanshu Shekhar Singh", "Muthusamy Chelliah", "Nikesh Garera"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-23", "url": "https://arxiv.org/abs/2402.15473", "pdf_url": "https://arxiv.org/pdf/2402.15473v2", "arxiv_id": "2402.15473", "doi": "10.48550/arXiv.2402.15473", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "5e73c6751d285744c7ad2ebed826244f4d3e1cc3d5689a164eb7e5154b432e70", "sources": ["arxiv", "semantic_scholar"], "title": "Generalizing Reward Modeling for Out-of-Distribution Preference Learning", "abstract": "Preference learning (PL) with large language models (LLMs) aims to align the LLMs' generations with human preferences. Previous work on reinforcement learning from human feedback (RLHF) has demonstrated promising results in in-distribution PL. However, due to the difficulty of obtaining human feedback, discretely training reward models for every encountered distribution is challenging. Thus, out-of-distribution (OOD) PL is practically useful for enhancing the generalization ability of LLMs with limited preference feedback. This work addresses OOD PL by optimizing a general reward model through a meta-learning approach. During meta-training, a bilevel optimization algorithm is utilized to learn a reward model capable of guiding policy learning to align with human preferences across various distributions. When encountering a test distribution, the meta-test procedure conducts regularized policy optimization using the learned reward model for PL. We theoretically demonstrate the convergence rate of the bilevel optimization algorithm under reasonable assumptions. Additionally, we conduct experiments on two text generation tasks across 20 held-out domains and outperform a variety of strong baselines across various evaluation metrics.", "authors": ["Chen Jia"], "categories": ["cs.LG", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-22", "url": "https://arxiv.org/abs/2402.14760", "pdf_url": "https://arxiv.org/pdf/2402.14760v2", "arxiv_id": "2402.14760", "doi": "10.48550/arXiv.2402.14760", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2113}
{"id": "456352af445c65a95e7d8761b2c4a6b91289b39075a51160253ebf1cdd5e7b79", "sources": ["arxiv", "semantic_scholar"], "title": "Bayesian Reward Models for LLM Alignment", "abstract": "To ensure that large language model (LLM) responses are helpful and non-toxic, a reward model trained on human preference data is usually used. LLM responses with high rewards are then selected through best-of-$n$ (BoN) sampling or the LLM is further optimized to produce responses with high rewards through reinforcement learning from human feedback (RLHF). However, these processes are susceptible to reward overoptimization or `hacking', where responses receive high rewards due to imperfections in the reward model rather than true preference, particularly as prompts or responses deviate from the training data. To address these challenges, we propose to train a Bayesian reward model, which signals higher uncertainty further from the training data distribution. We trained Bayesian reward models using Laplace approximation on LoRA weights, and found that the resulting uncertainty estimates can effectively mitigate reward overoptimization in BoN sampling.", "authors": ["Adam X. Yang", "Maxime Robeyns", "Thomas Coste", "Zhengyan Shi", "Jun Wang", "Haitham Bou-Ammar", "Laurence Aitchison"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-20", "url": "https://arxiv.org/abs/2402.13210", "pdf_url": "https://arxiv.org/pdf/2402.13210v2", "arxiv_id": "2402.13210", "doi": "10.48550/arXiv.2402.13210", "citation_count": 36, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3921}
{"id": "6d2d18421de9b997bf35296abb61db11eec618b5fc3e167751302889da4de6a0", "sources": ["arxiv", "semantic_scholar"], "title": "Advancing Translation Preference Modeling with RLHF: A Step Towards Cost-Effective Solution", "abstract": "Faithfulness, expressiveness, and elegance is the constant pursuit in machine translation. However, traditional metrics like \\textit{BLEU} do not strictly align with human preference of translation quality. In this paper, we explore leveraging reinforcement learning with human feedback (\\textit{RLHF}) to improve translation quality. It is non-trivial to collect a large high-quality dataset of human comparisons between translations, especially for low-resource languages. To address this issue, we propose a cost-effective preference learning strategy, optimizing reward models by distinguishing between human and machine translations. In this manner, the reward model learns the deficiencies of machine translation compared to human and guides subsequent improvements in machine translation. Experimental results demonstrate that \\textit{RLHF} can effectively enhance translation quality and this improvement benefits other translation directions not trained with \\textit{RLHF}. Further analysis indicates that the model's language capabilities play a crucial role in preference learning. A reward model with strong language capabilities can more sensitively learn the subtle differences in translation quality and align better with real human translation preferences.", "authors": ["Nuo Xu", "Jun Zhao", "Can Zu", "Sixian Li", "Lu Chen", "Zhihao Zhang", "Rui Zheng", "Shihan Dou", "Wenjuan Qin", "Tao Gui", "Qi Zhang", "Xuanjing Huang"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-18", "url": "https://arxiv.org/abs/2402.11525", "pdf_url": "https://arxiv.org/pdf/2402.11525v3", "arxiv_id": "2402.11525", "doi": "10.48550/arXiv.2402.11525", "citation_count": 17, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3138}
{"id": "1976e36470cd3fe02d21de08e9fcfaa004c89beae9eb8d7253dd15c31508209c", "sources": ["arxiv", "semantic_scholar"], "title": "Multi Task Inverse Reinforcement Learning for Common Sense Reward", "abstract": "One of the challenges in applying reinforcement learning in a complex real-world environment lies in providing the agent with a sufficiently detailed reward function. Any misalignment between the reward and the desired behavior can result in unwanted outcomes. This may lead to issues like \"reward hacking\" where the agent maximizes rewards by unintended behavior. In this work, we propose to disentangle the reward into two distinct parts. A simple task-specific reward, outlining the particulars of the task at hand, and an unknown common-sense reward, indicating the expected behavior of the agent within the environment. We then explore how this common-sense reward can be learned from expert demonstrations. We first show that inverse reinforcement learning, even when it succeeds in training an agent, does not learn a useful reward function. That is, training a new agent with the learned reward does not impair the desired behaviors. We then demonstrate that this problem can be solved by training simultaneously on multiple tasks. That is, multi-task inverse reinforcement learning can be applied to learn a useful reward function.", "authors": ["Neta Glazer", "Aviv Navon", "Aviv Shamsian", "Ethan Fetaya"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-17", "url": "https://arxiv.org/abs/2402.11367", "pdf_url": "https://arxiv.org/pdf/2402.11367v2", "arxiv_id": "2402.11367", "doi": "10.48550/arXiv.2402.11367", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "a1b4dece3f641b0d68a83c59f9f5137f29c12f59caa825279f8fb138fb674a25", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Generalization in RLHF: A Topological Perspective", "abstract": "Existing alignment methods share a common topology of information flow, where reward information is collected from humans, modeled with preference learning, and used to tune language models. However, this shared topology has not been systematically characterized, nor have its alternatives been thoroughly explored, leaving the problems of low data efficiency and unreliable generalization unaddressed. As a solution, we introduce a theory of reward generalization in reinforcement learning from human feedback (RLHF), focusing on the topology of information flow at both macro and micro levels. At the macro level, we portray the RLHF information flow as an autoencoding process over behavior distributions, formalizing the RLHF objective of distributional consistency between human preference and model behavior. At the micro level, we present induced Bayesian networks to model the impact of dataset topologies on reward generalization. Combining analysis on both levels, we propose reward modeling from tree-structured preference information. It is shown to reduce reward uncertainty by up to $Θ(\\log n/\\log\\log n)$ times compared to baselines, where $n$ is the dataset size. Validation on three NLP tasks shows that it achieves an average win rate of 65% against baselines, thus improving reward generalization for free via topology design, while reducing the amount of data requiring annotation.", "authors": ["Tianyi Qiu", "Fanzhi Zeng", "Jiaming Ji", "Dong Yan", "Kaile Wang", "Jiayi Zhou", "Yang Han", "Josef Dai", "Xuehai Pan", "Yaodong Yang"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.DM"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-15", "url": "https://arxiv.org/abs/2402.10184", "pdf_url": "https://arxiv.org/pdf/2402.10184v7", "arxiv_id": "2402.10184", "doi": "10.18653/v1/2025.findings-acl.820", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Annual Meeting of the Association for Computational Linguistics", "quality_score": 0.25}
{"id": "15776f809a85473bc441f84f9028f3920e496375885c38652cba8cd3dddf2b44", "sources": ["arxiv", "semantic_scholar"], "title": "InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling", "abstract": "Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models with human values, reward hacking, also termed reward overoptimization, remains a critical challenge. This issue primarily arises from reward misgeneralization, where reward models (RMs) compute reward using spurious features that are irrelevant to human preferences. In this work, we tackle this problem from an information-theoretic perspective and propose a framework for reward modeling, namely InfoRM, by introducing a variational information bottleneck objective to filter out irrelevant information. Notably, we further identify a correlation between overoptimization and outliers in the IB latent space of InfoRM, establishing it as a promising tool for detecting reward overoptimization. Inspired by this finding, we propose the Cluster Separation Index (CSI), which quantifies deviations in the IB latent space, as an indicator of reward overoptimization to facilitate the development of online mitigation strategies. Extensive experiments on a wide range of settings and RM scales (70M, 440M, 1.4B, and 7B) demonstrate the effectiveness of InfoRM. Further analyses reveal that InfoRM's overoptimization detection mechanism is not only effective but also robust across a broad range of datasets, signifying a notable advancement in the field of RLHF. The code will be released upon acceptance.", "authors": ["Yuchun Miao", "Sen Zhang", "Liang Ding", "Rong Bao", "Lefei Zhang", "Dacheng Tao"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-14", "url": "https://arxiv.org/abs/2402.09345", "pdf_url": "https://arxiv.org/pdf/2402.09345v5", "arxiv_id": "2402.09345", "doi": "10.52202/079017-4270", "citation_count": 88, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5}
{"id": "36799779e7f85232ee255e18319c74bbdaf1743587e6b5db7c279749d756f271", "sources": ["arxiv", "semantic_scholar"], "title": "MaxMin-RLHF: Alignment with Diverse Human Preferences", "abstract": "Reinforcement Learning from Human Feedback (RLHF) aligns language models to human preferences by employing a singular reward model derived from preference data. However, such an approach overlooks the rich diversity of human preferences inherent in data collected from multiple users. In this work, we first derive an impossibility result of alignment with single reward RLHF, thereby highlighting its insufficiency in representing diverse human preferences. To provide an equitable solution to the problem, we learn a mixture of preference distributions via an expectation-maximization algorithm and propose a MaxMin alignment objective for policy learning inspired by the Egalitarian principle in social choice theory to better represent diverse human preferences. We elucidate the connection of our proposed approach to distributionally robust optimization and general utility RL, thereby highlighting the generality and robustness of our proposed solution. We present comprehensive experimental results on small-scale (GPT-2) and large-scale language models (with Tulu2-7B) and show the efficacy of the proposed approach in the presence of diversity among human preferences. Our algorithm achieves an average improvement of more than 16% in win-rates over conventional RLHF algorithms and improves the win-rate (accuracy) for minority groups by over 33% without compromising the performance of majority groups, showcasing the robustness and fairness of our approach. We remark that our findings in this work are not only limited to language models but also extend to reinforcement learning in general.", "authors": ["Souradip Chakraborty", "Jiahao Qiu", "Hui Yuan", "Alec Koppel", "Furong Huang", "Dinesh Manocha", "Amrit Singh Bedi", "Mengdi Wang"], "categories": ["cs.CL", "cs.AI", "cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-14", "url": "https://arxiv.org/abs/2402.08925", "pdf_url": "https://arxiv.org/pdf/2402.08925v2", "arxiv_id": "2402.08925", "doi": null, "citation_count": 117, "influential_citation_count": 13, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.5731}
{"id": "87177ba80f04820725beba973ce8cdb206604ac01d601f6b75165b9207ecc273", "sources": ["arxiv", "semantic_scholar"], "title": "PRDP: Proximal Reward Difference Prediction for Large-Scale Reward Finetuning of Diffusion Models", "abstract": "Reward finetuning has emerged as a promising approach to aligning foundation models with downstream objectives. Remarkable success has been achieved in the language domain by using reinforcement learning (RL) to maximize rewards that reflect human preference. However, in the vision domain, existing RL-based reward finetuning methods are limited by their instability in large-scale training, rendering them incapable of generalizing to complex, unseen prompts. In this paper, we propose Proximal Reward Difference Prediction (PRDP), enabling stable black-box reward finetuning for diffusion models for the first time on large-scale prompt datasets with over 100K prompts. Our key innovation is the Reward Difference Prediction (RDP) objective that has the same optimal solution as the RL objective while enjoying better training stability. Specifically, the RDP objective is a supervised regression objective that tasks the diffusion model with predicting the reward difference of generated image pairs from their denoising trajectories. We theoretically prove that the diffusion model that obtains perfect reward difference prediction is exactly the maximizer of the RL objective. We further develop an online algorithm with proximal updates to stably optimize the RDP objective. In experiments, we demonstrate that PRDP can match the reward maximization ability of well-established RL-based methods in small-scale training. Furthermore, through large-scale training on text prompts from the Human Preference Dataset v2 and the Pick-a-Pic v1 dataset, PRDP achieves superior generation quality on a diverse set of complex, unseen prompts whereas RL-based methods completely fail.", "authors": ["Fei Deng", "Qifei Wang", "Wei Wei", "Matthias Grundmann", "Tingbo Hou"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-13", "url": "https://arxiv.org/abs/2402.08714", "pdf_url": "https://arxiv.org/pdf/2402.08714v2", "arxiv_id": "2402.08714", "doi": "10.1109/CVPR52733.2024.00709", "citation_count": 41, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.4225}
{"id": "290bc2f81b543659be187667d3461ec03494f7509023186e2dc089a65eab9e3d", "sources": ["arxiv", "semantic_scholar"], "title": "Online Iterative Reinforcement Learning from Human Feedback with General Preference Model", "abstract": "We investigate Reinforcement Learning from Human Feedback (RLHF) in the context of a general preference oracle. In particular, we do not assume the existence of a reward function and an oracle preference signal drawn from the Bradley-Terry model as most of the prior works do. We consider a standard mathematical formulation, the reverse-KL regularized minimax game between two LLMs for RLHF under general preference oracle. The learning objective of this formulation is to find a policy so that it is consistently preferred by the KL-regularized preference oracle over any competing LLMs. We show that this framework is strictly more general than the reward-based one, and propose sample-efficient algorithms for both the offline learning from a pre-collected preference dataset and online learning where we can query the preference oracle along the way of training. Empirical studies verify the effectiveness of the proposed framework.", "authors": ["Chenlu Ye", "Wei Xiong", "Yuheng Zhang", "Hanze Dong", "Nan Jiang", "Tong Zhang"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-02-11", "url": "https://arxiv.org/abs/2402.07314", "pdf_url": "https://arxiv.org/pdf/2402.07314v3", "arxiv_id": "2402.07314", "doi": "10.52202/079017-2598", "citation_count": 47, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.4225}
{"id": "afea05c550631e67146000470050ae25e0493f77b762485e9ec77c5e8cbd3ae8", "sources": ["arxiv", "semantic_scholar"], "title": "ODIN: Disentangled Reward Mitigates Hacking in RLHF", "abstract": "In this work, we study the issue of reward hacking on the response length, a challenge emerging in Reinforcement Learning from Human Feedback (RLHF) on LLMs. A well-formatted, verbose but less helpful response from the LLMs can often deceive LLMs or even human evaluators to achieve high scores. The same issue also holds for some reward models in RL. To address the challenges in both training and evaluation, we establish a more reliable evaluation protocol for comparing different training configurations, which inspects the trade-off between LLM evaluation score and response length obtained by varying training hyperparameters. Based on this evaluation, we conduct large-scale studies, where the results shed insights into the efficacy of hyperparameters and tricks used in RL on mitigating length bias. We further propose to improve the reward model by jointly training two linear heads on shared feature representations to predict the rewards, one trained to correlate with length, and the other trained to decorrelate with length and therefore focus more on the actual content. We then discard the length head in RL to prevent reward hacking on length. Experiments demonstrate that our approach almost eliminates the reward correlation with length, and improves the obtained policy by a significant margin.", "authors": ["Lichang Chen", "Chen Zhu", "Davit Soselia", "Jiuhai Chen", "Tianyi Zhou", "Tom Goldstein", "Heng Huang", "Mohammad Shoeybi", "Bryan Catanzaro"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-11", "url": "https://arxiv.org/abs/2402.07319", "pdf_url": "https://arxiv.org/pdf/2402.07319v1", "arxiv_id": "2402.07319", "doi": "10.48550/arXiv.2402.07319", "citation_count": 127, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.5268}
{"id": "a25824b158133086b76ae0402be96282204df0e93ff67bb60f7ebe4c04443e6f", "sources": ["arxiv", "semantic_scholar"], "title": "Informativeness of Reward Functions in Reinforcement Learning", "abstract": "Reward functions are central in specifying the task we want a reinforcement learning agent to perform. Given a task and desired optimal behavior, we study the problem of designing informative reward functions so that the designed rewards speed up the agent's convergence. In particular, we consider expert-driven reward design settings where an expert or teacher seeks to provide informative and interpretable rewards to a learning agent. Existing works have considered several different reward design formulations; however, the key challenge is formulating a reward informativeness criterion that adapts w.r.t. the agent's current policy and can be optimized under specified structural constraints to obtain interpretable rewards. In this paper, we propose a novel reward informativeness criterion, a quantitative measure that captures how the agent's current policy will improve if it receives rewards from a specific reward function. We theoretically showcase the utility of the proposed informativeness criterion for adaptively designing rewards for an agent. Experimental results on two navigation tasks demonstrate the effectiveness of our adaptive reward informativeness criterion.", "authors": ["Rati Devidze", "Parameswaran Kamalaruban", "Adish Singla"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-10", "url": "https://arxiv.org/abs/2402.07019", "pdf_url": "https://arxiv.org/pdf/2402.07019v1", "arxiv_id": "2402.07019", "doi": "10.48550/arXiv.2402.07019", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Adaptive Agents and Multi-Agent Systems", "quality_score": 0.1747}
{"id": "c0fe6a7f35cef4e42d10f3735fe60331308b7f7263cb7cbf78b372a498c49ecf", "sources": ["arxiv", "semantic_scholar"], "title": "Explaining Learned Reward Functions with Counterfactual Trajectories", "abstract": "Learning rewards from human behaviour or feedback is a promising approach to aligning AI systems with human values but fails to consistently extract correct reward functions. Interpretability tools could enable users to understand and evaluate possible flaws in learned reward functions. We propose Counterfactual Trajectory Explanations (CTEs) to interpret reward functions in reinforcement learning by contrasting an original with a counterfactual partial trajectory and the rewards they each receive. We derive six quality criteria for CTEs and propose a novel Monte-Carlo-based algorithm for generating CTEs that optimises these quality criteria. Finally, we measure how informative the generated explanations are to a proxy-human model by training it on CTEs. CTEs are demonstrably informative for the proxy-human model, increasing the similarity between its predictions and the reward function on unseen trajectories. Further, it learns to accurately judge differences in rewards between trajectories and generalises to out-of-distribution examples. Although CTEs do not lead to a perfect understanding of the reward, our method, and more generally the adaptation of XAI methods, are presented as a fruitful approach for interpreting learned reward functions.", "authors": ["Jan Wehner", "Frans Oliehoek", "Luciano Cavalcante Siebert"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-07", "url": "https://arxiv.org/abs/2402.04856", "pdf_url": "https://arxiv.org/pdf/2402.04856v4", "arxiv_id": "2402.04856", "doi": "10.48550/arXiv.2402.04856", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "2aa9e89ca4d5a046cdf6c133fe8ecb5bd8dab5e6a61b5fa75fcffb927ff752f3", "sources": ["arxiv", "semantic_scholar"], "title": "Reinforcement Learning from Bagged Reward", "abstract": "In Reinforcement Learning (RL), it is commonly assumed that an immediate reward signal is generated for each action taken by the agent, helping the agent maximize cumulative rewards to obtain the optimal policy. However, in many real-world scenarios, designing immediate reward signals is difficult; instead, agents receive a single reward that is contingent upon a partial sequence or a complete trajectory. In this work, we define this challenging problem as RL from Bagged Reward (RLBR), where sequences of data are treated as bags with non-Markovian bagged rewards, leading to the formulation of Bagged Reward Markov Decision Processes (BRMDPs). Theoretically, we demonstrate that RLBR can be addressed by solving a standard MDP with properly redistributed bagged rewards allocated to each instance within a bag. Empirically, we find that reward redistribution becomes more challenging as the bag length increases, due to reduced informational granularity. Existing reward redistribution methods are insufficient to address these challenges. Therefore, we propose a novel reward redistribution method equipped with a bidirectional attention mechanism, enabling the accurate interpretation of contextual nuances and temporal dependencies within each bag. We experimentally demonstrate that our proposed method consistently outperforms existing approaches.", "authors": ["Yuting Tang", "Xin-Qiang Cai", "Yao-Xiang Ding", "Qiyu Wu", "Guoqing Liu", "Masashi Sugiyama"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-06", "url": "https://arxiv.org/abs/2402.03771", "pdf_url": "https://arxiv.org/pdf/2402.03771v3", "arxiv_id": "2402.03771", "doi": "10.48550/arXiv.2402.03771", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "7325666e47cd871e8f24b7388dbc8f722f1de0e188940d25ea9f0dfed8f3600d", "sources": ["arxiv", "semantic_scholar"], "title": "A Theoretical Framework for Partially Observed Reward-States in RLHF", "abstract": "The growing deployment of reinforcement learning from human feedback (RLHF) calls for a deeper theoretical investigation of its underlying models. The prevalent models of RLHF do not account for neuroscience-backed, partially-observed \"internal states\" that can affect human feedback, nor do they accommodate intermediate feedback during an interaction. Both of these can be instrumental in speeding up learning and improving alignment. To address these limitations, we model RLHF as reinforcement learning with partially observed reward-states (PORRL). We accommodate two kinds of feedback $-$ cardinal and dueling feedback. We first demonstrate that PORRL subsumes a wide class of RL problems, including traditional RL, RLHF, and reward machines. For cardinal feedback, we present two model-based methods (POR-UCRL, POR-UCBVI). We give both cardinal regret and sample complexity guarantees for the methods, showing that they improve over naive history-summarization. We then discuss the benefits of a model-free method like GOLF with naive history-summarization in settings with recursive internal states and dense intermediate feedback. For this purpose, we define a new history aware version of the Bellman-eluder dimension and give a new guarantee for GOLF in our setting, which can be exponentially sharper in illustrative examples. For dueling feedback, we show that a naive reduction to cardinal feedback fails to achieve sublinear dueling regret. We then present the first explicit reduction that converts guarantees for cardinal regret to dueling regret. In both feedback settings, we show that our models and guarantees generalize and extend existing ones.", "authors": ["Chinmaya Kausik", "Mirco Mutti", "Aldo Pacchiano", "Ambuj Tewari"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-02-05", "url": "https://arxiv.org/abs/2402.03282", "pdf_url": "https://arxiv.org/pdf/2402.03282v3", "arxiv_id": "2402.03282", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "e6e8562837170ab283e123aba36f4f3acd8344ecdb666fb994b5049d61340dcb", "sources": ["arxiv", "semantic_scholar"], "title": "To the Max: Reinventing Reward in Reinforcement Learning", "abstract": "In reinforcement learning (RL), different reward functions can define the same optimal policy but result in drastically different learning performance. For some, the agent gets stuck with a suboptimal behavior, and for others, it solves the task efficiently. Choosing a good reward function is hence an extremely important yet challenging problem. In this paper, we explore an alternative approach for using rewards for learning. We introduce \\textit{max-reward RL}, where an agent optimizes the maximum rather than the cumulative reward. Unlike earlier works, our approach works for deterministic and stochastic environments and can be easily combined with state-of-the-art RL algorithms. In the experiments, we study the performance of max-reward RL algorithms in two goal-reaching environments from Gymnasium-Robotics and demonstrate its benefits over standard RL. The code is available at https://github.com/veviurko/To-the-Max.", "authors": ["Grigorii Veviurko", "Wendelin Böhmer", "Mathijs de Weerdt"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-02", "url": "https://arxiv.org/abs/2402.01361", "pdf_url": "https://arxiv.org/pdf/2402.01361v2", "arxiv_id": "2402.01361", "doi": "10.48550/arXiv.2402.01361", "citation_count": 15, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/veviurko/To-the-Max", "venue": "International Conference on Machine Learning", "quality_score": 0.301}
{"id": "c33ac5f5bf21455655482789cf38aa5f6a5b01c39bbfa7ba3024ad21e6826cc4", "sources": ["arxiv", "semantic_scholar"], "title": "Preference Poisoning Attacks on Reward Model Learning", "abstract": "Learning reward models from pairwise comparisons is a fundamental component in a number of domains, including autonomous control, conversational agents, and recommendation systems, as part of a broad goal of aligning automated decisions with user preferences. These approaches entail collecting preference information from people, with feedback often provided anonymously. Since preferences are subjective, there is no gold standard to compare against; yet, reliance of high-impact systems on preference learning creates a strong motivation for malicious actors to skew data collected in this fashion to their ends. We investigate the nature and extent of this vulnerability by considering an attacker who can flip a small subset of preference comparisons to either promote or demote a target outcome. We propose two classes of algorithmic approaches for these attacks: a gradient-based framework, and several variants of rank-by-distance methods. Next, we evaluate the efficacy of best attacks in both these classes in successfully achieving malicious goals on datasets from three domains: autonomous control, recommendation system, and textual prompt-response preference learning. We find that the best attacks are often highly successful, achieving in the most extreme case 100\\% success rate with only 0.3\\% of the data poisoned. However, \\emph{which} attack is best can vary significantly across domains. In addition, we observe that the simpler and more scalable rank-by-distance approaches are often competitive with, and on occasion significantly outperform, gradient-based methods. Finally, we show that state-of-the-art defenses against other classes of poisoning attacks exhibit limited efficacy in our setting.", "authors": ["Junlin Wu", "Jiongxiao Wang", "Chaowei Xiao", "Chenguang Wang", "Ning Zhang", "Yevgeniy Vorobeychik"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-02", "url": "https://arxiv.org/abs/2402.01920", "pdf_url": "https://arxiv.org/pdf/2402.01920v2", "arxiv_id": "2402.01920", "doi": "10.1109/SP61157.2025.00094", "citation_count": 15, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "IEEE Symposium on Security and Privacy", "quality_score": 0.301}
{"id": "7111ac37fcf6a411a4afbee08195d79ef4141ee34a1d6a786fa6e35d9ebe13cb", "sources": ["arxiv", "semantic_scholar"], "title": "Dense Reward for Free in Reinforcement Learning from Human Feedback", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has been credited as the key advance that has allowed Large Language Models (LLMs) to effectively follow instructions and produce useful assistance. Classically, this involves generating completions from the LLM in response to a query before using a separate reward model to assign a score to the full completion. As an auto-regressive process, the LLM has to take many \"actions\" (selecting individual tokens) and only receives a single, sparse reward at the end of an episode, a setup that is known to be difficult to optimise in traditional reinforcement learning. In this work we leverage the fact that the reward model contains more information than just its scalar output, in particular, it calculates an attention map over tokens as part of the transformer architecture. We use these attention weights to redistribute the reward along the whole completion, effectively densifying the signal and highlighting the most important tokens, all without incurring extra computational cost or requiring any additional modelling. We demonstrate that, theoretically, this approach is equivalent to potential-based reward shaping, ensuring that the optimal policy remains unchanged. Empirically, we show that it stabilises training, accelerates the rate of learning, and, in practical cases, may lead to better local optima.", "authors": ["Alex J. Chan", "Hao Sun", "Samuel Holt", "Mihaela van der Schaar"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-01", "url": "https://arxiv.org/abs/2402.00782", "pdf_url": "https://arxiv.org/pdf/2402.00782v1", "arxiv_id": "2402.00782", "doi": "10.48550/arXiv.2402.00782", "citation_count": 78, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4771}
{"id": "5a21725555b1d0eb1136ef406cc2727572e9c3e3f2d13e86ea058dfb9dd7a3a6", "sources": ["arxiv", "semantic_scholar"], "title": "Transforming and Combining Rewards for Aligning Large Language Models", "abstract": "A common approach for aligning language models to human preferences is to first learn a reward model from preference data, and then use this reward model to update the language model. We study two closely related problems that arise in this approach. First, any monotone transformation of the reward model preserves preference ranking; is there a choice that is ``better'' than others? Second, we often wish to align language models to multiple properties: how should we combine multiple reward models? Using a probabilistic interpretation of the alignment procedure, we identify a natural choice for transformation for (the common case of) rewards learned from Bradley-Terry preference models. The derived transformation is straightforward: we apply a log-sigmoid function to the centered rewards, a method we term ``LSC-transformation'' (log-sigmoid-centered transformation). This transformation has two important properties. First, it emphasizes improving poorly-performing outputs, rather than outputs that already score well. This mitigates both underfitting (where some prompts are not improved) and reward hacking (where the model learns to exploit misspecification of the reward model). Second, it enables principled aggregation of rewards by linking summation to logical conjunction: the sum of transformed rewards corresponds to the probability that the output is ``good'' in all measured properties, in a sense we make precise. Experiments aligning language models to be both helpful and harmless using RLHF show substantial improvements over the baseline (non-transformed) approach.", "authors": ["Zihao Wang", "Chirag Nagpal", "Jonathan Berant", "Jacob Eisenstein", "Alex D'Amour", "Sanmi Koyejo", "Victor Veitch"], "categories": ["cs.CL", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-02-01", "url": "https://arxiv.org/abs/2402.00742", "pdf_url": "https://arxiv.org/pdf/2402.00742v2", "arxiv_id": "2402.00742", "doi": "10.48550/arXiv.2402.00742", "citation_count": 32, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3796}
{"id": "0148cb0889c3697e44705bb896f85f23dd462398ae4deab61825c49809f27863", "sources": ["arxiv", "semantic_scholar"], "title": "Improving Reinforcement Learning from Human Feedback with Efficient Reward Model Ensemble", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is a widely adopted approach for aligning large language models with human values. However, RLHF relies on a reward model that is trained with a limited amount of human preference data, which could lead to inaccurate predictions. As a result, RLHF may produce outputs that are misaligned with human values. To mitigate this issue, we contribute a reward ensemble method that allows the reward model to make more accurate predictions. As using an ensemble of large language model-based reward models can be computationally and resource-expensive, we explore efficient ensemble methods including linear-layer ensemble and LoRA-based ensemble. Empirically, we run Best-of-$n$ and Proximal Policy Optimization with our ensembled reward models, and verify that our ensemble methods help improve the alignment performance of RLHF outputs.", "authors": ["Shun Zhang", "Zhenfang Chen", "Sunli Chen", "Yikang Shen", "Zhiqing Sun", "Chuang Gan"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-30", "url": "https://arxiv.org/abs/2401.16635", "pdf_url": "https://arxiv.org/pdf/2401.16635v3", "arxiv_id": "2401.16635", "doi": "10.48550/arXiv.2401.16635", "citation_count": 46, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.418}
{"id": "41956d5228f85635fd4e85ca5443fbf667b09e71b60a70ebe8190504bfbbbbfd", "sources": ["arxiv", "semantic_scholar"], "title": "Iterative Data Smoothing: Mitigating Reward Overfitting and Overoptimization in RLHF", "abstract": "Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique that aligns language models closely with human-centric values. The initial phase of RLHF involves learning human values using a reward model from ranking data. It is observed that the performance of the reward model degrades after one epoch of training, and optimizing too much against the learned reward model eventually hinders the true objective. This paper delves into these issues, leveraging the theoretical insights to design improved reward learning algorithm termed 'Iterative Data Smoothing' (IDS). The core idea is that during each training epoch, we not only update the model with the data, but also update the date using the model, replacing hard labels with soft labels. Our empirical findings highlight the superior performance of this approach over the traditional methods.", "authors": ["Banghua Zhu", "Michael I. Jordan", "Jiantao Jiao"], "categories": ["cs.LG", "cs.AI", "cs.CL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2024-01-29", "url": "https://arxiv.org/abs/2401.16335", "pdf_url": "https://arxiv.org/pdf/2401.16335v1", "arxiv_id": "2401.16335", "doi": "10.48550/arXiv.2401.16335", "citation_count": 55, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.437}
{"id": "f6dc10d20cb85eb721acd62a6b473ab660a1e37e58c1a1f9ff8b83239a7b2701", "sources": ["arxiv", "semantic_scholar"], "title": "Reward-Relevance-Filtered Linear Offline Reinforcement Learning", "abstract": "This paper studies offline reinforcement learning with linear function approximation in a setting with decision-theoretic, but not estimation sparsity. The structural restrictions of the data-generating process presume that the transitions factor into a sparse component that affects the reward and could affect additional exogenous dynamics that do not affect the reward. Although the minimally sufficient adjustment set for estimation of full-state transition properties depends on the whole state, the optimal policy and therefore state-action value function depends only on the sparse component: we call this causal/decision-theoretic sparsity. We develop a method for reward-filtering the estimation of the state-action value function to the sparse component by a modification of thresholded lasso in least-squares policy evaluation. We provide theoretical guarantees for our reward-filtered linear fitted-Q-iteration, with sample complexity depending only on the size of the sparse component.", "authors": ["Angela Zhou"], "categories": ["stat.ML", "cs.LG", "math.OC"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2024-01-23", "url": "https://arxiv.org/abs/2401.12934", "pdf_url": "https://arxiv.org/pdf/2401.12934v1", "arxiv_id": "2401.12934", "doi": "10.48550/arXiv.2401.12934", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.1505}
{"id": "acd0d5f550133228b614f5bcd723ea260fc1b2210d58984e4fb315ec1c55c320", "sources": ["arxiv", "semantic_scholar"], "title": "West-of-N: Synthetic Preferences for Self-Improving Reward Models", "abstract": "The success of reinforcement learning from human feedback (RLHF) in language model alignment is strongly dependent on the quality of the underlying reward model. In this paper, we present a novel approach to improve reward model quality by generating synthetic preference data, thereby augmenting the training dataset with on-policy, high-quality preference pairs. Motivated by the promising results of Best-of-N sampling strategies in language model training, we extend their application to reward model training. This results in a self-training strategy to generate preference pairs by selecting the best and worst candidates in a pool of responses to a given query. Empirically, we find that this approach improves the performance of any reward model, with an effect comparable to the addition of a similar quantity of human preference data. This work opens up new avenues of research for improving RLHF for language model alignment, by offering synthetic preference generation as a solution to reward modeling challenges.", "authors": ["Alizée Pace", "Jonathan Mallinson", "Eric Malmi", "Sebastian Krause", "Aliaksei Severyn"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-22", "url": "https://arxiv.org/abs/2401.12086", "pdf_url": "https://arxiv.org/pdf/2401.12086v2", "arxiv_id": "2401.12086", "doi": null, "citation_count": 23, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3451}
{"id": "4a501dbc8958d00f3a8d902664e87078afee4d742c39259e74dac01510f2f8d7", "sources": ["arxiv", "semantic_scholar"], "title": "WARM: On the Benefits of Weight Averaged Reward Models", "abstract": "Aligning large language models (LLMs) with human preferences through reinforcement learning (RLHF) can lead to reward hacking, where LLMs exploit failures in the reward model (RM) to achieve seemingly high rewards without meeting the underlying objectives. We identify two primary challenges when designing RMs to mitigate reward hacking: distribution shifts during the RL process and inconsistencies in human preferences. As a solution, we propose Weight Averaged Reward Models (WARM), first fine-tuning multiple RMs, then averaging them in the weight space. This strategy follows the observation that fine-tuned weights remain linearly mode connected when sharing the same pre-training. By averaging weights, WARM improves efficiency compared to the traditional ensembling of predictions, while improving reliability under distribution shifts and robustness to preference inconsistencies. Our experiments on summarization tasks, using best-of-N and RL methods, shows that WARM improves the overall quality and alignment of LLM predictions; for example, a policy RL fine-tuned with WARM has a 79.4% win rate against a policy RL fine-tuned with a single RM.", "authors": ["Alexandre Ramé", "Nino Vieillard", "Léonard Hussenot", "Robert Dadashi", "Geoffrey Cideron", "Olivier Bachem", "Johan Ferret"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-22", "url": "https://arxiv.org/abs/2401.12187", "pdf_url": "https://arxiv.org/pdf/2401.12187v1", "arxiv_id": "2401.12187", "doi": "10.48550/arXiv.2401.12187", "citation_count": 150, "influential_citation_count": 10, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.5447}
{"id": "ee6d23bf62872cc775b640315e7ce284615663a9f9874dccc7681b0c1fdcb0ef", "sources": ["arxiv", "semantic_scholar"], "title": "Crowd-PrefRL: Preference-Based Reward Learning from Crowds", "abstract": "Preference-based reinforcement learning (RL) provides a framework to train AI agents using human feedback through preferences over pairs of behaviors, enabling agents to learn desired behaviors when it is difficult to specify a numerical reward function. While this paradigm leverages human feedback, it typically treats the feedback as given by a single human user. However, different users may desire multiple AI behaviors and modes of interaction. Meanwhile, incorporating preference feedback from crowds (i.e. ensembles of users) in a robust manner remains a challenge, and the problem of training RL agents using feedback from multiple human users remains understudied. In this work, we introduce a conceptual framework, Crowd-PrefRL, that integrates preference-based RL approaches with techniques from unsupervised crowdsourcing to enable training of autonomous system behaviors from crowdsourced feedback. We show preliminary results suggesting that Crowd-PrefRL can learn reward functions and agent policies from preference feedback provided by crowds of unknown expertise and reliability. We also show that in most cases, agents trained with Crowd-PrefRL outperform agents trained with majority-vote preferences or preferences from any individual user, especially when the spread of user error rates among the crowd is large. Results further suggest that our method can identify the presence of minority viewpoints within the crowd in an unsupervised manner.", "authors": ["David Chhan", "Ellen Novoseller", "Vernon J. Lawhern"], "categories": ["cs.HC", "cs.LG", "cs.SI"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-17", "url": "https://arxiv.org/abs/2401.10941", "pdf_url": "https://arxiv.org/pdf/2401.10941v2", "arxiv_id": "2401.10941", "doi": "10.48550/arXiv.2401.10941", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "802ed4ce619353fae5820e18c5b313770871c8ed9bf622415ae4c78cc5514f08", "sources": ["arxiv", "semantic_scholar"], "title": "Secrets of RLHF in Large Language Models Part II: Reward Modeling", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has become a crucial technology for aligning language models with human values and intentions, enabling models to produce more helpful and harmless responses. Reward models are trained as proxies for human preferences to drive reinforcement learning optimization. While reward models are often considered central to achieving high performance, they face the following challenges in practical applications: (1) Incorrect and ambiguous preference pairs in the dataset may hinder the reward model from accurately capturing human intent. (2) Reward models trained on data from a specific distribution often struggle to generalize to examples outside that distribution and are not suitable for iterative RLHF training. In this report, we attempt to address these two issues. (1) From a data perspective, we propose a method to measure the strength of preferences within the data, based on a voting mechanism of multiple reward models. Experimental results confirm that data with varying preference strengths have different impacts on reward model performance. We introduce a series of novel methods to mitigate the influence of incorrect and ambiguous preferences in the dataset and fully leverage high-quality preference data. (2) From an algorithmic standpoint, we introduce contrastive learning to enhance the ability of reward models to distinguish between chosen and rejected responses, thereby improving model generalization. Furthermore, we employ meta-learning to enable the reward model to maintain the ability to differentiate subtle differences in out-of-distribution samples, and this approach can be utilized for iterative RLHF optimization.", "authors": ["Binghai Wang", "Rui Zheng", "Lu Chen", "Yan Liu", "Shihan Dou", "Caishuang Huang", "Wei Shen", "Senjie Jin", "Enyu Zhou", "Chenyu Shi", "Songyang Gao", "Nuo Xu", "Yuhao Zhou", "Xiaoran Fan", "Zhiheng Xi", "Jun Zhao", "Xiao Wang", "Tao Ji", "Hang Yan", "Lixing Shen", "Zhan Chen", "Tao Gui", "Qi Zhang", "Xipeng Qiu", "Xuanjing Huang", "Zuxuan Wu", "Yu-Gang Jiang"], "categories": ["cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-11", "url": "https://arxiv.org/abs/2401.06080", "pdf_url": "https://arxiv.org/pdf/2401.06080v2", "arxiv_id": "2401.06080", "doi": "10.48550/arXiv.2401.06080", "citation_count": 166, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5557}
{"id": "154971cc1734b6d4ff0847f9c13709396e902e5b3dd975a7f3f871dd59f2df09", "sources": ["arxiv", "semantic_scholar"], "title": "The Distributional Reward Critic Framework for Reinforcement Learning Under Perturbed Rewards", "abstract": "The reward signal plays a central role in defining the desired behaviors of agents in reinforcement learning (RL). Rewards collected from realistic environments could be perturbed, corrupted, or noisy due to an adversary, sensor error, or because they come from subjective human feedback. Thus, it is important to construct agents that can learn under such rewards. Existing methodologies for this problem make strong assumptions, including that the perturbation is known in advance, clean rewards are accessible, or that the perturbation preserves the optimal policy. We study a new, more general, class of unknown perturbations, and introduce a distributional reward critic framework for estimating reward distributions and perturbations during training. Our proposed methods are compatible with any RL algorithm. Despite their increased generality, we show that they achieve comparable or better rewards than existing methods in a variety of environments, including those with clean rewards. Under the challenging and generalized perturbations we study, we win/tie the highest return in 44/48 tested settings (compared to 11/48 for the best baseline). Our results broaden and deepen our ability to perform RL in reward-perturbed environments.", "authors": ["Xi Chen", "Zhihui Zhu", "Andrew Perrault"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2024-01-11", "url": "https://arxiv.org/abs/2401.05710", "pdf_url": "https://arxiv.org/pdf/2401.05710v3", "arxiv_id": "2401.05710", "doi": "10.1609/aaai.v39i15.33745", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.0}
{"id": "09e95ff3a160797f188310547ff612a458db371b2390687922e28e3f42274ea0", "sources": ["arxiv", "semantic_scholar"], "title": "Uncertainty-Penalized Reinforcement Learning from Human Feedback with Diverse Reward LoRA Ensembles", "abstract": "Reinforcement learning from human feedback (RLHF) emerges as a promising paradigm for aligning large language models (LLMs). However, a notable challenge in RLHF is overoptimization, where beyond a certain threshold, the pursuit of higher rewards leads to a decline in human preferences. In this paper, we observe the weakness of KL regularization which is commonly employed in existing RLHF methods to address overoptimization. To mitigate this limitation, we scrutinize the RLHF objective in the offline dataset and propose uncertainty-penalized RLHF (UP-RLHF), which incorporates uncertainty regularization during RL-finetuning. To enhance the uncertainty quantification abilities for reward models, we first propose a diverse low-rank adaptation (LoRA) ensemble by maximizing the nuclear norm of LoRA matrix concatenations. Then we optimize policy models utilizing penalized rewards, determined by both rewards and uncertainties provided by the diverse reward LoRA ensembles. Our experimental results, based on two real human preference datasets, showcase the effectiveness of diverse reward LoRA ensembles in quantifying reward uncertainty. Additionally, uncertainty regularization in UP-RLHF proves to be pivotal in mitigating overoptimization, thereby contributing to the overall performance.", "authors": ["Yuanzhao Zhai", "Han Zhang", "Yu Lei", "Yue Yu", "Kele Xu", "Dawei Feng", "Bo Ding", "Huaimin Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-30", "url": "https://arxiv.org/abs/2401.00243", "pdf_url": "https://arxiv.org/pdf/2401.00243v1", "arxiv_id": "2401.00243", "doi": "10.48550/arXiv.2401.00243", "citation_count": 49, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4247}
{"id": "cb3f62c582cd2c86ba039e42d965fead943993acddb13cc7f9d9d4ea50bbe054", "sources": ["arxiv", "semantic_scholar"], "title": "Preference as Reward, Maximum Preference Optimization with Importance Sampling", "abstract": "Preference learning is a key technology for aligning language models with human values. Reinforcement Learning from Human Feedback (RLHF) is a model-based algorithm to optimize preference learning, which first fits a reward model for preference scores and then optimizes the generating policy with an on-policy PPO algorithm to maximize the reward. The processing of RLHF is complex, time-consuming, and unstable. The Direct Preference Optimization (DPO) algorithm uses an off-policy algorithm to directly optimize the generating policy and eliminates the need for a reward model. DPO is more data-efficient and stable. However, DPO has a drawback of overfitting to the preference data and ignoring the KL-regularization term when the preference is deterministic. Identity mapping Preference Optimization(IPO) uses a root-finding MSE loss to incorporate KL-regularization. However, both DPO and IPO fail to properly address the KL-regularization term because the support of the preference distribution is not equal to the reference distribution. In this paper, we propose a simple and intuitive off-policy preference optimization algorithm from an importance sampling view, which we call Maximum Preference Optimization (MPO). MPO incorporates the off-policy KL-regularization term, making regularization truly effective. MPO achieves the best of both worlds by combining the objectives of RLHF and IPO while being an off-policy algorithm. Furthermore, MPO eliminates the need for a reward model and reference policy, simplifying the learning process and reducing memory usage.", "authors": ["Zaifan Jiang", "Xing Huang", "Chao Wei"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-27", "url": "https://arxiv.org/abs/2312.16430", "pdf_url": "https://arxiv.org/pdf/2312.16430v5", "arxiv_id": "2312.16430", "doi": "10.48550/arXiv.2312.16430", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "a95eaa3ecbb33c49964c4f81fd7a635fdf45088d7733f7465ab64aebcf776877", "sources": ["arxiv", "semantic_scholar"], "title": "REBEL: Reward Regularization-Based Approach for Robotic Reinforcement Learning from Human Feedback", "abstract": "The effectiveness of reinforcement learning (RL) agents in continuous control robotics tasks is mainly dependent on the design of the underlying reward function, which is highly prone to reward hacking. A misalignment between the reward function and underlying human preferences (values, social norms) can lead to catastrophic outcomes in the real world especially in the context of robotics for critical decision making. Recent methods aim to mitigate misalignment by learning reward functions from human preferences and subsequently performing policy optimization. However, these methods inadvertently introduce a distribution shift during reward learning due to ignoring the dependence of agent-generated trajectories on the reward learning objective, ultimately resulting in sub-optimal alignment. Hence, in this work, we address this challenge by advocating for the adoption of regularized reward functions that more accurately mirror the intended behaviors of the agent. We propose a novel concept of reward regularization within the robotic RLHF (RL from Human Feedback) framework, which we refer to as \\emph{agent preferences}. Our approach uniquely incorporates not just human feedback in the form of preferences but also considers the preferences of the RL agent itself during the reward function learning process. This dual consideration significantly mitigates the issue of distribution shift in RLHF with a computationally tractable algorithm. We provide a theoretical justification for the proposed algorithm by formulating the robotic RLHF problem as a bilevel optimization problem and developing a computationally tractable version of the same. We demonstrate the efficiency of our algorithm {\\ours} in several continuous control benchmarks in DeepMind Control Suite \\cite{tassa2018deepmind}.", "authors": ["Souradip Chakraborty", "Anukriti Singh", "Amisha Bhaskar", "Pratap Tokekar", "Dinesh Manocha", "Amrit Singh Bedi"], "categories": ["cs.RO", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-22", "url": "https://arxiv.org/abs/2312.14436", "pdf_url": "https://arxiv.org/pdf/2312.14436v3", "arxiv_id": "2312.14436", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "53e7468f7e566f6a1c0758f6abdfb7aeee57c6a0cbf21c056cdb4411175a1d55", "sources": ["arxiv", "semantic_scholar"], "title": "Diffusion Reward: Learning Rewards via Conditional Video Diffusion", "abstract": "Learning rewards from expert videos offers an affordable and effective solution to specify the intended behaviors for reinforcement learning (RL) tasks. In this work, we propose Diffusion Reward, a novel framework that learns rewards from expert videos via conditional video diffusion models for solving complex visual RL problems. Our key insight is that lower generative diversity is exhibited when conditioning diffusion on expert trajectories. Diffusion Reward is accordingly formalized by the negative of conditional entropy that encourages productive exploration of expert behaviors. We show the efficacy of our method over robotic manipulation tasks in both simulation platforms and the real world with visual input. Moreover, Diffusion Reward can even solve unseen tasks successfully and effectively, largely surpassing baseline methods. Project page and code: https://diffusion-reward.github.io.", "authors": ["Tao Huang", "Guangqi Jiang", "Yanjie Ze", "Huazhe Xu"], "categories": ["cs.LG", "cs.CV", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-21", "url": "https://arxiv.org/abs/2312.14134", "pdf_url": "https://arxiv.org/pdf/2312.14134v3", "arxiv_id": "2312.14134", "doi": "10.48550/arXiv.2312.14134", "citation_count": 57, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "European Conference on Computer Vision", "quality_score": 0.4515}
{"id": "eeb9badca48d86134f5592ce95011db6752043b186eb64332ead5173456711b8", "sources": ["arxiv", "semantic_scholar"], "title": "Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint", "abstract": "This paper studies the alignment process of generative models with Reinforcement Learning from Human Feedback (RLHF). We first identify the primary challenges of existing popular methods like offline PPO and offline DPO as lacking in strategical exploration of the environment. Then, to understand the mathematical principle of RLHF, we consider a standard mathematical formulation, the reverse-KL regularized contextual bandit for RLHF. Despite its widespread practical application, a rigorous theoretical analysis of this formulation remains open. We investigate its behavior in three distinct settings -- offline, online, and hybrid -- and propose efficient algorithms with finite-sample theoretical guarantees. Moving towards practical applications, our framework, with a robust approximation of the information-theoretical policy improvement oracle, naturally gives rise to several novel RLHF algorithms. This includes an iterative version of the Direct Preference Optimization (DPO) algorithm for online settings, and a multi-step rejection sampling strategy for offline scenarios. Our empirical evaluations on real-world alignment experiment of large language model demonstrate that these proposed methods significantly surpass existing strong baselines, such as DPO and Rejection Sampling Optimization (RSO), showcasing the connections between solid theoretical foundations and their potent practical implementations.", "authors": ["Wei Xiong", "Hanze Dong", "Chenlu Ye", "Ziqi Wang", "Han Zhong", "Heng Ji", "Nan Jiang", "Tong Zhang"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-12-18", "url": "https://arxiv.org/abs/2312.11456", "pdf_url": "https://arxiv.org/pdf/2312.11456v4", "arxiv_id": "2312.11456", "doi": null, "citation_count": 372, "influential_citation_count": 32, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.7593}
{"id": "8b2ca1d23b255aa4b159e567ff6b67cdc0107d8e62cd781bb71356cdc00dae3c", "sources": ["arxiv", "semantic_scholar"], "title": "Helping or Herding? Reward Model Ensembles Mitigate but do not Eliminate Reward Hacking", "abstract": "Reward models play a key role in aligning language model applications towards human preferences. However, this setup creates an incentive for the language model to exploit errors in the reward model to achieve high estimated reward, a phenomenon often termed \\emph{reward hacking}. A natural mitigation is to train an ensemble of reward models, aggregating over model outputs to obtain a more robust reward estimate. We explore the application of reward ensembles to alignment at both training time (through reinforcement learning) and inference time (through reranking). First, we show that reward models are \\emph{underspecified}: reward models that perform similarly in-distribution can yield very different rewards when used in alignment, due to distribution shift. Second, underspecification results in overoptimization, where alignment to one reward model does not improve reward as measured by another reward model trained on the same data. Third, overoptimization is mitigated by the use of reward ensembles, and ensembles that vary by their \\emph{pretraining} seeds lead to better generalization than ensembles that differ only by their \\emph{fine-tuning} seeds, with both outperforming individual reward models. However, even pretrain reward ensembles do not eliminate reward hacking: we show several qualitative reward hacking phenomena that are not mitigated by ensembling because all reward models in the ensemble exhibit similar error patterns.", "authors": ["Jacob Eisenstein", "Chirag Nagpal", "Alekh Agarwal", "Ahmad Beirami", "Alex D'Amour", "DJ Dvijotham", "Adam Fisch", "Katherine Heller", "Stephen Pfohl", "Deepak Ramachandran", "Peter Shaw", "Jonathan Berant"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-12-14", "url": "https://arxiv.org/abs/2312.09244", "pdf_url": "https://arxiv.org/pdf/2312.09244v3", "arxiv_id": "2312.09244", "doi": "10.48550/arXiv.2312.09244", "citation_count": 180, "influential_citation_count": 14, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.588}
{"id": "17c9d55f0849938e92566832ce5186d7e9233196917da61ea7a579b42e822f79", "sources": ["arxiv", "semantic_scholar"], "title": "Global Rewards in Multi-Agent Deep Reinforcement Learning for Autonomous Mobility on Demand Systems", "abstract": "We study vehicle dispatching in autonomous mobility on demand (AMoD) systems, where a central operator assigns vehicles to customer requests or rejects these with the aim of maximizing its total profit. Recent approaches use multi-agent deep reinforcement learning (MADRL) to realize scalable yet performant algorithms, but train agents based on local rewards, which distorts the reward signal with respect to the system-wide profit, leading to lower performance. We therefore propose a novel global-rewards-based MADRL algorithm for vehicle dispatching in AMoD systems, which resolves so far existing goal conflicts between the trained agents and the operator by assigning rewards to agents leveraging a counterfactual baseline. Our algorithm shows statistically significant improvements across various settings on real-world data compared to state-of-the-art MADRL algorithms with local rewards. We further provide a structural analysis which shows that the utilization of global rewards can improve implicit vehicle balancing and demand forecasting abilities. Our code is available at https://github.com/tumBAIS/GR-MADRL-AMoD.", "authors": ["Heiko Hoppe", "Tobias Enders", "Quentin Cappart", "Maximilian Schiffer"], "categories": ["cs.LG", "cs.MA", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-12-14", "url": "https://arxiv.org/abs/2312.08884", "pdf_url": "https://arxiv.org/pdf/2312.08884v2", "arxiv_id": "2312.08884", "doi": "10.48550/arXiv.2312.08884", "citation_count": 13, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/tumBAIS/GR-MADRL-AMoD", "venue": "Conference on Learning for Dynamics & Control", "quality_score": 0.2865}
{"id": "74281349b9f211bee9d0dff61fbdc5a86e71be3a1810951710538948cbcab8a8", "sources": ["arxiv", "semantic_scholar"], "title": "Distributional Preference Learning: Understanding and Accounting for Hidden Context in RLHF", "abstract": "In practice, preference learning from human feedback depends on incomplete data with hidden context. Hidden context refers to data that affects the feedback received, but which is not represented in the data used to train a preference model. This captures common issues of data collection, such as having human annotators with varied preferences, cognitive processes that result in seemingly irrational behavior, and combining data labeled according to different criteria. We prove that standard applications of preference learning, including reinforcement learning from human feedback (RLHF), implicitly aggregate over hidden contexts according to a well-known voting rule called Borda count. We show this can produce counter-intuitive results that are very different from other methods which implicitly aggregate via expected utility. Furthermore, our analysis formalizes the way that preference learning from users with diverse values tacitly implements a social choice function. A key implication of this result is that annotators have an incentive to misreport their preferences in order to influence the learned model, leading to vulnerabilities in the deployment of RLHF. As a step towards mitigating these problems, we introduce a class of methods called distributional preference learning (DPL). DPL methods estimate a distribution of possible score values for each alternative in order to better account for hidden context. Experimental results indicate that applying DPL to RLHF for LLM chatbots identifies hidden context in the data and significantly reduces subsequent jailbreak vulnerability. Our code and data are available at https://github.com/cassidylaidlaw/hidden-context", "authors": ["Anand Siththaranjan", "Cassidy Laidlaw", "Dylan Hadfield-Menell"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-12-13", "url": "https://arxiv.org/abs/2312.08358", "pdf_url": "https://arxiv.org/pdf/2312.08358v2", "arxiv_id": "2312.08358", "doi": "10.48550/arXiv.2312.08358", "citation_count": 123, "influential_citation_count": 24, "has_code": true, "code_url": "https://github.com/cassidylaidlaw/hidden-context", "venue": "International Conference on Learning Representations", "quality_score": 0.699}
{"id": "2172745df8a4f8628adcf68e1fd7427611ff87f77695cae5618f490ddbf683e2", "sources": ["arxiv", "semantic_scholar"], "title": "When is Off-Policy Evaluation (Reward Modeling) Useful in Contextual Bandits? A Data-Centric Perspective", "abstract": "Evaluating the value of a hypothetical target policy with only a logged dataset is important but challenging. On the one hand, it brings opportunities for safe policy improvement under high-stakes scenarios like clinical guidelines. On the other hand, such opportunities raise a need for precise off-policy evaluation (OPE). While previous work on OPE focused on improving the algorithm in value estimation, in this work, we emphasize the importance of the offline dataset, hence putting forward a data-centric framework for evaluating OPE problems. We propose DataCOPE, a data-centric framework for evaluating OPE, that answers the questions of whether and to what extent we can evaluate a target policy given a dataset. DataCOPE (1) forecasts the overall performance of OPE algorithms without access to the environment, which is especially useful before real-world deployment where evaluating OPE is impossible; (2) identifies the sub-group in the dataset where OPE can be inaccurate; (3) permits evaluations of datasets or data-collection strategies for OPE problems. Our empirical analysis of DataCOPE in the logged contextual bandit settings using healthcare datasets confirms its ability to evaluate both machine-learning and human expert policies like clinical guidelines. Finally, we apply DataCOPE to the task of reward modeling in Large Language Model alignment to demonstrate its scalability in real-world applications.", "authors": ["Hao Sun", "Alex J. Chan", "Nabeel Seedat", "Alihan Hüyük", "Mihaela van der Schaar"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-23", "url": "https://arxiv.org/abs/2311.14110", "pdf_url": "https://arxiv.org/pdf/2311.14110v2", "arxiv_id": "2311.14110", "doi": null, "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1505}
{"id": "325b4d60cc95a021ea32fc1b4dd805ef70d1af09191c46e14e0539959a1169a8", "sources": ["arxiv", "semantic_scholar"], "title": "Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model", "abstract": "Using reinforcement learning with human feedback (RLHF) has shown significant promise in fine-tuning diffusion models. Previous methods start by training a reward model that aligns with human preferences, then leverage RL techniques to fine-tune the underlying models. However, crafting an efficient reward model demands extensive datasets, optimal architecture, and manual hyperparameter tuning, making the process both time and cost-intensive. The direct preference optimization (DPO) method, effective in fine-tuning large language models, eliminates the necessity for a reward model. However, the extensive GPU memory requirement of the diffusion model's denoising process hinders the direct application of the DPO method. To address this issue, we introduce the Direct Preference for Denoising Diffusion Policy Optimization (D3PO) method to directly fine-tune diffusion models. The theoretical analysis demonstrates that although D3PO omits training a reward model, it effectively functions as the optimal reward model trained using human feedback data to guide the learning process. This approach requires no training of a reward model, proving to be more direct, cost-effective, and minimizing computational overhead. In experiments, our method uses the relative scale of objectives as a proxy for human preference, delivering comparable results to methods using ground-truth rewards. Moreover, D3PO demonstrates the ability to reduce image distortion rates and generate safer images, overcoming challenges lacking robust reward models. Our code is publicly available at https://github.com/yk7333/D3PO.", "authors": ["Kai Yang", "Jian Tao", "Jiafei Lyu", "Chunjiang Ge", "Jiaxin Chen", "Qimai Li", "Weihan Shen", "Xiaolong Zhu", "Xiu Li"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-22", "url": "https://arxiv.org/abs/2311.13231", "pdf_url": "https://arxiv.org/pdf/2311.13231v3", "arxiv_id": "2311.13231", "doi": "10.1109/CVPR52733.2024.00854", "citation_count": 267, "influential_citation_count": 28, "has_code": true, "code_url": "https://github.com/yk7333/D3PO", "venue": "Computer Vision and Pattern Recognition", "quality_score": 0.7312}
{"id": "b62f57af3a138b85aa35d8e34ca9178f46eea3f412d0677a306d10eb6784cacb", "sources": ["arxiv", "semantic_scholar"], "title": "Optimal Cooperative Multiplayer Learning Bandits with Noisy Rewards and No Communication", "abstract": "We consider a cooperative multiplayer bandit learning problem where the players are only allowed to agree on a strategy beforehand, but cannot communicate during the learning process. In this problem, each player simultaneously selects an action. Based on the actions selected by all players, the team of players receives a reward. The actions of all the players are commonly observed. However, each player receives a noisy version of the reward which cannot be shared with other players. Since players receive potentially different rewards, there is an asymmetry in the information used to select their actions. In this paper, we provide an algorithm based on upper and lower confidence bounds that the players can use to select their optimal actions despite the asymmetry in the reward information. We show that this algorithm can achieve logarithmic $O(\\frac{\\log T}{Δ_{\\bm{a}}})$ (gap-dependent) regret as well as $O(\\sqrt{T\\log T})$ (gap-independent) regret. This is asymptotically optimal in $T$. We also show that it performs empirically better than the current state of the art algorithm for this environment.", "authors": ["William Chang", "Yuanhao Lu"], "categories": ["cs.LG", "cs.MA", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-11-10", "url": "https://arxiv.org/abs/2311.06210", "pdf_url": "https://arxiv.org/pdf/2311.06210v1", "arxiv_id": "2311.06210", "doi": "10.1109/CDC56724.2024.10885907", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Conference on Decision and Control", "quality_score": 0.1505}
{"id": "3c09d16a1bddf023594d771316221a41d3240383d4b83d4d7a8790ff0447014e", "sources": ["arxiv", "semantic_scholar"], "title": "DreamSmooth: Improving Model-based Reinforcement Learning via Reward Smoothing", "abstract": "Model-based reinforcement learning (MBRL) has gained much attention for its ability to learn complex behaviors in a sample-efficient way: planning actions by generating imaginary trajectories with predicted rewards. Despite its success, we found that surprisingly, reward prediction is often a bottleneck of MBRL, especially for sparse rewards that are challenging (or even ambiguous) to predict. Motivated by the intuition that humans can learn from rough reward estimates, we propose a simple yet effective reward smoothing approach, DreamSmooth, which learns to predict a temporally-smoothed reward, instead of the exact reward at the given timestep. We empirically show that DreamSmooth achieves state-of-the-art performance on long-horizon sparse-reward tasks both in sample efficiency and final performance without losing performance on common benchmarks, such as Deepmind Control Suite and Atari benchmarks.", "authors": ["Vint Lee", "Pieter Abbeel", "Youngwoon Lee"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2023-11-02", "url": "https://arxiv.org/abs/2311.01450", "pdf_url": "https://arxiv.org/pdf/2311.01450v2", "arxiv_id": "2311.01450", "doi": "10.48550/arXiv.2311.01450", "citation_count": 10, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2603}
{"id": "20b3d09ca4a0672e41d749be1116b73534005c79ff560e853b12e39dd0cf4978", "sources": ["arxiv", "semantic_scholar"], "title": "Differentially Private Reward Estimation with Preference Feedback", "abstract": "Learning from preference-based feedback has recently gained considerable traction as a promising approach to align generative models with human interests. Instead of relying on numerical rewards, the generative models are trained using reinforcement learning with human feedback (RLHF). These approaches first solicit feedback from human labelers typically in the form of pairwise comparisons between two possible actions, then estimate a reward model using these comparisons, and finally employ a policy based on the estimated reward model. An adversarial attack in any step of the above pipeline might reveal private and sensitive information of human labelers. In this work, we adopt the notion of label differential privacy (DP) and focus on the problem of reward estimation from preference-based feedback while protecting privacy of each individual labelers. Specifically, we consider the parametric Bradley-Terry-Luce (BTL) model for such pairwise comparison feedback involving a latent reward parameter $θ^* \\in \\mathbb{R}^d$. Within a standard minimax estimation framework, we provide tight upper and lower bounds on the error in estimating $θ^*$ under both local and central models of DP. We show, for a given privacy budget $ε$ and number of samples $n$, that the additional cost to ensure label-DP under local model is $Θ\\big(\\frac{1}{ e^ε-1}\\sqrt{\\frac{d}{n}}\\big)$, while it is $Θ\\big(\\frac{\\text{poly}(d)}{εn} \\big)$ under the weaker central model. We perform simulations on synthetic data that corroborate these theoretical results.", "authors": ["Sayak Ray Chowdhury", "Xingyu Zhou", "Nagarajan Natarajan"], "categories": ["cs.LG", "cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-30", "url": "https://arxiv.org/abs/2310.19733", "pdf_url": "https://arxiv.org/pdf/2310.19733v1", "arxiv_id": "2310.19733", "doi": "10.48550/arXiv.2310.19733", "citation_count": 12, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.301}
{"id": "972ad0d6246d31b2206dacb84165386bc1b969da5656724a1872cedfb7b6e081", "sources": ["arxiv", "semantic_scholar"], "title": "Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning", "abstract": "Reinforcement learning (RL) requires either manually specifying a reward function, which is often infeasible, or learning a reward model from a large amount of human feedback, which is often very expensive. We study a more sample-efficient alternative: using pretrained vision-language models (VLMs) as zero-shot reward models (RMs) to specify tasks via natural language. We propose a natural and general approach to using VLMs as reward models, which we call VLM-RMs. We use VLM-RMs based on CLIP to train a MuJoCo humanoid to learn complex tasks without a manually specified reward function, such as kneeling, doing the splits, and sitting in a lotus position. For each of these tasks, we only provide a single sentence text prompt describing the desired task with minimal prompt engineering. We provide videos of the trained agents at: https://sites.google.com/view/vlm-rm. We can improve performance by providing a second \"baseline\" prompt and projecting out parts of the CLIP embedding space irrelevant to distinguish between goal and baseline. Further, we find a strong scaling effect for VLM-RMs: larger VLMs trained with more compute and data are better reward models. The failure modes of VLM-RMs we encountered are all related to known capability limitations of current VLMs, such as limited spatial reasoning ability or visually unrealistic environments that are far off-distribution for the VLM. We find that VLM-RMs are remarkably robust as long as the VLM is large enough. This suggests that future VLMs will become more and more useful reward models for a wide range of RL applications.", "authors": ["Juan Rocamonde", "Victoriano Montesinos", "Elvis Nava", "Ethan Perez", "David Lindner"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-19", "url": "https://arxiv.org/abs/2310.12921", "pdf_url": "https://arxiv.org/pdf/2310.12921v2", "arxiv_id": "2310.12921", "doi": "10.48550/arXiv.2310.12921", "citation_count": 166, "influential_citation_count": 19, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.6505}
{"id": "490965d397dd2d3b539e39bf1e1b2eb0bb5c5351fe0bc0f29e5225ddd78d8e87", "sources": ["arxiv", "semantic_scholar"], "title": "SALMON: Self-Alignment with Instructable Reward Models", "abstract": "Supervised Fine-Tuning (SFT) on response demonstrations combined with Reinforcement Learning from Human Feedback (RLHF) constitutes a powerful paradigm for aligning LLM-based AI agents. However, a significant limitation of such an approach is its dependency on high-quality human annotations, making its application to intricate tasks challenging due to difficulties in obtaining consistent response demonstrations and in-distribution response preferences. This paper presents a novel approach, namely SALMON, to align base language models with minimal human supervision, using only a small set of human-defined principles, yet achieving superior performance. Central to our approach is an instructable reward model. Trained on synthetic preference data, this model can generate reward scores based on arbitrary human-defined principles. By merely adjusting these principles during the RL training phase, we gain full control over the preferences with the instructable reward model, subsequently influencing the behavior of the RL-trained policy models, and reducing the reliance on the collection of online human preferences. Applying our method to the LLaMA-2-70b base language model, we developed an AI assistant named Dromedary-2. With only 6 exemplars for in-context learning and 31 human-defined principles, Dromedary-2 significantly surpasses the performance of several state-of-the-art AI systems, including LLaMA-2-Chat-70b, on various benchmark datasets. We have open-sourced the code and model weights to encourage further research into aligning LLM-based AI agents with enhanced supervision efficiency, improved controllability, and scalable oversight.", "authors": ["Zhiqing Sun", "Yikang Shen", "Hongxin Zhang", "Qinhong Zhou", "Zhenfang Chen", "David Cox", "Yiming Yang", "Chuang Gan"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-09", "url": "https://arxiv.org/abs/2310.05910", "pdf_url": "https://arxiv.org/pdf/2310.05910v2", "arxiv_id": "2310.05910", "doi": null, "citation_count": 62, "influential_citation_count": 7, "has_code": true, "code_url": "https://github.com/IBM/SALMON", "venue": "International Conference on Learning Representations", "quality_score": 0.4515}
{"id": "11f98bd59f14b1f66c6c13c3806c26bfd641ad8574b0b6b86350f4f0ea770632", "sources": ["arxiv", "semantic_scholar"], "title": "Loose lips sink ships: Mitigating Length Bias in Reinforcement Learning from Human Feedback", "abstract": "Reinforcement learning from human feedback serves as a crucial bridge, aligning large language models with human and societal values. This alignment requires a vast corpus of human feedback to learn a reward model, which is subsequently used to finetune language models. However, we have identified that the reward model often finds shortcuts to bypass its intended objectives, misleadingly assuming that humans prefer longer responses. The emergence of length bias often induces the model to favor longer outputs, yet it doesn't equate to an increase in helpful information within these outputs. In this paper, we propose an innovative solution, applying the Product-of-Experts (PoE) technique to separate reward modeling from the influence of sequence length. In our framework, the main expert concentrates on understanding human intents, while the biased expert targets the identification and capture of length bias. To further enhance the learning of bias, we introduce perturbations into the bias-focused expert, disrupting the flow of semantic information. Experimental results validate the effectiveness of our approach, indicating that language model performance is improved, irrespective of sequence length.", "authors": ["Wei Shen", "Rui Zheng", "Wenyu Zhan", "Jun Zhao", "Shihan Dou", "Tao Gui", "Qi Zhang", "Xuanjing Huang"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-08", "url": "https://arxiv.org/abs/2310.05199", "pdf_url": "https://arxiv.org/pdf/2310.05199v5", "arxiv_id": "2310.05199", "doi": "10.48550/arXiv.2310.05199", "citation_count": 89, "influential_citation_count": 13, "has_code": false, "code_url": null, "venue": "Conference on Empirical Methods in Natural Language Processing", "quality_score": 0.5731}
{"id": "4f3d78bf7c8af30ba789eece6d011030c68437889dd73471a7f38a56db977e2d", "sources": ["arxiv", "semantic_scholar"], "title": "Confronting Reward Model Overoptimization with Constrained RLHF", "abstract": "Large language models are typically aligned with human preferences by optimizing $\\textit{reward models}$ (RMs) fitted to human feedback. However, human preferences are multi-faceted, and it is increasingly common to derive reward from a composition of simpler reward models which each capture a different aspect of language quality. This itself presents a challenge, as it is difficult to appropriately weight these component RMs when combining them. Compounding this difficulty, because any RM is only a proxy for human evaluation, this process is vulnerable to $\\textit{overoptimization}$, wherein past a certain point, accumulating higher reward is associated with worse human ratings. In this paper, we perform, to our knowledge, the first study on overoptimization in composite RMs, showing that correlation between component RMs has a significant effect on the locations of these points. We then introduce an approach to solve this issue using constrained reinforcement learning as a means of preventing the agent from exceeding each RM's threshold of usefulness. Our method addresses the problem of weighting component RMs by learning dynamic weights, naturally expressed by Lagrange multipliers. As a result, each RM stays within the range at which it is an effective proxy, improving evaluation performance. Finally, we introduce an adaptive method using gradient-free optimization to identify and optimize towards these points during a single run.", "authors": ["Ted Moskovitz", "Aaditya K. Singh", "DJ Strouse", "Tuomas Sandholm", "Ruslan Salakhutdinov", "Anca D. Dragan", "Stephen McAleer"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-06", "url": "https://arxiv.org/abs/2310.04373", "pdf_url": "https://arxiv.org/pdf/2310.04373v2", "arxiv_id": "2310.04373", "doi": "10.48550/arXiv.2310.04373", "citation_count": 101, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.5022}
{"id": "eca0856dd29389e446bf337c58bda2b68ebe9ac172aaf3e4b4889103d2214e0e", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Model Ensembles Help Mitigate Overoptimization", "abstract": "Reinforcement learning from human feedback (RLHF) is a standard approach for fine-tuning large language models to follow instructions. As part of this process, learned reward models are used to approximately model human preferences. However, as imperfect representations of the \"true\" reward, these learned reward models are susceptible to overoptimization. Gao et al. (2023) studied this phenomenon in a synthetic human feedback setup with a significantly larger \"gold\" reward model acting as the true reward (instead of humans) and showed that overoptimization remains a persistent problem regardless of the size of the proxy reward model and training data used. Using a similar setup, we conduct a systematic study to evaluate the efficacy of using ensemble-based conservative optimization objectives, specifically worst-case optimization (WCO) and uncertainty-weighted optimization (UWO), for mitigating reward model overoptimization when using two optimization methods: (a) best-of-n sampling (BoN) (b) proximal policy optimization (PPO). We additionally extend the setup of Gao et al. (2023) to include 25% label noise to better mirror real-world conditions. Both with and without label noise, we find that conservative optimization practically eliminates overoptimization and improves performance by up to 70% for BoN sampling. For PPO, ensemble-based conservative optimization always reduces overoptimization and outperforms single reward model optimization. Moreover, combining it with a small KL penalty successfully prevents overoptimization at no performance cost. Overall, our results demonstrate that ensemble-based conservative optimization can effectively counter overoptimization.", "authors": ["Thomas Coste", "Usman Anwar", "Robert Kirk", "David Krueger"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-04", "url": "https://arxiv.org/abs/2310.02743", "pdf_url": "https://arxiv.org/pdf/2310.02743v2", "arxiv_id": "2310.02743", "doi": "10.48550/arXiv.2310.02743", "citation_count": 228, "influential_citation_count": 30, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.7457}
{"id": "b71c3de2b01a628f56455de2952fb37f5a77d1d50bd4b5f8854e4e5ebe12c088", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Optimal Advantage from Preferences and Mistaking it for Reward", "abstract": "We consider algorithms for learning reward functions from human preferences over pairs of trajectory segments, as used in reinforcement learning from human feedback (RLHF). Most recent work assumes that human preferences are generated based only upon the reward accrued within those segments, or their partial return. Recent work casts doubt on the validity of this assumption, proposing an alternative preference model based upon regret. We investigate the consequences of assuming preferences are based upon partial return when they actually arise from regret. We argue that the learned function is an approximation of the optimal advantage function, $\\hat{A^*_r}$, not a reward function. We find that if a specific pitfall is addressed, this incorrect assumption is not particularly harmful, resulting in a highly shaped reward function. Nonetheless, this incorrect usage of $\\hat{A^*_r}$ is less desirable than the appropriate and simpler approach of greedy maximization of $\\hat{A^*_r}$. From the perspective of the regret preference model, we also provide a clearer interpretation of fine tuning contemporary large language models with RLHF. This paper overall provides insight regarding why learning under the partial return preference model tends to work so well in practice, despite it conforming poorly to how humans give preferences.", "authors": ["W. Bradley Knox", "Stephane Hatgis-Kessell", "Sigurdur Orn Adalgeirsson", "Serena Booth", "Anca Dragan", "Peter Stone", "Scott Niekum"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-03", "url": "https://arxiv.org/abs/2310.02456", "pdf_url": "https://arxiv.org/pdf/2310.02456v1", "arxiv_id": "2310.02456", "doi": "10.48550/arXiv.2310.02456", "citation_count": 20, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.3306}
{"id": "9b769764e5a3242eecaea9c9630e0a509f6e8bbe4efcdef5efc820b03ab7d048", "sources": ["arxiv", "semantic_scholar"], "title": "Tool-Augmented Reward Modeling", "abstract": "Reward modeling (a.k.a., preference modeling) is instrumental for aligning large language models with human preferences, particularly within the context of reinforcement learning from human feedback (RLHF). While conventional reward models (RMs) have exhibited remarkable scalability, they oft struggle with fundamental functionality such as arithmetic computation, code execution, and factual lookup. In this paper, we propose a tool-augmented preference modeling approach, named Themis, to address these limitations by empowering RMs with access to external environments, including calculators and search engines. This approach not only fosters synergy between tool utilization and reward grading but also enhances interpretive capacity and scoring reliability. Our study delves into the integration of external tools into RMs, enabling them to interact with diverse external sources and construct task-specific tool engagement and reasoning traces in an autoregressive manner. We validate our approach across a wide range of domains, incorporating seven distinct external tools. Our experimental results demonstrate a noteworthy overall improvement of 17.7% across eight tasks in preference ranking. Furthermore, our approach outperforms Gopher 280B by 7.3% on TruthfulQA task in zero-shot evaluation. In human evaluations, RLHF trained with Themis attains an average win rate of 32% when compared to baselines across four distinct tasks. Additionally, we provide a comprehensive collection of tool-related RM datasets, incorporating data from seven distinct tool APIs, totaling 15,000 instances. We have made the code, data, and model checkpoints publicly available to facilitate and inspire further research advancements\\footnote{\\url{https://github.com/ernie-research/Tool-Augmented-Reward-Model}}.", "authors": ["Lei Li", "Yekun Chai", "Shuohuan Wang", "Yu Sun", "Hao Tian", "Ningyu Zhang", "Hua Wu"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-10-02", "url": "https://arxiv.org/abs/2310.01045", "pdf_url": "https://arxiv.org/pdf/2310.01045v2", "arxiv_id": "2310.01045", "doi": "10.48550/arXiv.2310.01045", "citation_count": 23, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/ernie-research/Tool-Augmented-Reward-Model}}", "venue": "International Conference on Learning Representations", "quality_score": 0.3451}
{"id": "81905379d79ed510734d47c0cb9cab246b42e1ab1264f35f7bd6c780983afb26", "sources": ["arxiv", "semantic_scholar"], "title": "The Trickle-down Impact of Reward (In-)consistency on RLHF", "abstract": "Standard practice within Reinforcement Learning from Human Feedback (RLHF) involves optimizing against a Reward Model (RM), which itself is trained to reflect human preferences for desirable generations. A notable subject that is understudied is the (in-)consistency of RMs -- whether they can recognize the semantic changes to different prompts and appropriately adapt their reward assignments -- and their impact on the downstream RLHF model. In this paper, we visit a series of research questions relevant to RM inconsistency: (1) How can we measure the consistency of reward models? (2) How consistent are the existing RMs and how can we improve them? (3) In what ways does reward inconsistency influence the chatbots resulting from the RLHF model training? We propose Contrast Instructions -- a benchmarking strategy for the consistency of RM. Each example in Contrast Instructions features a pair of lexically similar instructions with different ground truth responses. A consistent RM is expected to rank the corresponding instruction and response higher than other combinations. We observe that current RMs trained with the standard ranking objective fail miserably on Contrast Instructions compared to average humans. To show that RM consistency can be improved efficiently without using extra training budget, we propose two techniques ConvexDA and RewardFusion, which enhance reward consistency through extrapolation during the RM training and inference stage, respectively. We show that RLHF models trained with a more consistent RM yield more useful responses, suggesting that reward inconsistency exhibits a trickle-down effect on the downstream RLHF process.", "authors": ["Lingfeng Shen", "Sihao Chen", "Linfeng Song", "Lifeng Jin", "Baolin Peng", "Haitao Mi", "Daniel Khashabi", "Dong Yu"], "categories": ["cs.CL", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-28", "url": "https://arxiv.org/abs/2309.16155", "pdf_url": "https://arxiv.org/pdf/2309.16155v1", "arxiv_id": "2309.16155", "doi": "10.48550/arXiv.2309.16155", "citation_count": 30, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3728}
{"id": "b1eecaff2eeb0244f0cca6608d9d7bc142b2fe2000ac119bf5c755913f7ad23c", "sources": ["arxiv", "semantic_scholar"], "title": "Text2Reward: Reward Shaping with Language Models for Reinforcement Learning", "abstract": "Designing reward functions is a longstanding challenge in reinforcement learning (RL); it requires specialized knowledge or domain data, leading to high costs for development. To address this, we introduce Text2Reward, a data-free framework that automates the generation and shaping of dense reward functions based on large language models (LLMs). Given a goal described in natural language, Text2Reward generates shaped dense reward functions as an executable program grounded in a compact representation of the environment. Unlike inverse RL and recent work that uses LLMs to write sparse reward codes or unshaped dense rewards with a constant function across timesteps, Text2Reward produces interpretable, free-form dense reward codes that cover a wide range of tasks, utilize existing packages, and allow iterative refinement with human feedback. We evaluate Text2Reward on two robotic manipulation benchmarks (ManiSkill2, MetaWorld) and two locomotion environments of MuJoCo. On 13 of the 17 manipulation tasks, policies trained with generated reward codes achieve similar or better task success rates and convergence speed than expert-written reward codes. For locomotion tasks, our method learns six novel locomotion behaviors with a success rate exceeding 94%. Furthermore, we show that the policies trained in the simulator with our method can be deployed in the real world. Finally, Text2Reward further improves the policies by refining their reward functions with human feedback. Video results are available at https://text-to-reward.github.io/ .", "authors": ["Tianbao Xie", "Siheng Zhao", "Chen Henry Wu", "Yitao Liu", "Qian Luo", "Victor Zhong", "Yanchao Yang", "Tao Yu"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-20", "url": "https://arxiv.org/abs/2309.11489", "pdf_url": "https://arxiv.org/pdf/2309.11489v3", "arxiv_id": "2309.11489", "doi": null, "citation_count": 159, "influential_citation_count": 15, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.6021}
{"id": "546b9eacc9b18f606e3f5e355c5d0a8858a1d96d7ba2102a219c57c8156a2ca9", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Engineering for Generating Semi-structured Explanation", "abstract": "Semi-structured explanation depicts the implicit process of a reasoner with an explicit representation. This explanation highlights how available information in a specific query is utilised and supplemented with information a reasoner produces from its internal weights towards generating an answer. Despite the recent improvements in generative capabilities of language models, producing structured explanations to verify a model's true reasoning capabilities remains a challenge. This issue is particularly pronounced for not-so-large LMs (e.g., FLAN-T5-XXL). In this work, we first underscore the limitations of supervised fine-tuning (SFT) in tackling this challenge, and then introduce a carefully crafted reward engineering method in reinforcement learning (RL) to better address this problem. We investigate multiple reward aggregation methods and provide a detailed discussion which sheds light on the promising potential of RL for future research. Our proposed method on two semi-structured explanation generation benchmarks (ExplaGraph and COPA-SSE) achieves new state-of-the-art results.", "authors": ["Jiuzhou Han", "Wray Buntine", "Ehsan Shareghi"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-15", "url": "https://arxiv.org/abs/2309.08347", "pdf_url": "https://arxiv.org/pdf/2309.08347v2", "arxiv_id": "2309.08347", "doi": "10.48550/arXiv.2309.08347", "citation_count": 1, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/Jiuzhouh/Reward-Engineering-for-Generating-SEG", "venue": "Findings", "quality_score": 0.0753}
{"id": "59cceaf646d8612f3854c352443ae3a68acc6ebafbe50e02863172eec517e44d", "sources": ["arxiv", "semantic_scholar"], "title": "Everyone Deserves A Reward: Learning Customized Human Preferences", "abstract": "Reward models (RMs) are essential for aligning large language models (LLMs) with human preferences to improve interaction quality. However, the real world is pluralistic, which leads to diversified human preferences with respect to different religions, politics, cultures, etc. Moreover, each individual can have their unique preferences on various topics. Neglecting the diversity of human preferences, current human feedback aligning methods only consider a general reward model, which is below satisfaction for customized or personalized application scenarios. To explore customized preference learning, we collect a domain-specific preference (DSP) dataset, which includes preferred responses for each given query from four practical domains. Besides, from the perspective of data efficiency, we propose a three-stage customized RM learning scheme, then empirically verify its effectiveness on both general preference datasets and our DSP set. Furthermore, we test multiple training and data strategies on the three learning stages. We find several ways to better preserve the general preferring ability while training the customized RMs, especially general preference enrichment, and customized preference imitation learning. The DSP dataset and code are available at https://github.com/Linear95/DSP.", "authors": ["Pengyu Cheng", "Jiawen Xie", "Ke Bai", "Yong Dai", "Nan Du"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-06", "url": "https://arxiv.org/abs/2309.03126", "pdf_url": "https://arxiv.org/pdf/2309.03126v2", "arxiv_id": "2309.03126", "doi": "10.48550/arXiv.2309.03126", "citation_count": 48, "influential_citation_count": 5, "has_code": true, "code_url": "https://github.com/Linear95/DSP", "venue": "arXiv.org", "quality_score": 0.4225}
{"id": "818ba0ef479d5c8bdca13fd00eefc63cdff07cad8656fff2c076ddc7d69557e9", "sources": ["arxiv", "semantic_scholar"], "title": "RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback", "abstract": "Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models (LLMs) with human preferences, but gathering high-quality preference labels is expensive. RL from AI Feedback (RLAIF), introduced in Bai et al., offers a promising alternative that trains the reward model (RM) on preferences generated by an off-the-shelf LLM. Across the tasks of summarization, helpful dialogue generation, and harmless dialogue generation, we show that RLAIF achieves comparable performance to RLHF. Furthermore, we take a step towards \"self-improvement\" by demonstrating that RLAIF can outperform a supervised fine-tuned baseline even when the AI labeler is the same size as the policy, or even the exact same checkpoint as the initial policy. Finally, we introduce direct-RLAIF (d-RLAIF) - a technique that circumvents RM training by obtaining rewards directly from an off-the-shelf LLM during RL, which achieves superior performance to canonical RLAIF. Our results suggest that RLAIF can achieve performance on-par with using human feedback, offering a potential solution to the scalability limitations of RLHF.", "authors": ["Harrison Lee", "Samrat Phatale", "Hassan Mansoor", "Thomas Mesnard", "Johan Ferret", "Kellie Lu", "Colton Bishop", "Ethan Hall", "Victor Carbune", "Abhinav Rastogi", "Sushant Prakash"], "categories": ["cs.CL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-09-01", "url": "https://arxiv.org/abs/2309.00267", "pdf_url": "https://arxiv.org/pdf/2309.00267v3", "arxiv_id": "2309.00267", "doi": null, "citation_count": 645, "influential_citation_count": 40, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.8064}
{"id": "bead925d6145fe6394e21e70bcec9c4600ac3f6e37868ed3f1a6387949dcadcd", "sources": ["arxiv", "semantic_scholar"], "title": "On Reward Structures of Markov Decision Processes", "abstract": "A Markov decision process can be parameterized by a transition kernel and a reward function. Both play essential roles in the study of reinforcement learning as evidenced by their presence in the Bellman equations. In our inquiry of various kinds of \"costs\" associated with reinforcement learning inspired by the demands in robotic applications, rewards are central to understanding the structure of a Markov decision process and reward-centric notions can elucidate important concepts in reinforcement learning. Specifically, we study the sample complexity of policy evaluation and develop a novel estimator with an instance-specific error bound of $\\tilde{O}(\\sqrt{\\frac{τ_s}{n}})$ for estimating a single state value. Under the online regret minimization setting, we refine the transition-based MDP constant, diameter, into a reward-based constant, maximum expected hitting cost, and with it, provide a theoretical explanation for how a well-known technique, potential-based reward shaping, could accelerate learning with expert knowledge. In an attempt to study safe reinforcement learning, we model hazardous environments with irrecoverability and proposed a quantitative notion of safe learning via reset efficiency. In this setting, we modify a classic algorithm to account for resets achieving promising preliminary numerical results. Lastly, for MDPs with multiple reward functions, we develop a planning algorithm that computationally efficiently finds Pareto-optimal stochastic policies.", "authors": ["Falcon Z. Dai"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-08-28", "url": "https://arxiv.org/abs/2308.14919", "pdf_url": "https://arxiv.org/pdf/2308.14919v2", "arxiv_id": "2308.14919", "doi": "10.48550/arXiv.2308.14919", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "3710392a797ab8900d3773ab86dcb67e5e91fe75dd8a2eda0f44e2a45aadd93b", "sources": ["arxiv", "semantic_scholar"], "title": "Language Reward Modulation for Pretraining Reinforcement Learning", "abstract": "Using learned reward functions (LRFs) as a means to solve sparse-reward reinforcement learning (RL) tasks has yielded some steady progress in task-complexity through the years. In this work, we question whether today's LRFs are best-suited as a direct replacement for task rewards. Instead, we propose leveraging the capabilities of LRFs as a pretraining signal for RL. Concretely, we propose $\\textbf{LA}$nguage Reward $\\textbf{M}$odulated $\\textbf{P}$retraining (LAMP) which leverages the zero-shot capabilities of Vision-Language Models (VLMs) as a $\\textit{pretraining}$ utility for RL as opposed to a downstream task reward. LAMP uses a frozen, pretrained VLM to scalably generate noisy, albeit shaped exploration rewards by computing the contrastive alignment between a highly diverse collection of language instructions and the image observations of an agent in its pretraining environment. LAMP optimizes these rewards in conjunction with standard novelty-seeking exploration rewards with reinforcement learning to acquire a language-conditioned, pretrained policy. Our VLM pretraining approach, which is a departure from previous attempts to use LRFs, can warmstart sample-efficient learning on robot manipulation tasks in RLBench.", "authors": ["Ademi Adeniji", "Amber Xie", "Carmelo Sferrazza", "Younggyo Seo", "Stephen James", "Pieter Abbeel"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-08-23", "url": "https://arxiv.org/abs/2308.12270", "pdf_url": "https://arxiv.org/pdf/2308.12270v1", "arxiv_id": "2308.12270", "doi": "10.48550/arXiv.2308.12270", "citation_count": 45, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/ademiadeniji/lamp", "venue": "arXiv.org", "quality_score": 0.4157}
{"id": "fcdcb3186923e6afe74ef601de2dc89f1bf615f312a5e9205e9fad42edfbfaed", "sources": ["arxiv", "semantic_scholar"], "title": "INFINITY: Neural Field Modeling for Reynolds-Averaged Navier-Stokes Equations", "abstract": "For numerical design, the development of efficient and accurate surrogate models is paramount. They allow us to approximate complex physical phenomena, thereby reducing the computational burden of direct numerical simulations. We propose INFINITY, a deep learning model that utilizes implicit neural representations (INRs) to address this challenge. Our framework encodes geometric information and physical fields into compact representations and learns a mapping between them to infer the physical fields. We use an airfoil design optimization problem as an example task and we evaluate our approach on the challenging AirfRANS dataset, which closely resembles real-world industrial use-cases. The experimental results demonstrate that our framework achieves state-of-the-art performance by accurately inferring physical fields throughout the volume and surface. Additionally we demonstrate its applicability in contexts such as design exploration and shape optimization: our model can correctly predict drag and lift coefficients while adhering to the equations.", "authors": ["Louis Serrano", "Leon Migus", "Yuan Yin", "Jocelyn Ahmed Mazari", "Patrick Gallinari"], "categories": ["cs.LG", "physics.flu-dyn"], "fields_of_study": ["Computer Science", "Physics"], "published_date": "2023-07-25", "url": "https://arxiv.org/abs/2307.13538", "pdf_url": "https://arxiv.org/pdf/2307.13538v1", "arxiv_id": "2307.13538", "doi": "10.48550/arXiv.2307.13538", "citation_count": 7, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2258}
{"id": "1d302e16a2cf112908bc425cda55b5ed834c9d81b2a410f77790baf5b60ff92e", "sources": ["arxiv", "semantic_scholar"], "title": "DIP-RL: Demonstration-Inferred Preference Learning in Minecraft", "abstract": "In machine learning for sequential decision-making, an algorithmic agent learns to interact with an environment while receiving feedback in the form of a reward signal. However, in many unstructured real-world settings, such a reward signal is unknown and humans cannot reliably craft a reward signal that correctly captures desired behavior. To solve tasks in such unstructured and open-ended environments, we present Demonstration-Inferred Preference Reinforcement Learning (DIP-RL), an algorithm that leverages human demonstrations in three distinct ways, including training an autoencoder, seeding reinforcement learning (RL) training batches with demonstration data, and inferring preferences over behaviors to learn a reward function to guide RL. We evaluate DIP-RL in a tree-chopping task in Minecraft. Results suggest that the method can guide an RL agent to learn a reward function that reflects human preferences and that DIP-RL performs competitively relative to baselines. DIP-RL is inspired by our previous work on combining demonstrations and pairwise preferences in Minecraft, which was awarded a research prize at the 2022 NeurIPS MineRL BASALT competition, Learning from Human Feedback in Minecraft. Example trajectory rollouts of DIP-RL and baselines are located at https://sites.google.com/view/dip-rl.", "authors": ["Ellen Novoseller", "Vinicius G. Goecks", "David Watkins", "Josh Miller", "Nicholas Waytowich"], "categories": ["cs.LG", "cs.AI", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-22", "url": "https://arxiv.org/abs/2307.12158", "pdf_url": "https://arxiv.org/pdf/2307.12158v1", "arxiv_id": "2307.12158", "doi": "10.48550/arXiv.2307.12158", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "71105a0145ae93831736a2f13205da847f729c10f2f812f3b0950d8b2c6a184b", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Formal Specifications from Membership and Preference Queries", "abstract": "Active learning is a well-studied approach to learning formal specifications, such as automata. In this work, we extend active specification learning by proposing a novel framework that strategically requests a combination of membership labels and pair-wise preferences, a popular alternative to membership labels. The combination of pair-wise preferences and membership labels allows for a more flexible approach to active specification learning, which previously relied on membership labels only. We instantiate our framework in two different domains, demonstrating the generality of our approach. Our results suggest that learning from both modalities allows us to robustly and conveniently identify specifications via membership and preferences.", "authors": ["Ameesh Shah", "Marcell Vazquez-Chanlatte", "Sebastian Junges", "Sanjit A. Seshia"], "categories": ["cs.FL", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-19", "url": "https://arxiv.org/abs/2307.10434", "pdf_url": "https://arxiv.org/pdf/2307.10434v2", "arxiv_id": "2307.10434", "doi": "10.48550/arXiv.2307.10434", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.2113}
{"id": "313deb4d873f0a88e1f0a1ab6c342c2aae5320c0f7503fb3fc151e68b9ae111b", "sources": ["arxiv", "semantic_scholar"], "title": "Benchmarking Potential Based Rewards for Learning Humanoid Locomotion", "abstract": "The main challenge in developing effective reinforcement learning (RL) pipelines is often the design and tuning the reward functions. Well-designed shaping reward can lead to significantly faster learning. Naively formulated rewards, however, can conflict with the desired behavior and result in overfitting or even erratic performance if not properly tuned. In theory, the broad class of potential based reward shaping (PBRS) can help guide the learning process without affecting the optimal policy. Although several studies have explored the use of potential based reward shaping to accelerate learning convergence, most have been limited to grid-worlds and low-dimensional systems, and RL in robotics has predominantly relied on standard forms of reward shaping. In this paper, we benchmark standard forms of shaping with PBRS for a humanoid robot. We find that in this high-dimensional system, PBRS has only marginal benefits in convergence speed. However, the PBRS reward terms are significantly more robust to scaling than typical reward shaping approaches, and thus easier to tune.", "authors": ["Se Hwan Jeon", "Steve Heim", "Charles Khazoom", "Sangbae Kim"], "categories": ["cs.RO", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-07-19", "url": "https://arxiv.org/abs/2307.10142", "pdf_url": "https://arxiv.org/pdf/2307.10142v1", "arxiv_id": "2307.10142", "doi": "10.1109/ICRA48891.2023.10160885", "citation_count": 30, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE International Conference on Robotics and Automation", "quality_score": 0.3728}
{"id": "9b8705228fd9008dc0fa048d650c6aca1c6fa5af47d3fd4877000decd8bd3f10", "sources": ["arxiv", "semantic_scholar"], "title": "Can Differentiable Decision Trees Enable Interpretable Reward Learning from Human Feedback?", "abstract": "Reinforcement Learning from Human Feedback (RLHF) has emerged as a popular paradigm for capturing human intent to alleviate the challenges of hand-crafting the reward values. Despite the increasing interest in RLHF, most works learn black box reward functions that while expressive are difficult to interpret and often require running the whole costly process of RL before we can even decipher if these frameworks are actually aligned with human preferences. We propose and evaluate a novel approach for learning expressive and interpretable reward functions from preferences using Differentiable Decision Trees (DDTs). Our experiments across several domains, including CartPole, Visual Gridworld environments and Atari games, provide evidence that the tree structure of our learned reward function is useful in determining the extent to which the reward function is aligned with human preferences. We also provide experimental evidence that not only shows that reward DDTs can often achieve competitive RL performance when compared with larger capacity deep neural network reward functions but also demonstrates the diagnostic utility of our framework in checking alignment of learned reward functions. We also observe that the choice between soft and hard (argmax) output of reward DDT reveals a tension between wanting highly shaped rewards to ensure good RL performance, while also wanting simpler, more interpretable rewards. Videos and code, are available at: https://sites.google.com/view/ddt-rlhf", "authors": ["Akansha Kalra", "Daniel S. Brown"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-22", "url": "https://arxiv.org/abs/2306.13004", "pdf_url": "https://arxiv.org/pdf/2306.13004v6", "arxiv_id": "2306.13004", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "49197981adf1b66fddf59c1a391ded337ee1fdfe82ea2cce2e586aec10b7427e", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Shaping via Diffusion Process in Reinforcement Learning", "abstract": "Reinforcement Learning (RL) models have continually evolved to navigate the exploration - exploitation trade-off in uncertain Markov Decision Processes (MDPs). In this study, I leverage the principles of stochastic thermodynamics and system dynamics to explore reward shaping via diffusion processes. This provides an elegant framework as a way to think about exploration-exploitation trade-off. This article sheds light on relationships between information entropy, stochastic system dynamics, and their influences on entropy production. This exploration allows us to construct a dual-pronged framework that can be interpreted as either a maximum entropy program for deriving efficient policies or a modified cost optimization program accounting for informational costs and benefits. This work presents a novel perspective on the physical nature of information and its implications for online learning in MDPs, consequently providing a better understanding of information-oriented formulations in RL.", "authors": ["Peeyush Kumar"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-20", "url": "https://arxiv.org/abs/2306.11885", "pdf_url": "https://arxiv.org/pdf/2306.11885v1", "arxiv_id": "2306.11885", "doi": "10.48550/arXiv.2306.11885", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "77c6e6c40fe02b83e36e1d785cbbac41dfdbc95f9f1f93f94e762a9f90bd9075", "sources": ["arxiv", "semantic_scholar"], "title": "Fairness in Preference-based Reinforcement Learning", "abstract": "In this paper, we address the issue of fairness in preference-based reinforcement learning (PbRL) in the presence of multiple objectives. The main objective is to design control policies that can optimize multiple objectives while treating each objective fairly. Toward this objective, we design a new fairness-induced preference-based reinforcement learning or FPbRL. The main idea of FPbRL is to learn vector reward functions associated with multiple objectives via new welfare-based preferences rather than reward-based preference in PbRL, coupled with policy learning via maximizing a generalized Gini welfare function. Finally, we provide experiment studies on three different environments to show that the proposed FPbRL approach can achieve both efficiency and equity for learning effective and fair policies.", "authors": ["Umer Siddique", "Abhinav Sinha", "Yongcan Cao"], "categories": ["cs.LG", "cs.AI", "cs.CY", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2023-06-16", "url": "https://arxiv.org/abs/2306.09995", "pdf_url": "https://arxiv.org/pdf/2306.09995v2", "arxiv_id": "2306.09995", "doi": "10.48550/arXiv.2306.09995", "citation_count": 8, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "28ee071fd7578a584599803970f89e61393c1ccc4e7118d6bc26a9eef3ddb117", "sources": ["arxiv", "semantic_scholar"], "title": "PEARL: Zero-shot Cross-task Preference Alignment and Robust Reward Learning for Robotic Manipulation", "abstract": "In preference-based Reinforcement Learning (RL), obtaining a large number of preference labels are both time-consuming and costly. Furthermore, the queried human preferences cannot be utilized for the new tasks. In this paper, we propose Zero-shot Cross-task Preference Alignment and Robust Reward Learning (PEARL), which learns policies from cross-task preference transfer without any human labels of the target task. Our contributions include two novel components that facilitate the transfer and learning process. The first is Cross-task Preference Alignment (CPA), which transfers the preferences between tasks via optimal transport. The key idea of CPA is to use Gromov-Wasserstein distance to align the trajectories between tasks, and the solved optimal transport matrix serves as the correspondence between trajectories. The target task preferences are computed as the weighted sum of source task preference labels with the correspondence as weights. Moreover, to ensure robust learning from these transferred labels, we introduce Robust Reward Learning (RRL), which considers both reward mean and uncertainty by modeling rewards as Gaussian distributions. Empirical results on robotic manipulation tasks from Meta-World and Robomimic demonstrate that our method is capable of transferring preference labels across tasks accurately and then learns well-behaved policies. Notably, our approach significantly exceeds existing methods when there are few human preferences. The code and videos of our method are available at: https://sites.google.com/view/pearl-preference.", "authors": ["Runze Liu", "Yali Du", "Fengshuo Bai", "Jiafei Lyu", "Xiu Li"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-06-06", "url": "https://arxiv.org/abs/2306.03615", "pdf_url": "https://arxiv.org/pdf/2306.03615v2", "arxiv_id": "2306.03615", "doi": null, "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.2785}
{"id": "8bbd59382ecefc037327569d6947c1cb5610afecf75ee11779359a24e3513787", "sources": ["arxiv", "semantic_scholar"], "title": "Direct Preference Optimization: Your Language Model is Secretly a Reward Model", "abstract": "While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the large unsupervised LM using reinforcement learning to maximize this estimated reward without drifting too far from the original model. In this paper we introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form, allowing us to solve the standard RLHF problem with only a simple classification loss. The resulting algorithm, which we call Direct Preference Optimization (DPO), is stable, performant, and computationally lightweight, eliminating the need for sampling from the LM during fine-tuning or performing significant hyperparameter tuning. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods. Notably, fine-tuning with DPO exceeds PPO-based RLHF in ability to control sentiment of generations, and matches or improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train.", "authors": ["Rafael Rafailov", "Archit Sharma", "Eric Mitchell", "Stefano Ermon", "Christopher D. Manning", "Chelsea Finn"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-29", "url": "https://arxiv.org/abs/2305.18290", "pdf_url": "https://arxiv.org/pdf/2305.18290v3", "arxiv_id": "2305.18290", "doi": "10.52202/075280-2338", "citation_count": 9191, "influential_citation_count": 1975, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 1.0}
{"id": "7039d1a52585449d7d84aa0cd06ec1462ae834ba517cf67edf396c816b493460", "sources": ["arxiv", "semantic_scholar"], "title": "Provable Reward-Agnostic Preference-Based Reinforcement Learning", "abstract": "Preference-based Reinforcement Learning (PbRL) is a paradigm in which an RL agent learns to optimize a task using pair-wise preference-based feedback over trajectories, rather than explicit reward signals. While PbRL has demonstrated practical success in fine-tuning language models, existing theoretical work focuses on regret minimization and fails to capture most of the practical frameworks. In this study, we fill in such a gap between theoretical PbRL and practical algorithms by proposing a theoretical reward-agnostic PbRL framework where exploratory trajectories that enable accurate learning of hidden reward functions are acquired before collecting any human feedback. Theoretical analysis demonstrates that our algorithm requires less human feedback for learning the optimal policy under preference-based models with linear parameterization and unknown transitions, compared to the existing theoretical literature. Specifically, our framework can incorporate linear and low-rank MDPs with efficient sample complexity. Additionally, we investigate reward-agnostic RL with action-based comparison feedback and introduce an efficient querying algorithm tailored to this scenario.", "authors": ["Wenhao Zhan", "Masatoshi Uehara", "Wen Sun", "Jason D. Lee"], "categories": ["cs.LG", "cs.AI", "math.ST", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-05-29", "url": "https://arxiv.org/abs/2305.18505", "pdf_url": "https://arxiv.org/pdf/2305.18505v3", "arxiv_id": "2305.18505", "doi": null, "citation_count": 18, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3197}
{"id": "8d6d30e11cbf6a248caa2607e2dbb8c9307a9a730afb2270fe7158d58ebc804d", "sources": ["arxiv", "semantic_scholar"], "title": "Beyond Reward: Offline Preference-guided Policy Optimization", "abstract": "This study focuses on the topic of offline preference-based reinforcement learning (PbRL), a variant of conventional reinforcement learning that dispenses with the need for online interaction or specification of reward functions. Instead, the agent is provided with fixed offline trajectories and human preferences between pairs of trajectories to extract the dynamics and task information, respectively. Since the dynamics and task information are orthogonal, a naive approach would involve using preference-based reward learning followed by an off-the-shelf offline RL algorithm. However, this requires the separate learning of a scalar reward function, which is assumed to be an information bottleneck of the learning process. To address this issue, we propose the offline preference-guided policy optimization (OPPO) paradigm, which models offline trajectories and preferences in a one-step process, eliminating the need for separately learning a reward function. OPPO achieves this by introducing an offline hindsight information matching objective for optimizing a contextual policy and a preference modeling objective for finding the optimal context. OPPO further integrates a well-performing decision policy by optimizing the two objectives iteratively. Our empirical results demonstrate that OPPO effectively models offline preferences and outperforms prior competing baselines, including offline RL algorithms performed over either true or pseudo reward function specifications. Our code is available on the project website: https://sites.google.com/view/oppo-icml-2023 .", "authors": ["Yachen Kang", "Diyuan Shi", "Jinxin Liu", "Li He", "Donglin Wang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-25", "url": "https://arxiv.org/abs/2305.16217", "pdf_url": "https://arxiv.org/pdf/2305.16217v2", "arxiv_id": "2305.16217", "doi": "10.48550/arXiv.2305.16217", "citation_count": 44, "influential_citation_count": 6, "has_code": true, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4225}
{"id": "30f973c8edb8e0a5496a5062a18b036c1277942a426ce525e05b239dcd749107", "sources": ["arxiv", "semantic_scholar"], "title": "Inverse Preference Learning: Preference-based RL without a Reward Function", "abstract": "Reward functions are difficult to design and often hard to align with human intent. Preference-based Reinforcement Learning (RL) algorithms address these problems by learning reward functions from human feedback. However, the majority of preference-based RL methods naïvely combine supervised reward models with off-the-shelf RL algorithms. Contemporary approaches have sought to improve performance and query complexity by using larger and more complex reward architectures such as transformers. Instead of using highly complex architectures, we develop a new and parameter-efficient algorithm, Inverse Preference Learning (IPL), specifically designed for learning from offline preference data. Our key insight is that for a fixed policy, the $Q$-function encodes all information about the reward function, effectively making them interchangeable. Using this insight, we completely eliminate the need for a learned reward function. Our resulting algorithm is simpler and more parameter-efficient. Across a suite of continuous control and robotics benchmarks, IPL attains competitive performance compared to more complex approaches that leverage transformer-based and non-Markovian reward functions while having fewer algorithmic hyperparameters and learned network parameters. Our code is publicly released.", "authors": ["Joey Hejna", "Dorsa Sadigh"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-24", "url": "https://arxiv.org/abs/2305.15363", "pdf_url": "https://arxiv.org/pdf/2305.15363v2", "arxiv_id": "2305.15363", "doi": "10.48550/arXiv.2305.15363", "citation_count": 83, "influential_citation_count": 13, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5731}
{"id": "f3b03923589ba300db26f169b3524ad4704241c3c37a744e497d2dc0cc61a242", "sources": ["arxiv", "semantic_scholar"], "title": "Video Prediction Models as Rewards for Reinforcement Learning", "abstract": "Specifying reward signals that allow agents to learn complex behaviors is a long-standing challenge in reinforcement learning. A promising approach is to extract preferences for behaviors from unlabeled videos, which are widely available on the internet. We present Video Prediction Rewards (VIPER), an algorithm that leverages pretrained video prediction models as action-free reward signals for reinforcement learning. Specifically, we first train an autoregressive transformer on expert videos and then use the video prediction likelihoods as reward signals for a reinforcement learning agent. VIPER enables expert-level control without programmatic task rewards across a wide range of DMC, Atari, and RLBench tasks. Moreover, generalization of the video prediction model allows us to derive rewards for an out-of-distribution environment where no expert data is available, enabling cross-embodiment generalization for tabletop manipulation. We see our work as starting point for scalable reward specification from unlabeled videos that will benefit from the rapid advances in generative modeling. Source code and datasets are available on the project website: https://escontrela.me/viper", "authors": ["Alejandro Escontrela", "Ademi Adeniji", "Wilson Yan", "Ajay Jain", "Xue Bin Peng", "Ken Goldberg", "Youngwoon Lee", "Danijar Hafner", "Pieter Abbeel"], "categories": ["cs.LG", "cs.AI", "cs.CV"], "fields_of_study": ["Computer Science"], "published_date": "2023-05-23", "url": "https://arxiv.org/abs/2305.14343", "pdf_url": "https://arxiv.org/pdf/2305.14343v2", "arxiv_id": "2305.14343", "doi": "10.48550/arXiv.2305.14343", "citation_count": 106, "influential_citation_count": 7, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5073}
{"id": "9f557b9017e5d29b7b3e2fddaecf9fdbc37a7f5de9056d1088049c04635437a3", "sources": ["arxiv", "semantic_scholar"], "title": "Minimax-Optimal Reward-Agnostic Exploration in Reinforcement Learning", "abstract": "This paper studies reward-agnostic exploration in reinforcement learning (RL) -- a scenario where the learner is unware of the reward functions during the exploration stage -- and designs an algorithm that improves over the state of the art. More precisely, consider a finite-horizon inhomogeneous Markov decision process with $S$ states, $A$ actions, and horizon length $H$, and suppose that there are no more than a polynomial number of given reward functions of interest. By collecting an order of \\begin{align*} \\frac{SAH^3}{\\varepsilon^2} \\text{ sample episodes (up to log factor)} \\end{align*} without guidance of the reward information, our algorithm is able to find $\\varepsilon$-optimal policies for all these reward functions, provided that $\\varepsilon$ is sufficiently small. This forms the first reward-agnostic exploration scheme in this context that achieves provable minimax optimality. Furthermore, once the sample size exceeds $\\frac{S^2AH^3}{\\varepsilon^2}$ episodes (up to log factor), our algorithm is able to yield $\\varepsilon$ accuracy for arbitrarily many reward functions (even when they are adversarially designed), a task commonly dubbed as ``reward-free exploration.'' The novelty of our algorithm design draws on insights from offline RL: the exploration scheme attempts to maximize a critical reward-agnostic quantity that dictates the performance of offline RL, while the policy learning paradigm leverages ideas from sample-optimal offline RL paradigms.", "authors": ["Gen Li", "Yuling Yan", "Yuxin Chen", "Jianqing Fan"], "categories": ["cs.LG", "cs.IT", "eess.SY", "math.ST", "stat.ML"], "fields_of_study": ["Computer Science", "Engineering", "Mathematics"], "published_date": "2023-04-14", "url": "https://arxiv.org/abs/2304.07278", "pdf_url": "https://arxiv.org/pdf/2304.07278v2", "arxiv_id": "2304.07278", "doi": "10.48550/arXiv.2304.07278", "citation_count": 16, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Annual Conference Computational Learning Theory", "quality_score": 0.3076}
{"id": "ad637aeab56d49f8f4c53246fe5e22231f2fe0121f3ddb7d11966a598d56606e", "sources": ["arxiv", "semantic_scholar"], "title": "RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment", "abstract": "Generative foundation models are susceptible to implicit biases that can arise from extensive unsupervised training data. Such biases can produce suboptimal samples, skewed outcomes, and unfairness, with potentially serious consequences. Consequently, aligning these models with human ethics and preferences is an essential step toward ensuring their responsible and effective deployment in real-world applications. Prior research has primarily employed Reinforcement Learning from Human Feedback (RLHF) to address this problem, where generative models are fine-tuned with RL algorithms guided by a human-feedback-informed reward model. However, the inefficiencies and instabilities associated with RL algorithms frequently present substantial obstacles to the successful alignment, necessitating the development of a more robust and streamlined approach. To this end, we introduce a new framework, Reward rAnked FineTuning (RAFT), designed to align generative models effectively. Utilizing a reward model and a sufficient number of samples, our approach selects the high-quality samples, discarding those that exhibit undesired behavior, and subsequently enhancing the model by fine-tuning on these filtered samples. Our studies show that RAFT can effectively improve the model performance in both reward learning and other automated metrics in both large language models and diffusion models.", "authors": ["Hanze Dong", "Wei Xiong", "Deepanshu Goyal", "Yihan Zhang", "Winnie Chow", "Rui Pan", "Shizhe Diao", "Jipeng Zhang", "Kashun Shum", "Tong Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CL", "cs.CV", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-04-13", "url": "https://arxiv.org/abs/2304.06767", "pdf_url": "https://arxiv.org/pdf/2304.06767v4", "arxiv_id": "2304.06767", "doi": "10.48550/arXiv.2304.06767", "citation_count": 738, "influential_citation_count": 68, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.9194}
{"id": "cf6d0c04d9eedb2b057b5e8950b42e198ec3d2540dea550364fcd7317c2ba475", "sources": ["arxiv", "semantic_scholar"], "title": "Full Gradient Deep Reinforcement Learning for Average-Reward Criterion", "abstract": "We extend the provably convergent Full Gradient DQN algorithm for discounted reward Markov decision processes from Avrachenkov et al. (2021) to average reward problems. We experimentally compare widely used RVI Q-Learning with recently proposed Differential Q-Learning in the neural function approximation setting with Full Gradient DQN and DQN. We also extend this to learn Whittle indices for Markovian restless multi-armed bandits. We observe a better convergence rate of the proposed Full Gradient variant across different tasks.", "authors": ["Tejas Pagare", "Vivek Borkar", "Konstantin Avrachenkov"], "categories": ["eess.SY", "cs.LG"], "fields_of_study": ["Engineering", "Computer Science"], "published_date": "2023-04-07", "url": "https://arxiv.org/abs/2304.03729", "pdf_url": "https://arxiv.org/pdf/2304.03729v1", "arxiv_id": "2304.03729", "doi": "10.48550/arXiv.2304.03729", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Learning for Dynamics & Control", "quality_score": 0.2113}
{"id": "49cc3cb48592253763df5f730fe38288271b1faf38f76d87fe761031f64a7262", "sources": ["arxiv", "semantic_scholar"], "title": "Decision-Focused Model-based Reinforcement Learning for Reward Transfer", "abstract": "Model-based reinforcement learning (MBRL) provides a way to learn a transition model of the environment, which can then be used to plan personalized policies for different patient cohorts and to understand the dynamics involved in the decision-making process. However, standard MBRL algorithms are either sensitive to changes in the reward function or achieve suboptimal performance on the task when the transition model is restricted. Motivated by the need to use simple and interpretable models in critical domains such as healthcare, we propose a novel robust decision-focused (RDF) algorithm that learns a transition model that achieves high returns while being robust to changes in the reward function. We demonstrate our RDF algorithm can be used with several model classes and planning algorithms. We also provide theoretical and empirical evidence, on a variety of simulators and real patient data, that RDF can learn simple yet effective models that can be used to plan personalized policies.", "authors": ["Abhishek Sharma", "Sonali Parbhoo", "Omer Gottesman", "Finale Doshi-Velez"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-04-06", "url": "https://arxiv.org/abs/2304.03365", "pdf_url": "https://arxiv.org/pdf/2304.03365v3", "arxiv_id": "2304.03365", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0753}
{"id": "375a5a225433dfd382927b6cabe0acc20cca85ec3baa3718ca4354308510908b", "sources": ["arxiv", "semantic_scholar"], "title": "BC-IRL: Learning Generalizable Reward Functions from Demonstrations", "abstract": "How well do reward functions learned with inverse reinforcement learning (IRL) generalize? We illustrate that state-of-the-art IRL algorithms, which maximize a maximum-entropy objective, learn rewards that overfit to the demonstrations. Such rewards struggle to provide meaningful rewards for states not covered by the demonstrations, a major detriment when using the reward to learn policies in new situations. We introduce BC-IRL a new inverse reinforcement learning method that learns reward functions that generalize better when compared to maximum-entropy IRL approaches. In contrast to the MaxEnt framework, which learns to maximize rewards around demonstrations, BC-IRL updates reward parameters such that the policy trained with the new reward matches the expert demonstrations better. We show that BC-IRL learns rewards that generalize better on an illustrative simple task and two continuous robotic control tasks, achieving over twice the success rate of baselines in challenging generalization settings.", "authors": ["Andrew Szot", "Amy Zhang", "Dhruv Batra", "Zsolt Kira", "Franziska Meier"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-03-28", "url": "https://arxiv.org/abs/2303.16194", "pdf_url": "https://arxiv.org/pdf/2303.16194v1", "arxiv_id": "2303.16194", "doi": "10.48550/arXiv.2303.16194", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2785}
{"id": "33e3af27cc1a3437267c73340782267bcaa7b6887b86ad9e73ec0166c67af22e", "sources": ["arxiv", "semantic_scholar"], "title": "Reinforcement Learning with Exogenous States and Rewards", "abstract": "Exogenous state variables and rewards can slow reinforcement learning by injecting uncontrolled variation into the reward signal. This paper formalizes exogenous state variables and rewards and shows that if the reward function decomposes additively into endogenous and exogenous components, the MDP can be decomposed into an exogenous Markov Reward Process (based on the exogenous reward) and an endogenous Markov Decision Process (optimizing the endogenous reward). Any optimal policy for the endogenous MDP is also an optimal policy for the original MDP, but because the endogenous reward typically has reduced variance, the endogenous MDP is easier to solve. We study settings where the decomposition of the state space into exogenous and endogenous state spaces is not given but must be discovered. The paper introduces and proves correctness of algorithms for discovering the exogenous and endogenous subspaces of the state space when they are mixed through linear combination. These algorithms can be applied during reinforcement learning to discover the exogenous subspace, remove the exogenous reward, and focus reinforcement learning on the endogenous MDP. Experiments on a variety of challenging synthetic MDPs show that these methods, applied online, discover large exogenous state spaces and produce substantial speedups in reinforcement learning.", "authors": ["George Trimponias", "Thomas G. Dietterich"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-03-22", "url": "https://arxiv.org/abs/2303.12957", "pdf_url": "https://arxiv.org/pdf/2303.12957v2", "arxiv_id": "2303.12957", "doi": "10.48550/arXiv.2303.12957", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "6461b996c31409c8be17331f9b7e488309aea71a26fd121366c4a7533844895a", "sources": ["arxiv", "semantic_scholar"], "title": "Preference Transformer: Modeling Human Preferences using Transformers for RL", "abstract": "Preference-based reinforcement learning (RL) provides a framework to train agents using human preferences between two behaviors. However, preference-based RL has been challenging to scale since it requires a large amount of human feedback to learn a reward function aligned with human intent. In this paper, we present Preference Transformer, a neural architecture that models human preferences using transformers. Unlike prior approaches assuming human judgment is based on the Markovian rewards which contribute to the decision equally, we introduce a new preference model based on the weighted sum of non-Markovian rewards. We then design the proposed preference model using a transformer architecture that stacks causal and bidirectional self-attention layers. We demonstrate that Preference Transformer can solve a variety of control tasks using real human preferences, while prior approaches fail to work. We also show that Preference Transformer can induce a well-specified reward and attend to critical events in the trajectory by automatically capturing the temporal dependencies in human decision-making. Code is available on the project website: https://sites.google.com/view/preference-transformer.", "authors": ["Changyeon Kim", "Jongjin Park", "Jinwoo Shin", "Honglak Lee", "Pieter Abbeel", "Kimin Lee"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2023-03-02", "url": "https://arxiv.org/abs/2303.00957", "pdf_url": "https://arxiv.org/pdf/2303.00957v1", "arxiv_id": "2303.00957", "doi": "10.48550/arXiv.2303.00957", "citation_count": 108, "influential_citation_count": 29, "has_code": true, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.7386}
{"id": "3f39c01c8cc2c15fbc578c9f1e771b73bf6109d38eeec3aff8fd91b5aa5c800d", "sources": ["arxiv", "semantic_scholar"], "title": "Active Reward Learning from Multiple Teachers", "abstract": "Reward learning algorithms utilize human feedback to infer a reward function, which is then used to train an AI system. This human feedback is often a preference comparison, in which the human teacher compares several samples of AI behavior and chooses which they believe best accomplishes the objective. While reward learning typically assumes that all feedback comes from a single teacher, in practice these systems often query multiple teachers to gather sufficient training data. In this paper, we investigate this disparity, and find that algorithmic evaluation of these different sources of feedback facilitates more accurate and efficient reward learning. We formally analyze the value of information (VOI) when reward learning from teachers with varying levels of rationality, and define and evaluate an algorithm that utilizes this VOI to actively select teachers to query for feedback. Surprisingly, we find that it is often more informative to query comparatively irrational teachers. By formalizing this problem and deriving an analytical solution, we hope to facilitate improvement in reward learning approaches to aligning AI behavior with human values.", "authors": ["Peter Barnett", "Rachel Freedman", "Justin Svegliato", "Stuart Russell"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-03-02", "url": "https://arxiv.org/abs/2303.00894", "pdf_url": "https://arxiv.org/pdf/2303.00894v1", "arxiv_id": "2303.00894", "doi": "10.48550/arXiv.2303.00894", "citation_count": 19, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3253}
{"id": "05a91af7a8449cb7e174670fd98c3a9a30feba7f7d147f0d560f461d408286b7", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Design with Language Models", "abstract": "Reward design in reinforcement learning (RL) is challenging since specifying human notions of desired behavior may be difficult via reward functions or require many expert demonstrations. Can we instead cheaply design rewards using a natural language interface? This paper explores how to simplify reward design by prompting a large language model (LLM) such as GPT-3 as a proxy reward function, where the user provides a textual prompt containing a few examples (few-shot) or a description (zero-shot) of the desired behavior. Our approach leverages this proxy reward function in an RL framework. Specifically, users specify a prompt once at the beginning of training. During training, the LLM evaluates an RL agent's behavior against the desired behavior described by the prompt and outputs a corresponding reward signal. The RL agent then uses this reward to update its behavior. We evaluate whether our approach can train agents aligned with user objectives in the Ultimatum Game, matrix games, and the DealOrNoDeal negotiation task. In all three tasks, we show that RL agents trained with our framework are well-aligned with the user's objectives and outperform RL agents trained with reward functions learned via supervised learning", "authors": ["Minae Kwon", "Sang Michael Xie", "Kalesha Bullard", "Dorsa Sadigh"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-27", "url": "https://arxiv.org/abs/2303.00001", "pdf_url": "https://arxiv.org/pdf/2303.00001v1", "arxiv_id": "2303.00001", "doi": "10.48550/arXiv.2303.00001", "citation_count": 314, "influential_citation_count": 9, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.6246}
{"id": "7b83bd41d827c63eccdd5501c0b468121a03c4a68e9503930c75eb521509da56", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Learning as Doubly Nonparametric Bandits: Optimal Design and Scaling Laws", "abstract": "Specifying reward functions for complex tasks like object manipulation or driving is challenging to do by hand. Reward learning seeks to address this by learning a reward model using human feedback on selected query policies. This shifts the burden of reward specification to the optimal design of the queries. We propose a theoretical framework for studying reward learning and the associated optimal experiment design problem. Our framework models rewards and policies as nonparametric functions belonging to subsets of Reproducing Kernel Hilbert Spaces (RKHSs). The learner receives (noisy) oracle access to a true reward and must output a policy that performs well under the true reward. For this setting, we first derive non-asymptotic excess risk bounds for a simple plug-in estimator based on ridge regression. We then solve the query design problem by optimizing these risk bounds with respect to the choice of query set and obtain a finite sample statistical rate, which depends primarily on the eigenvalue spectrum of a certain linear operator on the RKHSs. Despite the generality of these results, our bounds are stronger than previous bounds developed for more specialized problems. We specifically show that the well-studied problem of Gaussian process (GP) bandit optimization is a special case of our framework, and that our bounds either improve or are competitive with known regret guarantees for the Matérn kernel.", "authors": ["Kush Bhatia", "Wenshuo Guo", "Jacob Steinhardt"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2023-02-23", "url": "https://arxiv.org/abs/2302.12349", "pdf_url": "https://arxiv.org/pdf/2302.12349v1", "arxiv_id": "2302.12349", "doi": "10.48550/arXiv.2302.12349", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.0}
{"id": "3d16995ecccacfa17aaef6ac4403685db1b18978369088d4a9006af15e3c206b", "sources": ["arxiv", "semantic_scholar"], "title": "Data Driven Reward Initialization for Preference based Reinforcement Learning", "abstract": "Preference-based Reinforcement Learning (PbRL) methods utilize binary feedback from the human in the loop (HiL) over queried trajectory pairs to learn a reward model in an attempt to approximate the human's underlying reward function capturing their preferences. In this work, we investigate the issue of a high degree of variability in the initialized reward models which are sensitive to random seeds of the experiment. This further compounds the issue of degenerate reward functions PbRL methods already suffer from. We propose a data-driven reward initialization method that does not add any additional cost to the human in the loop and negligible cost to the PbRL agent and show that doing so ensures that the predicted rewards of the initialized reward model are uniform in the state space and this reduces the variability in the performance of the method across multiple runs and is shown to improve the overall performance compared to other initialization methods.", "authors": ["Mudit Verma", "Subbarao Kambhampati"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-17", "url": "https://arxiv.org/abs/2302.08733", "pdf_url": "https://arxiv.org/pdf/2302.08733v1", "arxiv_id": "2302.08733", "doi": "10.48550/arXiv.2302.08733", "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "e2385b6851e40786af691a97799557e5c195f0dfff1f7d3e2eda910a037ad3e5", "sources": ["arxiv", "semantic_scholar"], "title": "Adversarial Rewards in Universal Learning for Contextual Bandits", "abstract": "We study the fundamental limits of learning in contextual bandits, where a learner's rewards depend on their actions and a known context, which extends the canonical multi-armed bandit to the case where side-information is available. We are interested in universally consistent algorithms, which achieve sublinear regret compared to any measurable fixed policy, without any function class restriction. For stationary contextual bandits, when the underlying reward mechanism is time-invariant, Blanchard et. al (2022) characterized learnable context processes for which universal consistency is achievable; and further gave algorithms ensuring universal consistency whenever this is achievable, a property known as optimistic universal consistency. It is well understood, however, that reward mechanisms can evolve over time, possibly adversarially, and depending on the learner's actions. We show that optimistic universal learning for contextual bandits with adversarial rewards is impossible in general, contrary to all previously studied settings in online learning -- including standard supervised learning. We also give necessary and sufficient conditions for universal learning under various adversarial reward models, and an exact characterization for online rewards. In particular, the set of learnable processes for these reward models is still extremely general -- larger than i.i.d., stationary or ergodic -- but in general strictly smaller than that for supervised learning or stationary contextual bandits, shedding light on new adversarial phenomena.", "authors": ["Moise Blanchard", "Steve Hanneke", "Patrick Jaillet"], "categories": ["stat.ML", "cs.LG", "math.ST"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2023-02-14", "url": "https://arxiv.org/abs/2302.07186", "pdf_url": "https://arxiv.org/pdf/2302.07186v2", "arxiv_id": "2302.07186", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.1945}
{"id": "e4b8a670f931e7d0273cdc0661f0e8199a3969efbabcbb22ea130303030c78a7", "sources": ["arxiv", "semantic_scholar"], "title": "CLARE: Conservative Model-Based Reward Learning for Offline Inverse Reinforcement Learning", "abstract": "This work aims to tackle a major challenge in offline Inverse Reinforcement Learning (IRL), namely the reward extrapolation error, where the learned reward function may fail to explain the task correctly and misguide the agent in unseen environments due to the intrinsic covariate shift. Leveraging both expert data and lower-quality diverse data, we devise a principled algorithm (namely CLARE) that solves offline IRL efficiently via integrating \"conservatism\" into a learned reward function and utilizing an estimated dynamics model. Our theoretical analysis provides an upper bound on the return gap between the learned policy and the expert policy, based on which we characterize the impact of covariate shift by examining subtle two-tier tradeoffs between the exploitation (on both expert and diverse data) and exploration (on the estimated dynamics model). We show that CLARE can provably alleviate the reward extrapolation error by striking the right exploitation-exploration balance therein. Extensive experiments corroborate the significant performance gains of CLARE over existing state-of-the-art algorithms on MuJoCo continuous control tasks (especially with a small offline dataset), and the learned reward is highly instructive for further learning.", "authors": ["Sheng Yue", "Guanbo Wang", "Wei Shao", "Zhaofeng Zhang", "Sen Lin", "Ju Ren", "Junshan Zhang"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-09", "url": "https://arxiv.org/abs/2302.04782", "pdf_url": "https://arxiv.org/pdf/2302.04782v2", "arxiv_id": "2302.04782", "doi": "10.48550/arXiv.2302.04782", "citation_count": 26, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3578}
{"id": "5a156222dc68c69e799e3dde38f1d9863bcbd0fe9b0703bdf5817beb0724d502", "sources": ["arxiv", "semantic_scholar"], "title": "Internally Rewarded Reinforcement Learning", "abstract": "We study a class of reinforcement learning problems where the reward signals for policy learning are generated by an internal reward model that is dependent on and jointly optimized with the policy. This interdependence between the policy and the reward model leads to an unstable learning process because reward signals from an immature reward model are noisy and impede policy learning, and conversely, an under-optimized policy impedes reward estimation learning. We call this learning setting $\\textit{Internally Rewarded Reinforcement Learning}$ (IRRL) as the reward is not provided directly by the environment but $\\textit{internally}$ by a reward model. In this paper, we formally formulate IRRL and present a class of problems that belong to IRRL. We theoretically derive and empirically analyze the effect of the reward function in IRRL and based on these analyses propose the clipped linear reward function. Experimental results show that the proposed reward function can consistently stabilize the training process by reducing the impact of reward noise, which leads to faster convergence and higher performance compared with baselines in diverse tasks.", "authors": ["Mengdi Li", "Xufeng Zhao", "Jae Hee Lee", "Cornelius Weber", "Stefan Wermter"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-02-01", "url": "https://arxiv.org/abs/2302.00270", "pdf_url": "https://arxiv.org/pdf/2302.00270v3", "arxiv_id": "2302.00270", "doi": "10.48550/arXiv.2302.00270", "citation_count": 17, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3138}
{"id": "4bcb0eaa21092322ee2e1f77c50d3d33259180611cef6480b10794edeb9587ee", "sources": ["arxiv", "semantic_scholar"], "title": "Direct Preference-based Policy Optimization without Reward Modeling", "abstract": "Preference-based reinforcement learning (PbRL) is an approach that enables RL agents to learn from preference, which is particularly useful when formulating a reward function is challenging. Existing PbRL methods generally involve a two-step procedure: they first learn a reward model based on given preference data and then employ off-the-shelf reinforcement learning algorithms using the learned reward model. However, obtaining an accurate reward model solely from preference information, especially when the preference is from human teachers, can be difficult. Instead, we propose a PbRL algorithm that directly learns from preference without requiring any reward modeling. To achieve this, we adopt a contrastive learning framework to design a novel policy scoring metric that assigns a high score to policies that align with the given preferences. We apply our algorithm to offline RL tasks with actual human preference labels and show that our algorithm outperforms or is on par with the existing PbRL methods. Notably, on high-dimensional control tasks, our algorithm surpasses offline RL methods that learn with ground-truth reward information. Finally, we show that our algorithm can be successfully applied to fine-tune large language models.", "authors": ["Gaon An", "Junhyeok Lee", "Xingdong Zuo", "Norio Kosaka", "Kyung-Min Kim", "Hyun Oh Song"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-30", "url": "https://arxiv.org/abs/2301.12842", "pdf_url": "https://arxiv.org/pdf/2301.12842v3", "arxiv_id": "2301.12842", "doi": "10.52202/075280-3078", "citation_count": 57, "influential_citation_count": 11, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5396}
{"id": "b860aec5fccfacaf8f8485e3ac26a23f15ab5832fb93752ece14a0b3a035ce98", "sources": ["arxiv", "semantic_scholar"], "title": "Automatic Intrinsic Reward Shaping for Exploration in Deep Reinforcement Learning", "abstract": "We present AIRS: Automatic Intrinsic Reward Shaping that intelligently and adaptively provides high-quality intrinsic rewards to enhance exploration in reinforcement learning (RL). More specifically, AIRS selects shaping function from a predefined set based on the estimated task return in real-time, providing reliable exploration incentives and alleviating the biased objective problem. Moreover, we develop an intrinsic reward toolkit to provide efficient and reliable implementations of diverse intrinsic reward approaches. We test AIRS on various tasks of MiniGrid, Procgen, and DeepMind Control Suite. Extensive simulation demonstrates that AIRS can outperform the benchmarking schemes and achieve superior performance with simple architecture.", "authors": ["Mingqi Yuan", "Bo Li", "Xin Jin", "Wenjun Zeng"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-26", "url": "https://arxiv.org/abs/2301.10886", "pdf_url": "https://arxiv.org/pdf/2301.10886v5", "arxiv_id": "2301.10886", "doi": "10.48550/arXiv.2301.10886", "citation_count": 18, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.3197}
{"id": "05c77f88b9be0c60d129864e3f01461885ce16a56883c3e35da86c95b49c46ec", "sources": ["arxiv", "semantic_scholar"], "title": "On The Fragility of Learned Reward Functions", "abstract": "Reward functions are notoriously difficult to specify, especially for tasks with complex goals. Reward learning approaches attempt to infer reward functions from human feedback and preferences. Prior works on reward learning have mainly focused on the performance of policies trained alongside the reward function. This practice, however, may fail to detect learned rewards that are not capable of training new policies from scratch and thus do not capture the intended behavior. Our work focuses on demonstrating and studying the causes of these relearning failures in the domain of preference-based reward learning. We demonstrate with experiments in tabular and continuous control environments that the severity of relearning failures can be sensitive to changes in reward model design and the trajectory dataset composition. Based on our findings, we emphasize the need for more retraining-based evaluations in the literature.", "authors": ["Lev McKinney", "Yawen Duan", "David Krueger", "Adam Gleave"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-09", "url": "https://arxiv.org/abs/2301.03652", "pdf_url": "https://arxiv.org/pdf/2301.03652v1", "arxiv_id": "2301.03652", "doi": "10.48550/arXiv.2301.03652", "citation_count": 23, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3451}
{"id": "d6e8c1a089b78774580fe9769bdd538646101f70dc5e0e11c027c82249830a23", "sources": ["arxiv", "semantic_scholar"], "title": "Exploration in Model-based Reinforcement Learning with Randomized Reward", "abstract": "Model-based Reinforcement Learning (MBRL) has been widely adapted due to its sample efficiency. However, existing worst-case regret analysis typically requires optimistic planning, which is not realistic in general. In contrast, motivated by the theory, empirical study utilizes ensemble of models, which achieve state-of-the-art performance on various testing environments. Such deviation between theory and empirical study leads us to question whether randomized model ensemble guarantee optimism, and hence the optimal worst-case regret? This paper partially answers such question from the perspective of reward randomization, a scarcely explored direction of exploration with MBRL. We show that under the kernelized linear regulator (KNR) model, reward randomization guarantees a partial optimism, which further yields a near-optimal worst-case regret in terms of the number of interactions. We further extend our theory to generalized function approximation and identified conditions for reward randomization to attain provably efficient exploration. Correspondingly, we propose concrete examples of efficient reward randomization. To the best of our knowledge, our analysis establishes the first worst-case regret analysis on randomized MBRL with function approximation.", "authors": ["Lingxiao Wang", "Ping Li"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2023-01-09", "url": "https://arxiv.org/abs/2301.03142", "pdf_url": "https://arxiv.org/pdf/2301.03142v1", "arxiv_id": "2301.03142", "doi": "10.48550/arXiv.2301.03142", "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0}
{"id": "dfa636e7046a6b491dfc86d9be19c3980a8de068726a64c30b4b87ddc50690e7", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Goal-Conditioned Policies Offline with Self-Supervised Reward Shaping", "abstract": "Developing agents that can execute multiple skills by learning from pre-collected datasets is an important problem in robotics, where online interaction with the environment is extremely time-consuming. Moreover, manually designing reward functions for every single desired skill is prohibitive. Prior works targeted these challenges by learning goal-conditioned policies from offline datasets without manually specified rewards, through hindsight relabelling. These methods suffer from the issue of sparsity of rewards, and fail at long-horizon tasks. In this work, we propose a novel self-supervised learning phase on the pre-collected dataset to understand the structure and the dynamics of the model, and shape a dense reward function for learning policies offline. We evaluate our method on three continuous control tasks, and show that our model significantly outperforms existing approaches, especially on tasks that involve long-term planning.", "authors": ["Lina Mezghani", "Sainbayar Sukhbaatar", "Piotr Bojanowski", "Alessandro Lazaric", "Karteek Alahari"], "categories": ["cs.RO", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-05", "url": "https://arxiv.org/abs/2301.02099", "pdf_url": "https://arxiv.org/pdf/2301.02099v1", "arxiv_id": "2301.02099", "doi": "10.48550/arXiv.2301.02099", "citation_count": 29, "influential_citation_count": 1, "has_code": true, "code_url": "https://github.com/facebookresearch/go-fresh", "venue": "Conference on Robot Learning", "quality_score": 0.3693}
{"id": "ac27d8a13f67787a58c30af31ea16dd8c6b7455d4b06e943e024a2d8cbd38bc5", "sources": ["arxiv", "semantic_scholar"], "title": "Benchmarks and Algorithms for Offline Preference-Based Reward Learning", "abstract": "Learning a reward function from human preferences is challenging as it typically requires having a high-fidelity simulator or using expensive and potentially unsafe actual physical rollouts in the environment. However, in many tasks the agent might have access to offline data from related tasks in the same target environment. While offline data is increasingly being used to aid policy optimization via offline RL, our observation is that it can be a surprisingly rich source of information for preference learning as well. We propose an approach that uses an offline dataset to craft preference queries via pool-based active learning, learns a distribution over reward functions, and optimizes a corresponding policy via offline RL. Crucially, our proposed approach does not require actual physical rollouts or an accurate simulator for either the reward learning or policy optimization steps. To test our approach, we first evaluate existing offline RL benchmarks for their suitability for offline reward learning. Surprisingly, for many offline RL domains, we find that simply using a trivial reward function results good policy performance, making these domains ill-suited for evaluating learned rewards. To address this, we identify a subset of existing offline RL benchmarks that are well suited for offline reward learning and also propose new offline apprenticeship learning benchmarks which allow for more open-ended behaviors. When evaluated on this curated set of domains, our empirical results suggest that combining offline RL with learned human preferences can enable an agent to learn to perform novel tasks that were not explicitly shown in the offline data.", "authors": ["Daniel Shin", "Anca D. Dragan", "Daniel S. Brown"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2023-01-03", "url": "https://arxiv.org/abs/2301.01392", "pdf_url": "https://arxiv.org/pdf/2301.01392v1", "arxiv_id": "2301.01392", "doi": "10.48550/arXiv.2301.01392", "citation_count": 78, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.4744}
{"id": "f6166aef1114fc9df3f6896747571591d8e2e8d03af9d06b2112eb9431ca97a6", "sources": ["arxiv", "semantic_scholar"], "title": "Federated Multi-Agent Deep Reinforcement Learning Approach via Physics-Informed Reward for Multi-Microgrid Energy Management", "abstract": "The utilization of large-scale distributed renewable energy promotes the development of the multi-microgrid (MMG), which raises the need of developing an effective energy management method to minimize economic costs and keep self energy-sufficiency. The multi-agent deep reinforcement learning (MADRL) has been widely used for the energy management problem because of its real-time scheduling ability. However, its training requires massive energy operation data of microgrids (MGs), while gathering these data from different MGs would threaten their privacy and data security. Therefore, this paper tackles this practical yet challenging issue by proposing a federated multi-agent deep reinforcement learning (F-MADRL) algorithm via the physics-informed reward. In this algorithm, the federated learning (FL) mechanism is introduced to train the F-MADRL algorithm thus ensures the privacy and the security of data. In addition, a decentralized MMG model is built, and the energy of each participated MG is managed by an agent, which aims to minimize economic costs and keep self energy-sufficiency according to the physics-informed reward. At first, MGs individually execute the self-training based on local energy operation data to train their local agent models. Then, these local models are periodically uploaded to a server and their parameters are aggregated to build a global agent, which will be broadcasted to MGs and replace their local agents. In this way, the experience of each MG agent can be shared and the energy operation data is not explicitly transmitted, thus protecting the privacy and ensuring data security. Finally, experiments are conducted on Oak Ridge national laboratory distributed energy control communication lab microgrid (ORNL-MG) test system, and the comparisons are carried out to verify the effectiveness of introducing the FL mechanism and the outperformance of our proposed F-MADRL.", "authors": ["Yuanzheng Li", "Shangyang He", "Yang Li", "Yang Shi", "Zhigang Zeng"], "categories": ["eess.SY", "cs.LG"], "fields_of_study": ["Computer Science", "Medicine", "Engineering"], "published_date": "2022-12-29", "url": "https://arxiv.org/abs/2301.00641", "pdf_url": "https://arxiv.org/pdf/2301.00641v1", "arxiv_id": "2301.00641", "doi": "10.1109/TNNLS.2022.3232630", "citation_count": 108, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Neural Networks and Learning Systems", "quality_score": 0.5094}
{"id": "0c9af8c4316dbcd625d4a3192f55697fa46cca52994af420310a5ba92b15ddd7", "sources": ["arxiv", "semantic_scholar"], "title": "Tiered Reward: Designing Rewards for Specification and Fast Learning of Desired Behavior", "abstract": "Reinforcement-learning agents seek to maximize a reward signal through environmental interactions. As humans, our job in the learning process is to design reward functions to express desired behavior and enable the agent to learn such behavior swiftly. However, designing good reward functions to induce the desired behavior is generally hard, let alone the question of which rewards make learning fast. In this work, we introduce a family of a reward structures we call Tiered Reward that addresses both of these questions. We consider the reward-design problem in tasks formulated as reaching desirable states and avoiding undesirable states. To start, we propose a strict partial ordering of the policy space to resolve trade-offs in behavior preference. We prefer policies that reach the good states faster and with higher probability while avoiding the bad states longer. Next, we introduce Tiered Reward, a class of environment-independent reward functions and show it is guaranteed to induce policies that are Pareto-optimal according to our preference relation. Finally, we demonstrate that Tiered Reward leads to fast learning with multiple tabular and deep reinforcement-learning algorithms.", "authors": ["Zhiyuan Zhou", "Shreyas Sundara Raman", "Henry Sowerby", "Michael L. Littman"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-12-07", "url": "https://arxiv.org/abs/2212.03733", "pdf_url": "https://arxiv.org/pdf/2212.03733v3", "arxiv_id": "2212.03733", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/zhouzypaul/tiered-reward", "venue": "Reinforcement Learning Journal, vol. 1, no. 1, 2024, pp. TBD", "quality_score": 0.0}
{"id": "e86628f75ddde7edcce8d88d96c266c4bfba1c7198716681c4edb9197f64e2c8", "sources": ["arxiv", "semantic_scholar"], "title": "Automatic Evaluation of Excavator Operators using Learned Reward Functions", "abstract": "Training novice users to operate an excavator for learning different skills requires the presence of expert teachers. Considering the complexity of the problem, it is comparatively expensive to find skilled experts as the process is time-consuming and requires precise focus. Moreover, since humans tend to be biased, the evaluation process is noisy and will lead to high variance in the final score of different operators with similar skills. In this work, we address these issues and propose a novel strategy for the automatic evaluation of excavator operators. We take into account the internal dynamics of the excavator and the safety criterion at every time step to evaluate the performance. To further validate our approach, we use this score prediction model as a source of reward for a reinforcement learning agent to learn the task of maneuvering an excavator in a simulated environment that closely replicates the real-world dynamics. For a policy learned using these external reward prediction models, our results demonstrate safer solutions following the required dynamic constraints when compared to policy trained with task-based reward functions only, making it one step closer to real-life adoption. For future research, we release our codebase at https://github.com/pranavAL/InvRL_Auto-Evaluate and video results https://drive.google.com/file/d/1jR1otOAu8zrY8mkhUOUZW9jkBOAKK71Z/view?usp=share_link .", "authors": ["Pranav Agarwal", "Marek Teichmann", "Sheldon Andrews", "Samira Ebrahimi Kahou"], "categories": ["cs.RO", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-15", "url": "https://arxiv.org/abs/2211.07941", "pdf_url": "https://arxiv.org/pdf/2211.07941v1", "arxiv_id": "2211.07941", "doi": "10.48550/arXiv.2211.07941", "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/pranavAL/InvRL_Auto-Evaluate", "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "deaf80efa4489436699fa796aa1fa5ce35d8fb3476442612b57d9e62d1a1540d", "sources": ["arxiv", "semantic_scholar"], "title": "Rewards Encoding Environment Dynamics Improves Preference-based Reinforcement Learning", "abstract": "Preference-based reinforcement learning (RL) algorithms help avoid the pitfalls of hand-crafted reward functions by distilling them from human preference feedback, but they remain impractical due to the burdensome number of labels required from the human, even for relatively simple tasks. In this work, we demonstrate that encoding environment dynamics in the reward function (REED) dramatically reduces the number of preference labels required in state-of-the-art preference-based RL frameworks. We hypothesize that REED-based methods better partition the state-action space and facilitate generalization to state-action pairs not included in the preference dataset. REED iterates between encoding environment dynamics in a state-action representation via a self-supervised temporal consistency task, and bootstrapping the preference-based reward function from the state-action representation. Whereas prior approaches train only on the preference-labelled trajectory pairs, REED exposes the state-action representation to all transitions experienced during policy training. We explore the benefits of REED within the PrefPPO [1] and PEBBLE [2] preference learning frameworks and demonstrate improvements across experimental conditions to both the speed of policy learning and the final policy performance. For example, on quadruped-walk and walker-walk with 50 preference labels, REED-based reward functions recover 83% and 66% of ground truth reward policy performance and without REED only 38\\% and 21\\% are recovered. For some domains, REED-based reward functions result in policies that outperform policies trained on the ground truth reward.", "authors": ["Katherine Metcalf", "Miguel Sarabia", "Barry-John Theobald"], "categories": ["cs.LG", "cs.AI", "cs.HC"], "fields_of_study": ["Computer Science"], "published_date": "2022-11-12", "url": "https://arxiv.org/abs/2211.06527", "pdf_url": "https://arxiv.org/pdf/2211.06527v1", "arxiv_id": "2211.06527", "doi": "10.48550/arXiv.2211.06527", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "9fd153b6d3f47f9003d219593427881419ec92a7cf96f28cd699cee7a2dfe2f1", "sources": ["arxiv", "semantic_scholar"], "title": "Scaling Laws for Reward Model Overoptimization", "abstract": "In reinforcement learning from human feedback, it is common to optimize against a reward model trained to predict human preferences. Because the reward model is an imperfect proxy, optimizing its value too much can hinder ground truth performance, in accordance with Goodhart's law. This effect has been frequently observed, but not carefully measured due to the expense of collecting human preference data. In this work, we use a synthetic setup in which a fixed \"gold-standard\" reward model plays the role of humans, providing labels used to train a proxy reward model. We study how the gold reward model score changes as we optimize against the proxy reward model using either reinforcement learning or best-of-$n$ sampling. We find that this relationship follows a different functional form depending on the method of optimization, and that in both cases its coefficients scale smoothly with the number of reward model parameters. We also study the effect on this relationship of the size of the reward model dataset, the number of reward model and policy parameters, and the coefficient of the KL penalty added to the reward in the reinforcement learning setup. We explore the implications of these empirical results for theoretical considerations in AI alignment.", "authors": ["Leo Gao", "John Schulman", "Jacob Hilton"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-10-19", "url": "https://arxiv.org/abs/2210.10760", "pdf_url": "https://arxiv.org/pdf/2210.10760v1", "arxiv_id": "2210.10760", "doi": null, "citation_count": 1058, "influential_citation_count": 103, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 1.0}
{"id": "54827696a3c765d6bc0a8526a2e258b4dd8c1e4fb3dae001220195619a03c88c", "sources": ["arxiv", "semantic_scholar"], "title": "Symbol Guided Hindsight Priors for Reward Learning from Human Preferences", "abstract": "Specifying rewards for reinforcement learned (RL) agents is challenging. Preference-based RL (PbRL) mitigates these challenges by inferring a reward from feedback over sets of trajectories. However, the effectiveness of PbRL is limited by the amount of feedback needed to reliably recover the structure of the target reward. We present the PRIor Over Rewards (PRIOR) framework, which incorporates priors about the structure of the reward function and the preference feedback into the reward learning process. Imposing these priors as soft constraints on the reward learning objective reduces the amount of feedback required by half and improves overall reward recovery. Additionally, we demonstrate that using an abstract state space for the computation of the priors further improves the reward learning and the agent's performance.", "authors": ["Mudit Verma", "Katherine Metcalf"], "categories": ["cs.LG", "cs.AI", "cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-17", "url": "https://arxiv.org/abs/2210.09151", "pdf_url": "https://arxiv.org/pdf/2210.09151v2", "arxiv_id": "2210.09151", "doi": "10.48550/arXiv.2210.09151", "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "fc9ee52dcc9515e859cdef07e3b1b0602c0c97dcf3eb37d3d5d26b784fba6ae6", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Learning with Trees: Methods and Evaluation", "abstract": "Recent efforts to learn reward functions from human feedback have tended to use deep neural networks, whose lack of transparency hampers our ability to explain agent behaviour or verify alignment. We explore the merits of learning intrinsically interpretable tree models instead. We develop a recently proposed method for learning reward trees from preference labels, and show it to be broadly competitive with neural networks on challenging high-dimensional tasks, with good robustness to limited or corrupted data. Having found that reward tree learning can be done effectively in complex settings, we then consider why it should be used, demonstrating that the interpretable reward structure gives significant scope for traceability, verification and explanation.", "authors": ["Tom Bewley", "Jonathan Lawry", "Arthur Richards", "Rachel Craddock", "Ian Henderson"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-10-03", "url": "https://arxiv.org/abs/2210.01007", "pdf_url": "https://arxiv.org/pdf/2210.01007v1", "arxiv_id": "2210.01007", "doi": "10.48550/arXiv.2210.01007", "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "5502f15001f403b7947b857065fe9924689d6f790b0ddc214d004586a3213114", "sources": ["arxiv", "semantic_scholar"], "title": "Argumentative Reward Learning: Reasoning About Human Preferences", "abstract": "We define a novel neuro-symbolic framework, argumentative reward learning, which combines preference-based argumentation with existing approaches to reinforcement learning from human feedback. Our method improves prior work by generalising human preferences, reducing the burden on the user and increasing the robustness of the reward model. We demonstrate this with a number of experiments.", "authors": ["Francis Rhys Ward", "Francesco Belardinelli", "Francesca Toni"], "categories": ["cs.AI", "cs.HC", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-28", "url": "https://arxiv.org/abs/2209.14010", "pdf_url": "https://arxiv.org/pdf/2209.14010v1", "arxiv_id": "2209.14010", "doi": "10.48550/arXiv.2209.14010", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "b63d569bf7ee4ed6b25dafa6abf2bbd810ef111116ac092dcd59fb3329197f2d", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Delay Attacks on Deep Reinforcement Learning", "abstract": "Most reinforcement learning algorithms implicitly assume strong synchrony. We present novel attacks targeting Q-learning that exploit a vulnerability entailed by this assumption by delaying the reward signal for a limited time period. We consider two types of attack goals: targeted attacks, which aim to cause a target policy to be learned, and untargeted attacks, which simply aim to induce a policy with a low reward. We evaluate the efficacy of the proposed attacks through a series of experiments. Our first observation is that reward-delay attacks are extremely effective when the goal is simply to minimize reward. Indeed, we find that even naive baseline reward-delay attacks are also highly successful in minimizing the reward. Targeted attacks, on the other hand, are more challenging, although we nevertheless demonstrate that the proposed approaches remain highly effective at achieving the attacker's targets. In addition, we introduce a second threat model that captures a minimal mitigation that ensures that rewards cannot be used out of sequence. We find that this mitigation remains insufficient to ensure robustness to attacks that delay, but preserve the order, of rewards.", "authors": ["Anindya Sarkar", "Jiarui Feng", "Yevgeniy Vorobeychik", "Christopher Gill", "Ning Zhang"], "categories": ["cs.LG", "cs.AI", "cs.CR"], "fields_of_study": ["Computer Science"], "published_date": "2022-09-08", "url": "https://arxiv.org/abs/2209.03540", "pdf_url": "https://arxiv.org/pdf/2209.03540v1", "arxiv_id": "2209.03540", "doi": "10.48550/arXiv.2209.03540", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Decision and Game Theory for Security", "quality_score": 0.2603}
{"id": "9f0ef7bcd7eb2d4747e1529c875e16d56ab60fa8a84142514285774811d9debd", "sources": ["arxiv", "semantic_scholar"], "title": "Automatic Reward Design via Learning Motivation-Consistent Intrinsic Rewards", "abstract": "Reward design is a critical part of the application of reinforcement learning, the performance of which strongly depends on how well the reward signal frames the goal of the designer and how well the signal assesses progress in reaching that goal. In many cases, the extrinsic rewards provided by the environment (e.g., win or loss of a game) are very sparse and make it difficult to train agents directly. Researchers usually assist the learning of agents by adding some auxiliary rewards in practice. However, designing auxiliary rewards is often turned to a trial-and-error search for reward settings that produces acceptable results. In this paper, we propose to automatically generate goal-consistent intrinsic rewards for the agent to learn, by maximizing which the expected accumulative extrinsic rewards can be maximized. To this end, we introduce the concept of motivation which captures the underlying goal of maximizing certain rewards and propose the motivation based reward design method. The basic idea is to shape the intrinsic rewards by minimizing the distance between the intrinsic and extrinsic motivations. We conduct extensive experiments and show that our method performs better than the state-of-the-art methods in handling problems of delayed reward, exploration, and credit assignment.", "authors": ["Yixiang Wang", "Yujing Hu", "Feng Wu", "Yingfeng Chen"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-07-29", "url": "https://arxiv.org/abs/2207.14722", "pdf_url": "https://arxiv.org/pdf/2207.14722v1", "arxiv_id": "2207.14722", "doi": "10.48550/arXiv.2207.14722", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1505}
{"id": "a57020d3e4ecdad7d5d9c26c3eb0b58cc13d46ef64c15399b2fee84103383c21", "sources": ["arxiv", "semantic_scholar"], "title": "Interactively Learning Preference Constraints in Linear Bandits", "abstract": "We study sequential decision-making with known rewards and unknown constraints, motivated by situations where the constraints represent expensive-to-evaluate human preferences, such as safe and comfortable driving behavior. We formalize the challenge of interactively learning about these constraints as a novel linear bandit problem which we call constrained linear best-arm identification. To solve this problem, we propose the Adaptive Constraint Learning (ACOL) algorithm. We provide an instance-dependent lower bound for constrained linear best-arm identification and show that ACOL's sample complexity matches the lower bound in the worst-case. In the average case, ACOL's sample complexity bound is still significantly tighter than bounds of simpler approaches. In synthetic experiments, ACOL performs on par with an oracle solution and outperforms a range of baselines. As an application, we consider learning constraints to represent human preferences in a driving simulation. ACOL is significantly more sample efficient than alternatives for this application. Further, we find that learning preferences as constraints is more robust to changes in the driving scenario than encoding the preferences directly in the reward function.", "authors": ["David Lindner", "Sebastian Tschiatschek", "Katja Hofmann", "Andreas Krause"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-06-10", "url": "https://arxiv.org/abs/2206.05255", "pdf_url": "https://arxiv.org/pdf/2206.05255v1", "arxiv_id": "2206.05255", "doi": "10.48550/arXiv.2206.05255", "citation_count": 15, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.301}
{"id": "d9ceb43162eb407ffe0f901585632a1e12389ce33d292a115dfd1f23e6a31ac3", "sources": ["arxiv", "semantic_scholar"], "title": "Models of human preference for learning reward functions", "abstract": "The utility of reinforcement learning is limited by the alignment of reward functions with the interests of human stakeholders. One promising method for alignment is to learn the reward function from human-generated preferences between pairs of trajectory segments, a type of reinforcement learning from human feedback (RLHF). These human preferences are typically assumed to be informed solely by partial return, the sum of rewards along each segment. We find this assumption to be flawed and propose modeling human preferences instead as informed by each segment's regret, a measure of a segment's deviation from optimal decision-making. Given infinitely many preferences generated according to regret, we prove that we can identify a reward function equivalent to the reward function that generated those preferences, and we prove that the previous partial return model lacks this identifiability property in multiple contexts. We empirically show that our proposed regret preference model outperforms the partial return preference model with finite training data in otherwise the same setting. Additionally, we find that our proposed regret preference model better predicts real human preferences and also learns reward functions from these preferences that lead to policies that are better human-aligned. Overall, this work establishes that the choice of preference model is impactful, and our proposed regret preference model provides an improvement upon a core assumption of recent research. We have open sourced our experimental code, the human preferences dataset we gathered, and our training and preference elicitation interfaces for gathering a such a dataset.", "authors": ["W. Bradley Knox", "Stephane Hatgis-Kessell", "Serena Booth", "Scott Niekum", "Peter Stone", "Alessandro Allievi"], "categories": ["cs.LG", "cs.AI", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2022-06-05", "url": "https://arxiv.org/abs/2206.02231", "pdf_url": "https://arxiv.org/pdf/2206.02231v3", "arxiv_id": "2206.02231", "doi": "10.48550/arXiv.2206.02231", "citation_count": 65, "influential_citation_count": 1, "has_code": true, "code_url": null, "venue": null, "quality_score": 0.4549}
{"id": "e6067985c3fe4a520f3049ae983667d5cab67e6e5cdaa2d0e13cf3942c7529ec", "sources": ["arxiv", "semantic_scholar"], "title": "Transferable Reward Learning by Dynamics-Agnostic Discriminator Ensemble", "abstract": "Recovering reward function from expert demonstrations is a fundamental problem in reinforcement learning. The recovered reward function captures the motivation of the expert. Agents can imitate experts by following these reward functions in their environment, which is known as apprentice learning. However, the agents may face environments different from the demonstrations, and therefore, desire transferable reward functions. Classical reward learning methods such as inverse reinforcement learning (IRL) or, equivalently, adversarial imitation learning (AIL), recover reward functions coupled with training dynamics, which are hard to be transferable. Previous dynamics-agnostic reward learning methods rely on assumptions such as that the reward function has to be state-only, restricting their applicability. In this work, we present a dynamics-agnostic discriminator-ensemble reward learning method (DARL) within the AIL framework, capable of learning both state-action and state-only reward functions. DARL achieves this by decoupling the reward function from training dynamics, employing a dynamics-agnostic discriminator on a latent space derived from the original state-action space. This latent space is optimized to minimize information on the dynamics. We moreover discover the policy-dependency issue of the AIL framework that reduces the transferability. DARL represents the reward function as an ensemble of discriminators during training to eliminate policy dependencies. Empirical studies on MuJoCo tasks with changed dynamics show that DARL better recovers the reward function and results in better imitation performance in transferred environments, handling both state-only and state-action reward scenarios.", "authors": ["Fan-Ming Luo", "Xingchen Cao", "Rong-Jun Qin", "Yang Yu"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-06-01", "url": "https://arxiv.org/abs/2206.00238", "pdf_url": "https://arxiv.org/pdf/2206.00238v2", "arxiv_id": "2206.00238", "doi": "10.48550/arXiv.2206.00238", "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1945}
{"id": "3edecdd0048370c07633af47df3b2ac1e6cf44260c265a2610ada8b7a22b2020", "sources": ["arxiv", "semantic_scholar"], "title": "Designing Rewards for Fast Learning", "abstract": "To convey desired behavior to a Reinforcement Learning (RL) agent, a designer must choose a reward function for the environment, arguably the most important knob designers have in interacting with RL agents. Although many reward functions induce the same optimal behavior (Ng et al., 1999), in practice, some of them result in faster learning than others. In this paper, we look at how reward-design choices impact learning speed and seek to identify principles of good reward design that quickly induce target behavior. This reward-identification problem is framed as an optimization problem: Firstly, we advocate choosing state-based rewards that maximize the action gap, making optimal actions easy to distinguish from suboptimal ones. Secondly, we propose minimizing a measure of the horizon, something we call the \"subjective discount\", over which rewards need to be optimized to encourage agents to make optimal decisions with less lookahead. To solve this optimization problem, we propose a linear-programming based algorithm that efficiently finds a reward function that maximizes action gap and minimizes subjective discount. We test the rewards generated with the algorithm in tabular environments with Q-Learning, and empirically show they lead to faster learning. Although we only focus on Q-Learning because it is perhaps the simplest and most well understood RL algorithm, preliminary results with R-max (Brafman and Tennenholtz, 2000) suggest our results are much more general. Our experiments support three principles of reward design: 1) consistent with existing results, penalizing each step taken induces faster learning than rewarding the goal. 2) When rewarding subgoals along the target trajectory, rewards should gradually increase as the goal gets closer. 3) Dense reward that's nonzero on every state is only good if designed carefully.", "authors": ["Henry Sowerby", "Zhiyuan Zhou", "Michael L. Littman"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-05-30", "url": "https://arxiv.org/abs/2205.15400", "pdf_url": "https://arxiv.org/pdf/2205.15400v1", "arxiv_id": "2205.15400", "doi": "10.48550/arXiv.2205.15400", "citation_count": 18, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3197}
{"id": "90c00392964b8ff7295346791322f3f0283367ccbf194db8ef948c742a66df20", "sources": ["arxiv"], "title": "Reward Uncertainty for Exploration in Preference-based Reinforcement Learning", "abstract": "Conveying complex objectives to reinforcement learning (RL) agents often requires meticulous reward engineering. Preference-based RL methods are able to learn a more flexible reward model based on human preferences by actively incorporating human feedback, i.e. teacher's preferences between two clips of behaviors. However, poor feedback-efficiency still remains a problem in current preference-based RL algorithms, as tailored human feedback is very expensive. To handle this issue, previous methods have mainly focused on improving query selection and policy initialization. At the same time, recent exploration methods have proven to be a recipe for improving sample-efficiency in RL. We present an exploration method specifically for preference-based RL algorithms. Our main idea is to design an intrinsic reward by measuring the novelty based on learned reward. Specifically, we utilize disagreement across ensemble of learned reward models. Our intuition is that disagreement in learned reward model reflects uncertainty in tailored human feedback and could be useful for exploration. Our experiments show that exploration bonus from uncertainty in learned reward improves both feedback- and sample-efficiency of preference-based RL algorithms on complex robot manipulation tasks from MetaWorld benchmarks, compared with other existing exploration methods that measure the novelty of state visitation.", "authors": ["Xinran Liang", "Katherine Shu", "Kimin Lee", "Pieter Abbeel"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": [], "published_date": "2022-05-24", "url": "https://arxiv.org/abs/2205.12401", "pdf_url": "https://arxiv.org/pdf/2205.12401v1", "arxiv_id": "2205.12401", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "ac2ebeae2a1a728989b4c1e0df4e193e2d0d7f2e71ce85f7c042f5831e2ecf3f", "sources": ["arxiv", "semantic_scholar"], "title": "Causal Confusion and Reward Misidentification in Preference-Based Reward Learning", "abstract": "Learning policies via preference-based reward learning is an increasingly popular method for customizing agent behavior, but has been shown anecdotally to be prone to spurious correlations and reward hacking behaviors. While much prior work focuses on causal confusion in reinforcement learning and behavioral cloning, we focus on a systematic study of causal confusion and reward misidentification when learning from preferences. In particular, we perform a series of sensitivity and ablation analyses on several benchmark domains where rewards learned from preferences achieve minimal test error but fail to generalize to out-of-distribution states -- resulting in poor policy performance when optimized. We find that the presence of non-causal distractor features, noise in the stated preferences, and partial state observability can all exacerbate reward misidentification. We also identify a set of methods with which to interpret misidentified learned rewards. In general, we observe that optimizing misidentified rewards drives the policy off the reward's training distribution, resulting in high predicted (learned) rewards but low true rewards. These findings illuminate the susceptibility of preference learning to reward misidentification and causal confusion -- failure to consider even one of many factors can result in unexpected, undesirable behavior.", "authors": ["Jeremy Tien", "Jerry Zhi-Yang He", "Zackory Erickson", "Anca D. Dragan", "Daniel S. Brown"], "categories": ["cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2022-04-13", "url": "https://arxiv.org/abs/2204.06601", "pdf_url": "https://arxiv.org/pdf/2204.06601v4", "arxiv_id": "2204.06601", "doi": null, "citation_count": 59, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.4445}
{"id": "29f167f4a5669c9194c28e7a2ddb9cc12d557861943b5c0898ff8e4324f979f8", "sources": ["arxiv", "semantic_scholar"], "title": "Preprocessing Reward Functions for Interpretability", "abstract": "In many real-world applications, the reward function is too complex to be manually specified. In such cases, reward functions must instead be learned from human feedback. Since the learned reward may fail to represent user preferences, it is important to be able to validate the learned reward function prior to deployment. One promising approach is to apply interpretability tools to the reward function to spot potential deviations from the user's intention. Existing work has applied general-purpose interpretability tools to understand learned reward functions. We propose exploiting the intrinsic structure of reward functions by first preprocessing them into simpler but equivalent reward functions, which are then visualized. We introduce a general framework for such reward preprocessing and propose concrete preprocessing algorithms. Our empirical evaluation shows that preprocessed rewards are often significantly easier to understand than the original reward.", "authors": ["Erik Jenner", "Adam Gleave"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2022-03-25", "url": "https://arxiv.org/abs/2203.13553", "pdf_url": "https://arxiv.org/pdf/2203.13553v1", "arxiv_id": "2203.13553", "doi": "10.48550/arXiv.2203.13553", "citation_count": 8, "influential_citation_count": 2, "has_code": true, "code_url": "https://github.com/HumanCompatibleAI/reward-preprocessing", "venue": "arXiv.org", "quality_score": 0.2386}
{"id": "0ed1bef9acc6544b4cb85fe78a5d7e50df2a16a6ca466b1e8ddbedc9ba729326", "sources": ["arxiv"], "title": "SURF: Semi-supervised Reward Learning with Data Augmentation for Feedback-efficient Preference-based Reinforcement Learning", "abstract": "Preference-based reinforcement learning (RL) has shown potential for teaching agents to perform the target tasks without a costly, pre-defined reward function by learning the reward with a supervisor's preference between the two agent behaviors. However, preference-based learning often requires a large amount of human feedback, making it difficult to apply this approach to various applications. This data-efficiency problem, on the other hand, has been typically addressed by using unlabeled samples or data augmentation techniques in the context of supervised learning. Motivated by the recent success of these approaches, we present SURF, a semi-supervised reward learning framework that utilizes a large amount of unlabeled samples with data augmentation. In order to leverage unlabeled samples for reward learning, we infer pseudo-labels of the unlabeled samples based on the confidence of the preference predictor. To further improve the label-efficiency of reward learning, we introduce a new data augmentation that temporally crops consecutive subsequences from the original behaviors. Our experiments demonstrate that our approach significantly improves the feedback-efficiency of the state-of-the-art preference-based method on a variety of locomotion and robotic manipulation tasks.", "authors": ["Jongjin Park", "Younggyo Seo", "Jinwoo Shin", "Honglak Lee", "Pieter Abbeel", "Kimin Lee"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": [], "published_date": "2022-03-18", "url": "https://arxiv.org/abs/2203.10050", "pdf_url": "https://arxiv.org/pdf/2203.10050v1", "arxiv_id": "2203.10050", "doi": null, "citation_count": 0, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.0}
{"id": "202f3419146998a282f938147fedf9de160781d0bec2a453ac1aa3d98360964d", "sources": ["arxiv", "semantic_scholar"], "title": "Invariance in Policy Optimisation and Partial Identifiability in Reward Learning", "abstract": "It is often very challenging to manually design reward functions for complex, real-world tasks. To solve this, one can instead use reward learning to infer a reward function from data. However, there are often multiple reward functions that fit the data equally well, even in the infinite-data limit. This means that the reward function is only partially identifiable. In this work, we formally characterise the partial identifiability of the reward function given several popular reward learning data sources, including expert demonstrations and trajectory comparisons. We also analyse the impact of this partial identifiability for several downstream tasks, such as policy optimisation. We unify our results in a framework for comparing data sources and downstream tasks by their invariances, with implications for the design and selection of data sources for reward learning.", "authors": ["Joar Skalse", "Matthew Farrugia-Roberts", "Stuart Russell", "Alessandro Abate", "Adam Gleave"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-03-14", "url": "https://arxiv.org/abs/2203.07475", "pdf_url": "https://arxiv.org/pdf/2203.07475v2", "arxiv_id": "2203.07475", "doi": "10.48550/arXiv.2203.07475", "citation_count": 59, "influential_citation_count": 5, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4445}
{"id": "ab786f3d38398f42204553ecde53b39f49c5e73acbb0cd48e8aa1db1861f8d72", "sources": ["arxiv", "semantic_scholar"], "title": "Inferring Lexicographically-Ordered Rewards from Preferences", "abstract": "Modeling the preferences of agents over a set of alternatives is a principal concern in many areas. The dominant approach has been to find a single reward/utility function with the property that alternatives yielding higher rewards are preferred over alternatives yielding lower rewards. However, in many settings, preferences are based on multiple, often competing, objectives; a single reward function is not adequate to represent such preferences. This paper proposes a method for inferring multi-objective reward-based representations of an agent's observed preferences. We model the agent's priorities over different objectives as entering lexicographically, so that objectives with lower priorities matter only when the agent is indifferent with respect to objectives with higher priorities. We offer two example applications in healthcare, one inspired by cancer treatment, the other inspired by organ transplantation, to illustrate how the lexicographically-ordered rewards we learn can provide a better understanding of a decision-maker's preferences and help improve policies when used in reinforcement learning.", "authors": ["Alihan Hüyük", "William R. Zame", "Mihaela van der Schaar"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-02-21", "url": "https://arxiv.org/abs/2202.10153", "pdf_url": "https://arxiv.org/pdf/2202.10153v2", "arxiv_id": "2202.10153", "doi": "10.1609/aaai.v36i5.20516", "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.2113}
{"id": "90fef9f237a6488af4422ed81b3ca99db78852030a550e5c60374658f0547368", "sources": ["arxiv", "semantic_scholar"], "title": "Choices, Risks, and Reward Reports: Charting Public Policy for Reinforcement Learning Systems", "abstract": "In the long term, reinforcement learning (RL) is considered by many AI theorists to be the most promising path to artificial general intelligence. This places RL practitioners in a position to design systems that have never existed before and lack prior documentation in law and policy. Public agencies could intervene on complex dynamics that were previously too opaque to deliberate about, and long-held policy ambitions would finally be made tractable. In this whitepaper we illustrate this potential and how it might be technically enacted in the domains of energy infrastructure, social media recommender systems, and transportation. Alongside these unprecedented interventions come new forms of risk that exacerbate the harms already generated by standard machine learning tools. We correspondingly present a new typology of risks arising from RL design choices, falling under four categories: scoping the horizon, defining rewards, pruning information, and training multiple agents. Rather than allowing RL systems to unilaterally reshape human domains, policymakers need new mechanisms for the rule of reason, foreseeability, and interoperability that match the risks these systems pose. We argue that criteria for these choices may be drawn from emerging subfields within antitrust, tort, and administrative law. It will then be possible for courts, federal and state agencies, and non-governmental organizations to play more active roles in RL specification and evaluation. Building on the \"model cards\" and \"datasheets\" frameworks proposed by Mitchell et al. and Gebru et al., we argue the need for Reward Reports for AI systems. Reward Reports are living documents for proposed RL deployments that demarcate design choices.", "authors": ["Thomas Krendl Gilbert", "Sarah Dean", "Tom Zick", "Nathan Lambert"], "categories": ["cs.LG", "cs.CY"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-11", "url": "https://arxiv.org/abs/2202.05716", "pdf_url": "https://arxiv.org/pdf/2202.05716v1", "arxiv_id": "2202.05716", "doi": null, "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.294}
{"id": "a13408d28dfe581fd0292a234b0efbf2e0a4841b09015047a99db50874fb50d7", "sources": ["arxiv", "semantic_scholar"], "title": "Reward-Respecting Subtasks for Model-Based Reinforcement Learning", "abstract": "To achieve the ambitious goals of artificial intelligence, reinforcement learning must include planning with a model of the world that is abstract in state and time. Deep learning has made progress with state abstraction, but temporal abstraction has rarely been used, despite extensively developed theory based on the options framework. One reason for this is that the space of possible options is immense, and the methods previously proposed for option discovery do not take into account how the option models will be used in planning. Options are typically discovered by posing subsidiary tasks, such as reaching a bottleneck state or maximizing the cumulative sum of a sensory signal other than reward. Each subtask is solved to produce an option, and then a model of the option is learned and made available to the planning process. In most previous work, the subtasks ignore the reward on the original problem, whereas we propose subtasks that use the original reward plus a bonus based on a feature of the state at the time the option terminates. We show that option models obtained from such reward-respecting subtasks are much more likely to be useful in planning than eigenoptions, shortest path options based on bottleneck states, or reward-respecting options generated by the option-critic. Reward respecting subtasks strongly constrain the space of options and thereby also provide a partial solution to the problem of option discovery. Finally, we show how values, policies, options, and models can all be learned online and off-policy using standard algorithms and general value functions.", "authors": ["Richard S. Sutton", "Marlos C. Machado", "G. Zacharias Holland", "David Szepesvari", "Finbarr Timbers", "Brian Tanner", "Adam White"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-07", "url": "https://arxiv.org/abs/2202.03466", "pdf_url": "https://arxiv.org/pdf/2202.03466v4", "arxiv_id": "2202.03466", "doi": "10.1016/j.artint.2023.104001", "citation_count": 30, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Artificial Intelligence", "quality_score": 0.3728}
{"id": "f82ea56039b0ae2e9e354fc1dc24f631e211e5d6a18a1cb0fab34dbee01c618e", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Synthetic Environments and Reward Networks for Reinforcement Learning", "abstract": "We introduce Synthetic Environments (SEs) and Reward Networks (RNs), represented by neural networks, as proxy environment models for training Reinforcement Learning (RL) agents. We show that an agent, after being trained exclusively on the SE, is able to solve the corresponding real environment. While an SE acts as a full proxy to a real environment by learning about its state dynamics and rewards, an RN is a partial proxy that learns to augment or replace rewards. We use bi-level optimization to evolve SEs and RNs: the inner loop trains the RL agent, and the outer loop trains the parameters of the SE / RN via an evolution strategy. We evaluate our proposed new concept on a broad range of RL algorithms and classic control environments. In a one-to-one comparison, learning an SE proxy requires more interactions with the real environment than training agents only on the real environment. However, once such an SE has been learned, we do not need any interactions with the real environment to train new agents. Moreover, the learned SE proxies allow us to train agents with fewer interactions while maintaining the original task performance. Our empirical results suggest that SEs achieve this result by learning informed representations that bias the agents towards relevant states. Moreover, we find that these proxies are robust against hyperparameter variation and can also transfer to unseen agents.", "authors": ["Fabio Ferreira", "Thomas Nierhoff", "Andreas Saelinger", "Frank Hutter"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-02-06", "url": "https://arxiv.org/abs/2202.02790", "pdf_url": "https://arxiv.org/pdf/2202.02790v1", "arxiv_id": "2202.02790", "doi": null, "citation_count": 6, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.2113}
{"id": "6617933c177bbeb99b6907c2ede475edcc0a57c185d9b546136b0e16728c813f", "sources": ["arxiv", "semantic_scholar"], "title": "Dynamics-Aware Comparison of Learned Reward Functions", "abstract": "The ability to learn reward functions plays an important role in enabling the deployment of intelligent agents in the real world. However, comparing reward functions, for example as a means of evaluating reward learning methods, presents a challenge. Reward functions are typically compared by considering the behavior of optimized policies, but this approach conflates deficiencies in the reward function with those of the policy search algorithm used to optimize it. To address this challenge, Gleave et al. (2020) propose the Equivalent-Policy Invariant Comparison (EPIC) distance. EPIC avoids policy optimization, but in doing so requires computing reward values at transitions that may be impossible under the system dynamics. This is problematic for learned reward functions because it entails evaluating them outside of their training distribution, resulting in inaccurate reward values that we show can render EPIC ineffective at comparing rewards. To address this problem, we propose the Dynamics-Aware Reward Distance (DARD), a new reward pseudometric. DARD uses an approximate transition model of the environment to transform reward functions into a form that allows for comparisons that are invariant to reward shaping while only evaluating reward functions on transitions close to their training distribution. Experiments in simulated physical domains demonstrate that DARD enables reliable reward comparisons without policy optimization and is significantly more predictive than baseline methods of downstream policy performance when dealing with learned reward functions.", "authors": ["Blake Wulfe", "Ashwin Balakrishna", "Logan Ellis", "Jean Mercat", "Rowan McAllister", "Adrien Gaidon"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2022-01-25", "url": "https://arxiv.org/abs/2201.10081", "pdf_url": "https://arxiv.org/pdf/2201.10081v1", "arxiv_id": "2201.10081", "doi": null, "citation_count": 19, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.3253}
{"id": "76cadb28aa5e078f9ad9a8838eb86b321f657d1d7895ce9e4a693aea90be4dc6", "sources": ["arxiv", "semantic_scholar"], "title": "Operator Deep Q-Learning: Zero-Shot Reward Transferring in Reinforcement Learning", "abstract": "Reinforcement learning (RL) has drawn increasing interests in recent years due to its tremendous success in various applications. However, standard RL algorithms can only be applied for single reward function, and cannot adapt to an unseen reward function quickly. In this paper, we advocate a general operator view of reinforcement learning, which enables us to directly approximate the operator that maps from reward function to value function. The benefit of learning the operator is that we can incorporate any new reward function as input and attain its corresponding value function in a zero-shot manner. To approximate this special type of operator, we design a number of novel operator neural network architectures based on its theoretical properties. Our design of operator networks outperform the existing methods and the standard design of general purpose operator network, and we demonstrate the benefit of our operator deep Q-learning framework in several tasks including reward transferring for offline policy evaluation (OPE) and reward transferring for offline policy optimization in a range of tasks.", "authors": ["Ziyang Tang", "Yihao Feng", "Qiang Liu"], "categories": ["cs.LG", "cs.RO", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2022-01-01", "url": "https://arxiv.org/abs/2201.00236", "pdf_url": "https://arxiv.org/pdf/2201.00236v1", "arxiv_id": "2201.00236", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "1c917eecdd329a4cde4de8c3557803add30ea3af34a1a6f37f9f8d7b8a49a026", "sources": ["arxiv", "semantic_scholar"], "title": "Interpretable Preference-based Reinforcement Learning with Tree-Structured Reward Functions", "abstract": "The potential of reinforcement learning (RL) to deliver aligned and performant agents is partially bottlenecked by the reward engineering problem. One alternative to heuristic trial-and-error is preference-based RL (PbRL), where a reward function is inferred from sparse human feedback. However, prior PbRL methods lack interpretability of the learned reward structure, which hampers the ability to assess robustness and alignment. We propose an online, active preference learning algorithm that constructs reward functions with the intrinsically interpretable, compositional structure of a tree. Using both synthetic and human-provided feedback, we demonstrate sample-efficient learning of tree-structured reward functions in several environments, then harness the enhanced interpretability to explore and debug for alignment.", "authors": ["Tom Bewley", "Freddy Lecue"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2021-12-20", "url": "https://arxiv.org/abs/2112.11230", "pdf_url": "https://arxiv.org/pdf/2112.11230v1", "arxiv_id": "2112.11230", "doi": "10.5555/3535850.3535865", "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Adaptive Agents and Multi-Agent Systems", "quality_score": 0.294}
{"id": "8a7b7128403f23b092387236490ea013b095ee1e5dc8fe1436f633e3fefb674a", "sources": ["arxiv", "semantic_scholar"], "title": "Reward-Free Attacks in Multi-Agent Reinforcement Learning", "abstract": "We investigate how effective an attacker can be when it only learns from its victim's actions, without access to the victim's reward. In this work, we are motivated by the scenario where the attacker wants to behave strategically when the victim's motivations are unknown. We argue that one heuristic approach an attacker can use is to maximize the entropy of the victim's policy. The policy is generally not obfuscated, which implies it may be extracted simply by passively observing the victim. We provide such a strategy in the form of a reward-free exploration algorithm that maximizes the attacker's entropy during the exploration phase, and then maximizes the victim's empirical entropy during the planning phase. In our experiments, the victim agents are subverted through policy entropy maximization, implying an attacker might not need access to the victim's reward to succeed. Hence, reward-free attacks, which are based only on observing behavior, show the feasibility of an attacker to act strategically without knowledge of the victim's motives even if the victim's reward information is protected.", "authors": ["Ted Fujimoto", "Timothy Doster", "Adam Attarian", "Jill Brandenberger", "Nathan Hodas"], "categories": ["cs.LG", "cs.AI", "cs.MA"], "fields_of_study": ["Computer Science"], "published_date": "2021-12-02", "url": "https://arxiv.org/abs/2112.00940", "pdf_url": "https://arxiv.org/pdf/2112.00940v1", "arxiv_id": "2112.00940", "doi": null, "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "5c1f863b65c6422ebe88f13aebbbb76c83eb68a9cc91bfe5d0adeef9709ea697", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Long-Term Reward Redistribution via Randomized Return Decomposition", "abstract": "Many practical applications of reinforcement learning require agents to learn from sparse and delayed rewards. It challenges the ability of agents to attribute their actions to future outcomes. In this paper, we consider the problem formulation of episodic reinforcement learning with trajectory feedback. It refers to an extreme delay of reward signals, in which the agent can only obtain one reward signal at the end of each trajectory. A popular paradigm for this problem setting is learning with a designed auxiliary dense reward function, namely proxy reward, instead of sparse environmental signals. Based on this framework, this paper proposes a novel reward redistribution algorithm, randomized return decomposition (RRD), to learn a proxy reward function for episodic reinforcement learning. We establish a surrogate problem by Monte-Carlo sampling that scales up least-squares-based reward redistribution to long-horizon problems. We analyze our surrogate loss function by connection with existing methods in the literature, which illustrates the algorithmic properties of our approach. In experiments, we extensively evaluate our proposed method on a variety of benchmark tasks with episodic rewards and demonstrate substantial improvement over baseline algorithms.", "authors": ["Zhizhou Ren", "Ruihan Guo", "Yuan Zhou", "Jian Peng"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2021-11-26", "url": "https://arxiv.org/abs/2111.13485", "pdf_url": "https://arxiv.org/pdf/2111.13485v2", "arxiv_id": "2111.13485", "doi": null, "citation_count": 46, "influential_citation_count": 14, "has_code": false, "code_url": null, "venue": "International Conference on Learning Representations", "quality_score": 0.588}
{"id": "249ff650869cc49332d0d8ce03a04eec8d6eb78b0bb25d099e3f40ff117ad59b", "sources": ["arxiv", "semantic_scholar"], "title": "Versatile Inverse Reinforcement Learning via Cumulative Rewards", "abstract": "Inverse Reinforcement Learning infers a reward function from expert demonstrations, aiming to encode the behavior and intentions of the expert. Current approaches usually do this with generative and uni-modal models, meaning that they encode a single behavior. In the common setting, where there are various solutions to a problem and the experts show versatile behavior this severely limits the generalization capabilities of these methods. We propose a novel method for Inverse Reinforcement Learning that overcomes these problems by formulating the recovered reward as a sum of iteratively trained discriminators. We show on simulated tasks that our approach is able to recover general, high-quality reward functions and produces policies of the same quality as behavioral cloning approaches designed for versatile behavior.", "authors": ["Niklas Freymuth", "Philipp Becker", "Gerhard Neumann"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2021-11-15", "url": "https://arxiv.org/abs/2111.07667", "pdf_url": "https://arxiv.org/pdf/2111.07667v1", "arxiv_id": "2111.07667", "doi": null, "citation_count": 4, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1747}
{"id": "c1a2cb34ad86ce73ee379553f41c057498f42f51ce7bf04019c77c72f9b86c21", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Multimodal Rewards from Rankings", "abstract": "Learning from human feedback has shown to be a useful approach in acquiring robot reward functions. However, expert feedback is often assumed to be drawn from an underlying unimodal reward function. This assumption does not always hold including in settings where multiple experts provide data or when a single expert provides data for different tasks -- we thus go beyond learning a unimodal reward and focus on learning a multimodal reward function. We formulate the multimodal reward learning as a mixture learning problem and develop a novel ranking-based learning approach, where the experts are only required to rank a given set of trajectories. Furthermore, as access to interaction data is often expensive in robotics, we develop an active querying approach to accelerate the learning process. We conduct experiments and user studies using a multi-task variant of OpenAI's LunarLander and a real Fetch robot, where we collect data from multiple users with different preferences. The results suggest that our approach can efficiently learn multimodal reward functions, and improve data-efficiency over benchmark methods that we adapt to our learning problem.", "authors": ["Vivek Myers", "Erdem Bıyık", "Nima Anari", "Dorsa Sadigh"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2021-09-27", "url": "https://arxiv.org/abs/2109.12750", "pdf_url": "https://arxiv.org/pdf/2109.12750v2", "arxiv_id": "2109.12750", "doi": null, "citation_count": 66, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Conference on Robot Learning", "quality_score": 0.4565}
{"id": "b6ea7561d9b5f3e650fe3b0f5fefb31a93c468086f67436e4981a3d86e929168", "sources": ["arxiv", "semantic_scholar"], "title": "Knowledge is reward: Learning optimal exploration by predictive reward cashing", "abstract": "There is a strong link between the general concept of intelligence and the ability to collect and use information. The theory of Bayes-adaptive exploration offers an attractive optimality framework for training machines to perform complex information gathering tasks. However, the computational complexity of the resulting optimal control problem has limited the diffusion of the theory to mainstream deep AI research. In this paper we exploit the inherent mathematical structure of Bayes-adaptive problems in order to dramatically simplify the problem by making the reward structure denser while simultaneously decoupling the learning of exploitation and exploration policies. The key to this simplification comes from the novel concept of cross-value (i.e. the value of being in an environment while acting optimally according to another), which we use to quantify the value of currently available information. This results in a new denser reward structure that \"cashes in\" all future rewards that can be predicted from the current information state. In a set of experiments we show that the approach makes it possible to learn challenging information gathering tasks without the use of shaping and heuristic bonuses in situations where the standard RL algorithms fail.", "authors": ["Luca Ambrogioni"], "categories": ["stat.ML", "cs.LG"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2021-09-17", "url": "https://arxiv.org/abs/2109.08518", "pdf_url": "https://arxiv.org/pdf/2109.08518v1", "arxiv_id": "2109.08518", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}
{"id": "4dd9812b0706bce4e7531645a83856bd3153fba8406f55eb51d95396630e9fd6", "sources": ["arxiv", "semantic_scholar"], "title": "APReL: A Library for Active Preference-based Reward Learning Algorithms", "abstract": "Reward learning is a fundamental problem in human-robot interaction to have robots that operate in alignment with what their human user wants. Many preference-based learning algorithms and active querying techniques have been proposed as a solution to this problem. In this paper, we present APReL, a library for active preference-based reward learning algorithms, which enable researchers and practitioners to experiment with the existing techniques and easily develop their own algorithms for various modules of the problem. APReL is available at https://github.com/Stanford-ILIAD/APReL.", "authors": ["Erdem Bıyık", "Aditi Talati", "Dorsa Sadigh"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2021-08-16", "url": "https://arxiv.org/abs/2108.07259", "pdf_url": "https://arxiv.org/pdf/2108.07259v2", "arxiv_id": "2108.07259", "doi": "10.1109/HRI53351.2022.9889650", "citation_count": 40, "influential_citation_count": 4, "has_code": true, "code_url": "https://github.com/Stanford-ILIAD/APReL", "venue": "IEEE/ACM International Conference on Human-Robot Interaction", "quality_score": 0.4032}
{"id": "83f0743426cf218d07d88f835761d2d898a217dbefa77e06dda859e4e8357650", "sources": ["arxiv", "semantic_scholar"], "title": "Reward-Weighted Regression Converges to a Global Optimum", "abstract": "Reward-Weighted Regression (RWR) belongs to a family of widely known iterative Reinforcement Learning algorithms based on the Expectation-Maximization framework. In this family, learning at each iteration consists of sampling a batch of trajectories using the current policy and fitting a new policy to maximize a return-weighted log-likelihood of actions. Although RWR is known to yield monotonic improvement of the policy under certain circumstances, whether and under which conditions RWR converges to the optimal policy have remained open questions. In this paper, we provide for the first time a proof that RWR converges to a global optimum when no function approximation is used, in a general compact setting. Furthermore, for the simpler case with finite state and action spaces we prove R-linear convergence of the state-value function to the optimum.", "authors": ["Miroslav Štrupl", "Francesco Faccio", "Dylan R. Ashley", "Rupesh Kumar Srivastava", "Jürgen Schmidhuber"], "categories": ["stat.ML", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2021-07-19", "url": "https://arxiv.org/abs/2107.09088", "pdf_url": "https://arxiv.org/pdf/2107.09088v3", "arxiv_id": "2107.09088", "doi": "10.1609/aaai.v36i8.20811", "citation_count": 5, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/dylanashley/reward-weighted-regression", "venue": "AAAI Conference on Artificial Intelligence", "quality_score": 0.1945}
{"id": "b2957bce7c1494dc72678b4387e668713a09a1be94035fd1548191f8fa5bccf7", "sources": ["arxiv", "semantic_scholar"], "title": "Inferring Probabilistic Reward Machines from Non-Markovian Reward Processes for Reinforcement Learning", "abstract": "The success of reinforcement learning in typical settings is predicated on Markovian assumptions on the reward signal by which an agent learns optimal policies. In recent years, the use of reward machines has relaxed this assumption by enabling a structured representation of non-Markovian rewards. In particular, such representations can be used to augment the state space of the underlying decision process, thereby facilitating non-Markovian reinforcement learning. However, these reward machines cannot capture the semantics of stochastic reward signals. In this paper, we make progress on this front by introducing probabilistic reward machines (PRMs) as a representation of non-Markovian stochastic rewards. We present an algorithm to learn PRMs from the underlying decision process and prove results around its correctness and convergence.", "authors": ["Taylor Dohmen", "Noah Topper", "George Atia", "Andre Beckus", "Ashutosh Trivedi", "Alvaro Velasquez"], "categories": ["cs.LG", "cs.FL", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2021-07-09", "url": "https://arxiv.org/abs/2107.04633", "pdf_url": "https://arxiv.org/pdf/2107.04633v2", "arxiv_id": "2107.04633", "doi": "10.1609/icaps.v32i1.19844", "citation_count": 21, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "International Conference on Automated Planning and Scheduling", "quality_score": 0.3356}
{"id": "e5e1233785e03ccf781ea7df39116e813ee20f0721964f87b1113249325fb548", "sources": ["arxiv", "semantic_scholar"], "title": "Learning from an Exploring Demonstrator: Optimal Reward Estimation for Bandits", "abstract": "We introduce the \"inverse bandit\" problem of estimating the rewards of a multi-armed bandit instance from observing the learning process of a low-regret demonstrator. Existing approaches to the related problem of inverse reinforcement learning assume the execution of an optimal policy, and thereby suffer from an identifiability issue. In contrast, we propose to leverage the demonstrator's behavior en route to optimality, and in particular, the exploration phase, for reward estimation. We begin by establishing a general information-theoretic lower bound under this paradigm that applies to any demonstrator algorithm, which characterizes a fundamental tradeoff between reward estimation and the amount of exploration of the demonstrator. Then, we develop simple and efficient reward estimators for upper-confidence-based demonstrator algorithms that attain the optimal tradeoff, showing in particular that consistent reward estimation -- free of identifiability issues -- is possible under our paradigm. Extensive simulations on both synthetic and semi-synthetic data corroborate our theoretical results.", "authors": ["Wenshuo Guo", "Kumar Krishna Agrawal", "Aditya Grover", "Vidya Muthukumar", "Ashwin Pananjady"], "categories": ["stat.ML", "cs.AI", "cs.IT", "cs.LG", "cs.RO"], "fields_of_study": ["Mathematics", "Computer Science"], "published_date": "2021-06-28", "url": "https://arxiv.org/abs/2106.14866", "pdf_url": "https://arxiv.org/pdf/2106.14866v2", "arxiv_id": "2106.14866", "doi": null, "citation_count": 9, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.25}
{"id": "156fe58a33d9789b4519707c3e2ac6fd9ec540b9f528537bf3c17d2f1ea83857", "sources": ["arxiv", "semantic_scholar"], "title": "On-Policy Deep Reinforcement Learning for the Average-Reward Criterion", "abstract": "We develop theory and algorithms for average-reward on-policy Reinforcement Learning (RL). We first consider bounding the difference of the long-term average reward for two policies. We show that previous work based on the discounted return (Schulman et al., 2015; Achiam et al., 2017) results in a non-meaningful bound in the average-reward setting. By addressing the average-reward criterion directly, we then derive a novel bound which depends on the average divergence between the two policies and Kemeny's constant. Based on this bound, we develop an iterative procedure which produces a sequence of monotonically improved policies for the average reward criterion. This iterative procedure can then be combined with classic DRL (Deep Reinforcement Learning) methods, resulting in practical DRL algorithms that target the long-run average reward criterion. In particular, we demonstrate that Average-Reward TRPO (ATRPO), which adapts the on-policy TRPO algorithm to the average-reward criterion, significantly outperforms TRPO in the most challenging MuJuCo environments.", "authors": ["Yiming Zhang", "Keith W. Ross"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2021-06-14", "url": "https://arxiv.org/abs/2106.07329", "pdf_url": "https://arxiv.org/pdf/2106.07329v1", "arxiv_id": "2106.07329", "doi": null, "citation_count": 59, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4445}
{"id": "fc0dc38c4096639e0a20718fc911aafef48e1601861ca297bd4ffa8c1a5ed44d", "sources": ["arxiv", "semantic_scholar"], "title": "Reward prediction for representation learning and reward shaping", "abstract": "One of the fundamental challenges in reinforcement learning (RL) is the one of data efficiency: modern algorithms require a very large number of training samples, especially compared to humans, for solving environments with high-dimensional observations. The severity of this problem is increased when the reward signal is sparse. In this work, we propose learning a state representation in a self-supervised manner for reward prediction. The reward predictor learns to estimate either a raw or a smoothed version of the true reward signal in environment with a single, terminating, goal state. We augment the training of out-of-the-box RL agents by shaping the reward using our reward predictor during policy learning. Using our representation for preprocessing high-dimensional observations, as well as using the predictor for reward shaping, is shown to significantly enhance Actor Critic using Kronecker-factored Trust Region and Proximal Policy Optimization in single-goal environments with visual inputs.", "authors": ["Hlynur Davíð Hlynsson", "Laurenz Wiskott"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2021-05-07", "url": "https://arxiv.org/abs/2105.03172", "pdf_url": "https://arxiv.org/pdf/2105.03172v1", "arxiv_id": "2105.03172", "doi": "10.5220/0010640200003063", "citation_count": 3, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Joint Conference on Computational Intelligence", "quality_score": 0.1505}
{"id": "ea775b8090b33712a0c0f71414ff1615e189e1b4f063bb21b0d5c31114b5e052", "sources": ["arxiv", "semantic_scholar"], "title": "Generative Adversarial Reward Learning for Generalized Behavior Tendency Inference", "abstract": "Recent advances in reinforcement learning have inspired increasing interest in learning user modeling adaptively through dynamic interactions, e.g., in reinforcement learning based recommender systems. Reward function is crucial for most of reinforcement learning applications as it can provide the guideline about the optimization. However, current reinforcement-learning-based methods rely on manually-defined reward functions, which cannot adapt to dynamic and noisy environments. Besides, they generally use task-specific reward functions that sacrifice generalization ability. We propose a generative inverse reinforcement learning for user behavioral preference modelling, to address the above issues. Instead of using predefined reward functions, our model can automatically learn the rewards from user's actions based on discriminative actor-critic network and Wasserstein GAN. Our model provides a general way of characterizing and explaining underlying behavioral tendencies, and our experiments show our method outperforms state-of-the-art methods in a variety of scenarios, namely traffic signal control, online recommender systems, and scanpath prediction.", "authors": ["Xiaocong Chen", "Lina Yao", "Xianzhi Wang", "Aixin Sun", "Wenjie Zhang", "Quan Z. Sheng"], "categories": ["cs.LG", "cs.AI", "cs.IR"], "fields_of_study": ["Computer Science"], "published_date": "2021-05-03", "url": "https://arxiv.org/abs/2105.00822", "pdf_url": "https://arxiv.org/pdf/2105.00822v2", "arxiv_id": "2105.00822", "doi": "10.1109/TKDE.2022.3186920", "citation_count": 12, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Knowledge and Data Engineering", "quality_score": 0.2785}
{"id": "9473ecad2878352a26d3b70a9ff6a3f45455efcc272a0d0d3bc6f87a76a8d28c", "sources": ["arxiv", "semantic_scholar"], "title": "Subgoal-based Reward Shaping to Improve Efficiency in Reinforcement Learning", "abstract": "Reinforcement learning, which acquires a policy maximizing long-term rewards, has been actively studied. Unfortunately, this learning type is too slow and difficult to use in practical situations because the state-action space becomes huge in real environments. Many studies have incorporated human knowledge into reinforcement Learning. Though human knowledge on trajectories is often used, a human could be asked to control an AI agent, which can be difficult. Knowledge on subgoals may lessen this requirement because humans need only to consider a few representative states on an optimal trajectory in their minds. The essential factor for learning efficiency is rewards. Potential-based reward shaping is a basic method for enriching rewards. However, it is often difficult to incorporate subgoals for accelerating learning over potential-based reward shaping. This is because the appropriate potentials are not intuitive for humans. We extend potential-based reward shaping and propose a subgoal-based reward shaping. The method makes it easier for human trainers to share their knowledge of subgoals. To evaluate our method, we obtained a subgoal series from participants and conducted experiments in three domains, four-rooms(discrete states and discrete actions), pinball(continuous and discrete), and picking(both continuous). We compared our method with a baseline reinforcement learning algorithm and other subgoal-based methods, including random subgoal and naive subgoal-based reward shaping. As a result, we found out that our reward shaping outperformed all other methods in learning efficiency.", "authors": ["Takato Okudo", "Seiji Yamada"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2021-04-13", "url": "https://arxiv.org/abs/2104.06411", "pdf_url": "https://arxiv.org/pdf/2104.06411v1", "arxiv_id": "2104.06411", "doi": "10.1109/ACCESS.2021.3090364", "citation_count": 26, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "IEEE Access", "quality_score": 0.3578}
{"id": "451aa27b68f118ef38189f8cbb1a41ba6793454db76e38abec47c2ff17a531f8", "sources": ["arxiv", "semantic_scholar"], "title": "On Preference Learning Based on Sequential Bayesian Optimization with Pairwise Comparison", "abstract": "User preference learning is generally a hard problem. Individual preferences are typically unknown even to users themselves, while the space of choices is infinite. Here we study user preference learning from information-theoretic perspective. We model preference learning as a system with two interacting sub-systems, one representing a user with his/her preferences and another one representing an agent that has to learn these preferences. The user with his/her behaviour is modeled by a parametric preference function. To efficiently learn the preferences and reduce search space quickly, we propose the agent that interacts with the user to collect the most informative data for learning. The agent presents two proposals to the user for evaluation, and the user rates them based on his/her preference function. We show that the optimum agent strategy for data collection and preference learning is a result of maximin optimization of the normalized weighted Kullback-Leibler (KL) divergence between true and agent-assigned predictive user response distributions. The resulting value of KL-divergence, which we also call remaining system uncertainty (RSU), provides an efficient performance metric in the absence of the ground truth. This metric characterises how well the agent can predict user and, thus, the quality of the underlying learned user (preference) model. Our proposed agent comprises sequential mechanisms for user model inference and proposal generation. To infer the user model (preference function), Bayesian approximate inference is used in the agent. The data collection strategy is to generate proposals, responses to which help resolving uncertainty associated with prediction of the user responses the most. The efficiency of our approach is validated by numerical simulations. Also a real-life example of preference learning application is provided.", "authors": ["Tanya Ignatenko", "Kirill Kondrashov", "Marco Cox", "Bert de Vries"], "categories": ["cs.LG", "cs.AI", "cs.IT", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2021-03-24", "url": "https://arxiv.org/abs/2103.13192", "pdf_url": "https://arxiv.org/pdf/2103.13192v3", "arxiv_id": "2103.13192", "doi": "10.1016/j.artint.2025.104400", "citation_count": 9, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Artificial Intelligence", "quality_score": 0.25}
{"id": "d40626080fa3877e01932895e43ec6e7e6972090b1c50228278f2e7999970fc8", "sources": ["arxiv", "semantic_scholar"], "title": "Learning One Representation to Optimize All Rewards", "abstract": "We introduce the forward-backward (FB) representation of the dynamics of a reward-free Markov decision process. It provides explicit near-optimal policies for any reward specified a posteriori. During an unsupervised phase, we use reward-free interactions with the environment to learn two representations via off-the-shelf deep learning methods and temporal difference (TD) learning. In the test phase, a reward representation is estimated either from observations or an explicit reward description (e.g., a target state). The optimal policy for that reward is directly obtained from these representations, with no planning. We assume access to an exploration scheme or replay buffer for the first phase. The corresponding unsupervised loss is well-principled: if training is perfect, the policies obtained are provably optimal for any reward function. With imperfect training, the sub-optimality is proportional to the unsupervised approximation error. The FB representation learns long-range relationships between states and actions, via a predictive occupancy map, without having to synthesize states as in model-based approaches. This is a step towards learning controllable agents in arbitrary black-box stochastic environments. This approach compares well to goal-oriented RL algorithms on discrete and continuous mazes, pixel-based MsPacman, and the FetchReach virtual robot arm. We also illustrate how the agent can immediately adapt to new tasks beyond goal-oriented RL.", "authors": ["Ahmed Touati", "Yann Ollivier"], "categories": ["cs.LG", "cs.AI", "math.OC"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2021-03-14", "url": "https://arxiv.org/abs/2103.07945", "pdf_url": "https://arxiv.org/pdf/2103.07945v3", "arxiv_id": "2103.07945", "doi": null, "citation_count": 119, "influential_citation_count": 27, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.7236}
{"id": "8189d17e14a8f511f295005f9e823462aee926497c6294176b41128317152219", "sources": ["arxiv", "semantic_scholar"], "title": "Model-free Policy Learning with Reward Gradients", "abstract": "Despite the increasing popularity of policy gradient methods, they are yet to be widely utilized in sample-scarce applications, such as robotics. The sample efficiency could be improved by making best usage of available information. As a key component in reinforcement learning, the reward function is usually devised carefully to guide the agent. Hence, the reward function is usually known, allowing access to not only scalar reward signals but also reward gradients. To benefit from reward gradients, previous works require the knowledge of environment dynamics, which are hard to obtain. In this work, we develop the \\textit{Reward Policy Gradient} estimator, a novel approach that integrates reward gradients without learning a model. Bypassing the model dynamics allows our estimator to achieve a better bias-variance trade-off, which results in a higher sample efficiency, as shown in the empirical analysis. Our method also boosts the performance of Proximal Policy Optimization on different MuJoCo control tasks.", "authors": ["Qingfeng Lan", "Samuele Tosatto", "Homayoon Farrahi", "A. Rupam Mahmood"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2021-03-09", "url": "https://arxiv.org/abs/2103.05147", "pdf_url": "https://arxiv.org/pdf/2103.05147v4", "arxiv_id": "2103.05147", "doi": null, "citation_count": 9, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "International Conference on Artificial Intelligence and Statistics", "quality_score": 0.301}
{"id": "23cf557cab73f53441b6bbdda89f335fd027bce0b73624b316e52d58d2ff78d6", "sources": ["arxiv", "semantic_scholar"], "title": "Information Directed Reward Learning for Reinforcement Learning", "abstract": "For many reinforcement learning (RL) applications, specifying a reward is difficult. This paper considers an RL setting where the agent obtains information about the reward only by querying an expert that can, for example, evaluate individual states or provide binary preferences over trajectories. From such expensive feedback, we aim to learn a model of the reward that allows standard RL algorithms to achieve high expected returns with as few expert queries as possible. To this end, we propose Information Directed Reward Learning (IDRL), which uses a Bayesian model of the reward and selects queries that maximize the information gain about the difference in return between plausibly optimal policies. In contrast to prior active reward learning methods designed for specific types of queries, IDRL naturally accommodates different query types. Moreover, it achieves similar or better performance with significantly fewer queries by shifting the focus from reducing the reward approximation error to improving the policy induced by the reward model. We support our findings with extensive evaluations in multiple environments and with different query types.", "authors": ["David Lindner", "Matteo Turchetta", "Sebastian Tschiatschek", "Kamil Ciosek", "Andreas Krause"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2021-02-24", "url": "https://arxiv.org/abs/2102.12466", "pdf_url": "https://arxiv.org/pdf/2102.12466v3", "arxiv_id": "2102.12466", "doi": null, "citation_count": 26, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.3578}
{"id": "0fb6b8746a9fc0a851a1e5eaa8162fcf44bf6ca74ff2fd70fdeadf39659c8da9", "sources": ["arxiv", "semantic_scholar"], "title": "In-Loop Meta-Learning with Gradient-Alignment Reward", "abstract": "At the heart of the standard deep learning training loop is a greedy gradient step minimizing a given loss. We propose to add a second step to maximize training generalization. To do this, we optimize the loss of the next training step. While computing the gradient for this generally is very expensive and many interesting applications consider non-differentiable parameters (e.g. due to hard samples), we present a cheap-to-compute and memory-saving reward, the gradient-alignment reward (GAR), that can guide the optimization. We use this reward to optimize multiple distributions during model training. First, we present the application of GAR to choosing the data distribution as a mixture of multiple dataset splits in a small scale setting. Second, we show that it can successfully guide learning augmentation strategies competitive with state-of-the-art augmentation strategies on CIFAR-10 and CIFAR-100.", "authors": ["Samuel Müller", "André Biedenkapp", "Frank Hutter"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2021-02-05", "url": "https://arxiv.org/abs/2102.03275", "pdf_url": "https://arxiv.org/pdf/2102.03275v1", "arxiv_id": "2102.03275", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.1193}
{"id": "b90bcdd71d1410b7354a9524886a6f7eb8789e4bf8dd5e2b3b9a137045470466", "sources": ["arxiv", "semantic_scholar"], "title": "Prior Preference Learning from Experts:Designing a Reward with Active Inference", "abstract": "Active inference may be defined as Bayesian modeling of a brain with a biologically plausible model of the agent. Its primary idea relies on the free energy principle and the prior preference of the agent. An agent will choose an action that leads to its prior preference for a future observation. In this paper, we claim that active inference can be interpreted using reinforcement learning (RL) algorithms and find a theoretical connection between them. We extend the concept of expected free energy (EFE), which is a core quantity in active inference, and claim that EFE can be treated as a negative value function. Motivated by the concept of prior preference and a theoretical connection, we propose a simple but novel method for learning a prior preference from experts. This illustrates that the problem with inverse RL can be approached with a new perspective of active inference. Experimental results of prior preference learning show the possibility of active inference with EFE-based rewards and its application to an inverse RL problem.", "authors": ["Jin young Shin", "Cheolhyeong Kim", "Hyung Ju Hwang"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2021-01-22", "url": "https://arxiv.org/abs/2101.08937", "pdf_url": "https://arxiv.org/pdf/2101.08937v3", "arxiv_id": "2101.08937", "doi": "10.1016/j.neucom.2021.12.042", "citation_count": 13, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "Neurocomputing", "quality_score": 0.2865}
{"id": "9c251f776321488dbf5eea0ea87e3bbf7f7e7618f59fca2b6e763b76c2ad2861", "sources": ["arxiv", "semantic_scholar"], "title": "Semi-supervised reward learning for offline reinforcement learning", "abstract": "In offline reinforcement learning (RL) agents are trained using a logged dataset. It appears to be the most natural route to attack real-life applications because in domains such as healthcare and robotics interactions with the environment are either expensive or unethical. Training agents usually requires reward functions, but unfortunately, rewards are seldom available in practice and their engineering is challenging and laborious. To overcome this, we investigate reward learning under the constraint of minimizing human reward annotations. We consider two types of supervision: timestep annotations and demonstrations. We propose semi-supervised learning algorithms that learn from limited annotations and incorporate unlabelled data. In our experiments with a simulated robotic arm, we greatly improve upon behavioural cloning and closely approach the performance achieved with ground truth rewards. We further investigate the relationship between the quality of the reward model and the final policies. We notice, for example, that the reward models do not need to be perfect to result in useful policies.", "authors": ["Ksenia Konyushkova", "Konrad Zolna", "Yusuf Aytar", "Alexander Novikov", "Scott Reed", "Serkan Cabi", "Nando de Freitas"], "categories": ["cs.LG", "cs.AI", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2020-12-12", "url": "https://arxiv.org/abs/2012.06899", "pdf_url": "https://arxiv.org/pdf/2012.06899v1", "arxiv_id": "2012.06899", "doi": null, "citation_count": 26, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.3578}
{"id": "b969e6cc33e2294b996d79124ba9a815f1c783fae8659edba8af28f3be1cd58a", "sources": ["arxiv", "semantic_scholar"], "title": "Understanding Learned Reward Functions", "abstract": "In many real-world tasks, it is not possible to procedurally specify an RL agent's reward function. In such cases, a reward function must instead be learned from interacting with and observing humans. However, current techniques for reward learning may fail to produce reward functions which accurately reflect user preferences. Absent significant advances in reward learning, it is thus important to be able to audit learned reward functions to verify whether they truly capture user preferences. In this paper, we investigate techniques for interpreting learned reward functions. In particular, we apply saliency methods to identify failure modes and predict the robustness of reward functions. We find that learned reward functions often implement surprising algorithms that rely on contingent aspects of the environment. We also discover that existing interpretability techniques often attend to irrelevant changes in reward output, suggesting that reward interpretability may need significantly different methods from policy interpretability.", "authors": ["Eric J. Michaud", "Adam Gleave", "Stuart Russell"], "categories": ["cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2020-12-10", "url": "https://arxiv.org/abs/2012.05862", "pdf_url": "https://arxiv.org/pdf/2012.05862v1", "arxiv_id": "2012.05862", "doi": null, "citation_count": 40, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4032}
{"id": "6ff576ef76748ebd6d62a25caae2cc6739ff832de381ae33140dd5d161d2cc78", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Biased Maximum Likelihood Estimation for Reinforcement Learning", "abstract": "The Reward-Biased Maximum Likelihood Estimate (RBMLE) for adaptive control of Markov chains was proposed to overcome the central obstacle of what is variously called the fundamental \"closed-identifiability problem\" of adaptive control, the \"dual control problem\", or, contemporaneously, the \"exploration vs. exploitation problem\". It exploited the key observation that since the maximum likelihood parameter estimator can asymptotically identify the closed-transition probabilities under a certainty equivalent approach, the limiting parameter estimates must necessarily have an optimal reward that is less than the optimal reward attainable for the true but unknown system. Hence it proposed a counteracting reverse bias in favor of parameters with larger optimal rewards, providing a solution to the fundamental problem alluded to above. It thereby proposed an optimistic approach of favoring parameters with larger optimal rewards, now known as \"optimism in the face of uncertainty\". The RBMLE approach has been proved to be long-term average reward optimal in a variety of contexts. However, modern attention is focused on the much finer notion of \"regret\", or finite-time performance. Recent analysis of RBMLE for multi-armed stochastic bandits and linear contextual bandits has shown that it not only has state-of-the-art regret, but it also exhibits empirical performance comparable to or better than the best current contenders, and leads to strikingly simple index policies. Motivated by this, we examine the finite-time performance of RBMLE for reinforcement learning tasks that involve the general problem of optimal control of unknown Markov Decision Processes. We show that it has a regret of $\\mathcal{O}( \\log T)$ over a time horizon of $T$ steps, similar to state-of-the-art algorithms. Simulation studies show that RBMLE outperforms other algorithms such as UCRL2 and Thompson Sampling.", "authors": ["Akshay Mete", "Rahul Singh", "Xi Liu", "P. R. Kumar"], "categories": ["cs.LG", "eess.SY"], "fields_of_study": ["Computer Science", "Engineering"], "published_date": "2020-11-16", "url": "https://arxiv.org/abs/2011.07738", "pdf_url": "https://arxiv.org/pdf/2011.07738v3", "arxiv_id": "2011.07738", "doi": null, "citation_count": 25, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Conference on Learning for Dynamics & Control", "quality_score": 0.3537}
{"id": "d463e4213e11da7380a37d15edc0f77882b2bd99ace056158cb7096187c32cc4", "sources": ["arxiv", "semantic_scholar"], "title": "Active Reinforcement Learning: Observing Rewards at a Cost", "abstract": "Active reinforcement learning (ARL) is a variant on reinforcement learning where the agent does not observe the reward unless it chooses to pay a query cost c > 0. The central question of ARL is how to quantify the long-term value of reward information. Even in multi-armed bandits, computing the value of this information is intractable and we have to rely on heuristics. We propose and evaluate several heuristic approaches for ARL in multi-armed bandits and (tabular) Markov decision processes, and discuss and illustrate some challenging aspects of the ARL problem.", "authors": ["David Krueger", "Jan Leike", "Owain Evans", "John Salvatier"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-11-13", "url": "https://arxiv.org/abs/2011.06709", "pdf_url": "https://arxiv.org/pdf/2011.06709v2", "arxiv_id": "2011.06709", "doi": null, "citation_count": 40, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.4032}
{"id": "39ee978b81fd09b1b0d8283543d16269b91722d6076fc45e611f745cd8d94a3b", "sources": ["arxiv", "semantic_scholar"], "title": "Learning to Utilize Shaping Rewards: A New Approach of Reward Shaping", "abstract": "Reward shaping is an effective technique for incorporating domain knowledge into reinforcement learning (RL). Existing approaches such as potential-based reward shaping normally make full use of a given shaping reward function. However, since the transformation of human knowledge into numeric reward values is often imperfect due to reasons such as human cognitive bias, completely utilizing the shaping reward function may fail to improve the performance of RL algorithms. In this paper, we consider the problem of adaptively utilizing a given shaping reward function. We formulate the utilization of shaping rewards as a bi-level optimization problem, where the lower level is to optimize policy using the shaping rewards and the upper level is to optimize a parameterized shaping weight function for true reward maximization. We formally derive the gradient of the expected true reward with respect to the shaping weight function parameters and accordingly propose three learning algorithms based on different assumptions. Experiments in sparse-reward cartpole and MuJoCo environments show that our algorithms can fully exploit beneficial shaping rewards, and meanwhile ignore unbeneficial shaping rewards or even transform them into beneficial ones.", "authors": ["Yujing Hu", "Weixun Wang", "Hangtian Jia", "Yixiang Wang", "Yingfeng Chen", "Jianye Hao", "Feng Wu", "Changjie Fan"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2020-11-05", "url": "https://arxiv.org/abs/2011.02669", "pdf_url": "https://arxiv.org/pdf/2011.02669v1", "arxiv_id": "2011.02669", "doi": null, "citation_count": 238, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.5946}
{"id": "eb624bcbcc160f27f09c0f4abf4237e59188247fa098f7e50efda56704d93f7e", "sources": ["arxiv", "semantic_scholar"], "title": "Collaborative Machine Learning with Incentive-Aware Model Rewards", "abstract": "Collaborative machine learning (ML) is an appealing paradigm to build high-quality ML models by training on the aggregated data from many parties. However, these parties are only willing to share their data when given enough incentives, such as a guaranteed fair reward based on their contributions. This motivates the need for measuring a party's contribution and designing an incentive-aware reward scheme accordingly. This paper proposes to value a party's reward based on Shapley value and information gain on model parameters given its data. Subsequently, we give each party a model as a reward. To formally incentivize the collaboration, we define some desirable properties (e.g., fairness and stability) which are inspired by cooperative game theory but adapted for our model reward that is uniquely freely replicable. Then, we propose a novel model reward scheme to satisfy fairness and trade off between the desirable properties via an adjustable parameter. The value of each party's model reward determined by our scheme is attained by injecting Gaussian noise to the aggregated training data with an optimized noise variance. We empirically demonstrate interesting properties of our scheme and evaluate its performance using synthetic and real-world datasets.", "authors": ["Rachael Hwee Ling Sim", "Yehong Zhang", "Mun Choon Chan", "Bryan Kian Hsiang Low"], "categories": ["cs.LG", "cs.GT", "cs.MA", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-10-24", "url": "https://arxiv.org/abs/2010.12797", "pdf_url": "https://arxiv.org/pdf/2010.12797v1", "arxiv_id": "2010.12797", "doi": null, "citation_count": 161, "influential_citation_count": 12, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.557}
{"id": "fa824431e2fccf7c6d1a5c02a6275116004014a67bd4e02aea555a62ec443ed7", "sources": ["arxiv", "semantic_scholar"], "title": "Learning from Suboptimal Demonstration via Self-Supervised Reward Regression", "abstract": "Learning from Demonstration (LfD) seeks to democratize robotics by enabling non-roboticist end-users to teach robots to perform a task by providing a human demonstration. However, modern LfD techniques, e.g. inverse reinforcement learning (IRL), assume users provide at least stochastically optimal demonstrations. This assumption fails to hold in most real-world scenarios. Recent attempts to learn from sub-optimal demonstration leverage pairwise rankings and following the Luce-Shepard rule. However, we show these approaches make incorrect assumptions and thus suffer from brittle, degraded performance. We overcome these limitations in developing a novel approach that bootstraps off suboptimal demonstrations to synthesize optimality-parameterized data to train an idealized reward function. We empirically validate we learn an idealized reward function with ~0.95 correlation with ground-truth reward versus ~0.75 for prior work. We can then train policies achieving ~200% improvement over the suboptimal demonstration and ~90% improvement over prior work. We present a physical demonstration of teaching a robot a topspin strike in table tennis that achieves 32% faster returns and 40% more topspin than user demonstration.", "authors": ["Letian Chen", "Rohan Paleja", "Matthew Gombolay"], "categories": ["cs.RO", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2020-10-17", "url": "https://arxiv.org/abs/2010.11723", "pdf_url": "https://arxiv.org/pdf/2010.11723v3", "arxiv_id": "2010.11723", "doi": null, "citation_count": 134, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "Conference on Robot Learning", "quality_score": 0.5326}
{"id": "0ee461d1f8a65ea771251a605df5b7167d001f4d1311da2a0a2ae557ae0178d6", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Machines: Exploiting Reward Function Structure in Reinforcement Learning", "abstract": "Reinforcement learning (RL) methods usually treat reward functions as black boxes. As such, these methods must extensively interact with the environment in order to discover rewards and optimal policies. In most RL applications, however, users have to program the reward function and, hence, there is the opportunity to make the reward function visible -- to show the reward function's code to the RL agent so it can exploit the function's internal structure to learn optimal policies in a more sample efficient manner. In this paper, we show how to accomplish this idea in two steps. First, we propose reward machines, a type of finite state machine that supports the specification of reward functions while exposing reward function structure. We then describe different methodologies to exploit this structure to support learning, including automated reward shaping, task decomposition, and counterfactual reasoning with off-policy learning. Experiments on tabular and continuous domains, across different tasks and RL agents, show the benefits of exploiting reward structure with respect to sample efficiency and the quality of resultant policies. Finally, by virtue of being a form of finite state machine, reward machines have the expressive power of a regular language and as such support loops, sequences and conditionals, as well as the expression of temporally extended properties typical of linear temporal logic and non-Markovian reward specification.", "authors": ["Rodrigo Toro Icarte", "Toryn Q. Klassen", "Richard Valenzano", "Sheila A. McIlraith"], "categories": ["cs.LG", "cs.AI"], "fields_of_study": ["Computer Science"], "published_date": "2020-10-06", "url": "https://arxiv.org/abs/2010.03950", "pdf_url": "https://arxiv.org/pdf/2010.03950v2", "arxiv_id": "2010.03950", "doi": "10.1613/jair.1.12440", "citation_count": 318, "influential_citation_count": 40, "has_code": false, "code_url": null, "venue": "Journal of Artificial Intelligence Research", "quality_score": 0.8064}
{"id": "59a0e4e98d8d4f5ea1cf1e99304139971a062d60727f611fd3cfd9d3be3c3a86", "sources": ["arxiv", "semantic_scholar"], "title": "Deep Reinforcement Learning with a Stage Incentive Mechanism of Dense Reward for Robotic Trajectory Planning", "abstract": "(This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.) To improve the efficiency of deep reinforcement learning (DRL)-based methods for robot manipulator trajectory planning in random working environments, we present three dense reward functions. These rewards differ from the traditional sparse reward. First, a posture reward function is proposed to speed up the learning process with a more reasonable trajectory by modeling the distance and direction constraints, which can reduce the blindness of exploration. Second, a stride reward function is proposed to improve the stability of the learning process by modeling the distance and movement distance of joint constraints. Finally, in order to further improve learning efficiency, we are inspired by the cognitive process of human behavior and propose a stage incentive mechanism, including a hard stage incentive reward function and a soft stage incentive reward function. Extensive experiments show that the soft stage incentive reward function is able to improve the convergence rate by up to 46.9% with the state-of-the-art DRL methods. The percentage increase in the convergence mean reward was 4.4-15.5% and the percentage decreases with respect to standard deviation were 21.9-63.2%. In the evaluation experiments, the success rate of trajectory planning for a robot manipulator reached 99.6%.", "authors": ["Gang Peng", "Jin Yang", "Xinde Lia", "Mohammad Omar Khyam"], "categories": ["cs.AI", "cs.LG", "cs.RO"], "fields_of_study": ["Computer Science"], "published_date": "2020-09-25", "url": "https://arxiv.org/abs/2009.12068", "pdf_url": "https://arxiv.org/pdf/2009.12068v2", "arxiv_id": "2009.12068", "doi": "10.1109/TSMC.2022.3228901", "citation_count": 28, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3656}
{"id": "553ad16c1af559f5e6f9fb72807b6ecbe0540471aa739417c395ec04ae5f6db9", "sources": ["arxiv", "semantic_scholar"], "title": "Addressing reward bias in Adversarial Imitation Learning with neutral reward functions", "abstract": "Generative Adversarial Imitation Learning suffers from the fundamental problem of reward bias stemming from the choice of reward functions used in the algorithm. Different types of biases also affect different types of environments - which are broadly divided into survival and task-based environments. We provide a theoretical sketch of why existing reward functions would fail in imitation learning scenarios in task based environments with multiple terminal states. We also propose a new reward function for GAIL which outperforms existing GAIL methods on task based environments with single and multiple terminal states and effectively overcomes both survival and termination bias.", "authors": ["Rohit Jena", "Siddharth Agrawal", "Katia Sycara"], "categories": ["cs.LG", "cs.RO", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-09-20", "url": "https://arxiv.org/abs/2009.09467", "pdf_url": "https://arxiv.org/pdf/2009.09467v1", "arxiv_id": "2009.09467", "doi": null, "citation_count": 9, "influential_citation_count": 2, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.25}
{"id": "f5193d0072620109f3057b502f0e59c5cfa0669b21eac774e870bb12ab3a79e0", "sources": ["arxiv", "semantic_scholar"], "title": "Online Semi-Supervised Learning in Contextual Bandits with Episodic Reward", "abstract": "We considered a novel practical problem of online learning with episodically revealed rewards, motivated by several real-world applications, where the contexts are nonstationary over different episodes and the reward feedbacks are not always available to the decision making agents. For this online semi-supervised learning setting, we introduced Background Episodic Reward LinUCB (BerlinUCB), a solution that easily incorporates clustering as a self-supervision module to provide useful side information when rewards are not observed. Our experiments on a variety of datasets, both in stationary and nonstationary environments of six different scenarios, demonstrated clear advantages of the proposed approach over the standard contextual bandit. Lastly, we introduced a relevant real-life example where this problem setting is especially useful.", "authors": ["Baihan Lin"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-09-17", "url": "https://arxiv.org/abs/2009.08457", "pdf_url": "https://arxiv.org/pdf/2009.08457v2", "arxiv_id": "2009.08457", "doi": "10.1007/978-3-030-64984-5_32", "citation_count": 14, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.294}
{"id": "890e318151cf3e61af970c30d61fec137264f5e0d3913792f37f7a6203880061", "sources": ["arxiv", "semantic_scholar"], "title": "Predicting the Humorousness of Tweets Using Gaussian Process Preference Learning", "abstract": "Most humour processing systems to date make at best discrete, coarse-grained distinctions between the comical and the conventional, yet such notions are better conceptualized as a broad spectrum. In this paper, we present a probabilistic approach, a variant of Gaussian process preference learning (GPPL), that learns to rank and rate the humorousness of short texts by exploiting human preference judgments and automatically sourced linguistic annotations. We apply our system, which is similar to one that had previously shown good performance on English-language one-liners annotated with pairwise humorousness annotations, to the Spanish-language data set of the HAHA@IberLEF2019 evaluation campaign. We report system performance for the campaign's two subtasks, humour detection and funniness score prediction, and discuss some issues arising from the conversion between the numeric scores used in the HAHA@IberLEF2019 data and the pairwise judgment annotations required for our method.", "authors": ["Tristan Miller", "Erik-Lân Do Dinh", "Edwin Simpson", "Iryna Gurevych"], "categories": ["cs.CL"], "fields_of_study": ["Computer Science"], "published_date": "2020-08-03", "url": "https://arxiv.org/abs/2008.00853", "pdf_url": "https://arxiv.org/pdf/2008.00853v2", "arxiv_id": "2008.00853", "doi": null, "citation_count": 5, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Procesamiento del Lenguaje Natural, 64:37-44, March 2020", "quality_score": 0.1945}
{"id": "806fea24c877342c0e0b7bdf1512962a4cd078925e5811e198cde7517a4cbfb3", "sources": ["arxiv", "semantic_scholar"], "title": "PixL2R: Guiding Reinforcement Learning Using Natural Language by Mapping Pixels to Rewards", "abstract": "Reinforcement learning (RL), particularly in sparse reward settings, often requires prohibitively large numbers of interactions with the environment, thereby limiting its applicability to complex problems. To address this, several prior approaches have used natural language to guide the agent's exploration. However, these approaches typically operate on structured representations of the environment, and/or assume some structure in the natural language commands. In this work, we propose a model that directly maps pixels to rewards, given a free-form natural language description of the task, which can then be used for policy learning. Our experiments on the Meta-World robot manipulation domain show that language-based rewards significantly improves the sample efficiency of policy learning, both in sparse and dense reward settings.", "authors": ["Prasoon Goyal", "Scott Niekum", "Raymond J. Mooney"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-07-30", "url": "https://arxiv.org/abs/2007.15543", "pdf_url": "https://arxiv.org/pdf/2007.15543v2", "arxiv_id": "2007.15543", "doi": null, "citation_count": 60, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": "Conference on Robot Learning", "quality_score": 0.4463}
{"id": "20a0d18dbc0a0c88aa896da3169d5d42d18f349f37f51e8c793b5702015be9a7", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Abstract Models for Strategic Exploration and Fast Reward Transfer", "abstract": "Model-based reinforcement learning (RL) is appealing because (i) it enables planning and thus more strategic exploration, and (ii) by decoupling dynamics from rewards, it enables fast transfer to new reward functions. However, learning an accurate Markov Decision Process (MDP) over high-dimensional states (e.g., raw pixels) is extremely challenging because it requires function approximation, which leads to compounding errors. Instead, to avoid compounding errors, we propose learning an abstract MDP over abstract states: low-dimensional coarse representations of the state (e.g., capturing agent position, ignoring other objects). We assume access to an abstraction function that maps the concrete states to abstract states. In our approach, we construct an abstract MDP, which grows through strategic exploration via planning. Similar to hierarchical RL approaches, the abstract actions of the abstract MDP are backed by learned subpolicies that navigate between abstract states. Our approach achieves strong results on three of the hardest Arcade Learning Environment games (Montezuma's Revenge, Pitfall!, and Private Eye), including superhuman performance on Pitfall! without demonstrations. After training on one task, we can reuse the learned abstract MDP for new reward functions, achieving higher reward in 1000x fewer samples than model-free methods trained from scratch.", "authors": ["Evan Zheran Liu", "Ramtin Keramati", "Sudarshan Seshadri", "Kelvin Guu", "Panupong Pasupat", "Emma Brunskill", "Percy Liang"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-07-12", "url": "https://arxiv.org/abs/2007.05896", "pdf_url": "https://arxiv.org/pdf/2007.05896v1", "arxiv_id": "2007.05896", "doi": null, "citation_count": 6, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.2113}
{"id": "759a89111b23a23faf26a7a3d8ebf95d39a53f664be14d1e05d475da2dd7971f", "sources": ["arxiv", "semantic_scholar"], "title": "Intrinsic Reward Driven Imitation Learning via Generative Model", "abstract": "Imitation learning in a high-dimensional environment is challenging. Most inverse reinforcement learning (IRL) methods fail to outperform the demonstrator in such a high-dimensional environment, e.g., Atari domain. To address this challenge, we propose a novel reward learning module to generate intrinsic reward signals via a generative model. Our generative method can perform better forward state transition and backward action encoding, which improves the module's dynamics modeling ability in the environment. Thus, our module provides the imitation agent both the intrinsic intention of the demonstrator and a better exploration ability, which is critical for the agent to outperform the demonstrator. Empirical results show that our method outperforms state-of-the-art IRL methods on multiple Atari games, even with one-life demonstration. Remarkably, our method achieves performance that is up to 5 times the performance of the demonstration.", "authors": ["Xingrui Yu", "Yueming Lyu", "Ivor W. Tsang"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-06-26", "url": "https://arxiv.org/abs/2006.15061", "pdf_url": "https://arxiv.org/pdf/2006.15061v4", "arxiv_id": "2006.15061", "doi": null, "citation_count": 64, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.4771}
{"id": "2f362d2929aa11f38dbfa288f1580e3e4c0cced10ccbb1ac3aa7871cd5aa1a24", "sources": ["arxiv", "semantic_scholar"], "title": "Learning Reward Functions from Diverse Sources of Human Feedback: Optimally Integrating Demonstrations and Preferences", "abstract": "Reward functions are a common way to specify the objective of a robot. As designing reward functions can be extremely challenging, a more promising approach is to directly learn reward functions from human teachers. Importantly, data from human teachers can be collected either passively or actively in a variety of forms: passive data sources include demonstrations, (e.g., kinesthetic guidance), whereas preferences (e.g., comparative rankings) are actively elicited. Prior research has independently applied reward learning to these different data sources. However, there exist many domains where multiple sources are complementary and expressive. Motivated by this general problem, we present a framework to integrate multiple sources of information, which are either passively or actively collected from human users. In particular, we present an algorithm that first utilizes user demonstrations to initialize a belief about the reward function, and then actively probes the user with preference queries to zero-in on their true reward. This algorithm not only enables us combine multiple data sources, but it also informs the robot when it should leverage each type of information. Further, our approach accounts for the human's ability to provide data: yielding user-friendly preference queries which are also theoretically optimal. Our extensive simulated experiments and user studies on a Fetch mobile manipulator demonstrate the superiority and the usability of our integrated framework.", "authors": ["Erdem Bıyık", "Dylan P. Losey", "Malayandi Palan", "Nicholas C. Landolfi", "Gleb Shevchuk", "Dorsa Sadigh"], "categories": ["cs.RO", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2020-06-24", "url": "https://arxiv.org/abs/2006.14091", "pdf_url": "https://arxiv.org/pdf/2006.14091v2", "arxiv_id": "2006.14091", "doi": "10.1177/02783649211041652", "citation_count": 146, "influential_citation_count": 8, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.5418}
{"id": "bf037a65038a94fe95bdbf251ca7ec71e5414d8d0e57514d80ae52f16f21977b", "sources": ["arxiv", "semantic_scholar"], "title": "Risk-Sensitive Reinforcement Learning: a Martingale Approach to Reward Uncertainty", "abstract": "We introduce a novel framework to account for sensitivity to rewards uncertainty in sequential decision-making problems. While risk-sensitive formulations for Markov decision processes studied so far focus on the distribution of the cumulative reward as a whole, we aim at learning policies sensitive to the uncertain/stochastic nature of the rewards, which has the advantage of being conceptually more meaningful in some cases. To this end, we present a new decomposition of the randomness contained in the cumulative reward based on the Doob decomposition of a stochastic process, and introduce a new conceptual tool - the \\textit{chaotic variation} - which can rigorously be interpreted as the risk measure of the martingale component associated to the cumulative reward process. We innovate on the reinforcement learning side by incorporating this new risk-sensitive approach into model-free algorithms, both policy gradient and value function based, and illustrate its relevance on grid world and portfolio optimization problems.", "authors": ["Nelson Vadori", "Sumitra Ganesh", "Prashant Reddy", "Manuela Veloso"], "categories": ["cs.LG", "q-fin.RM", "stat.ML"], "fields_of_study": ["Computer Science", "Economics", "Mathematics"], "published_date": "2020-06-23", "url": "https://arxiv.org/abs/2006.12686", "pdf_url": "https://arxiv.org/pdf/2006.12686v2", "arxiv_id": "2006.12686", "doi": "10.1145/3383455.3422519", "citation_count": 12, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "International Conference on AI in Finance", "quality_score": 0.2785}
{"id": "8b18f8342b86edba9546c3e469db4ad11ff8c97fa2fd6ca3a7fbafbb420eedd7", "sources": ["arxiv", "semantic_scholar"], "title": "On Reward-Free Reinforcement Learning with Linear Function Approximation", "abstract": "Reward-free reinforcement learning (RL) is a framework which is suitable for both the batch RL setting and the setting where there are many reward functions of interest. During the exploration phase, an agent collects samples without using a pre-specified reward function. After the exploration phase, a reward function is given, and the agent uses samples collected during the exploration phase to compute a near-optimal policy. Jin et al. [2020] showed that in the tabular setting, the agent only needs to collect polynomial number of samples (in terms of the number states, the number of actions, and the planning horizon) for reward-free RL. However, in practice, the number of states and actions can be large, and thus function approximation schemes are required for generalization. In this work, we give both positive and negative results for reward-free RL with linear function approximation. We give an algorithm for reward-free RL in the linear Markov decision process setting where both the transition and the reward admit linear representations. The sample complexity of our algorithm is polynomial in the feature dimension and the planning horizon, and is completely independent of the number of states and actions. We further give an exponential lower bound for reward-free RL in the setting where only the optimal $Q$-function admits a linear representation. Our results imply several interesting exponential separations on the sample complexity of reward-free RL.", "authors": ["Ruosong Wang", "Simon S. Du", "Lin F. Yang", "Ruslan Salakhutdinov"], "categories": ["cs.LG", "cs.AI", "math.OC", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-06-19", "url": "https://arxiv.org/abs/2006.11274", "pdf_url": "https://arxiv.org/pdf/2006.11274v1", "arxiv_id": "2006.11274", "doi": null, "citation_count": 117, "influential_citation_count": 25, "has_code": false, "code_url": null, "venue": "Neural Information Processing Systems", "quality_score": 0.7075}
{"id": "88b561f3983f6f059af200b1e6819e1c2b7dcdb8bcd59a4dc018c3c625cdd9d6", "sources": ["arxiv", "semantic_scholar"], "title": "Wasserstein Distance guided Adversarial Imitation Learning with Reward Shape Exploration", "abstract": "The generative adversarial imitation learning (GAIL) has provided an adversarial learning framework for imitating expert policy from demonstrations in high-dimensional continuous tasks. However, almost all GAIL and its extensions only design a kind of reward function of logarithmic form in the adversarial training strategy with the Jensen-Shannon (JS) divergence for all complex environments. The fixed logarithmic type of reward function may be difficult to solve all complex tasks, and the vanishing gradients problem caused by the JS divergence will harm the adversarial learning process. In this paper, we propose a new algorithm named Wasserstein Distance guided Adversarial Imitation Learning (WDAIL) for promoting the performance of imitation learning (IL). There are three improvements in our method: (a) introducing the Wasserstein distance to obtain more appropriate measure in the adversarial training process, (b) using proximal policy optimization (PPO) in the reinforcement learning stage which is much simpler to implement and makes the algorithm more efficient, and (c) exploring different reward function shapes to suit different tasks for improving the performance. The experiment results show that the learning procedure remains remarkably stable, and achieves significant performance in the complex continuous control tasks of MuJoCo.", "authors": ["Ming Zhang", "Yawei Wang", "Xiaoteng Ma", "Li Xia", "Jun Yang", "Zhiheng Li", "Xiu Li"], "categories": ["cs.LG", "cs.RO", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-06-05", "url": "https://arxiv.org/abs/2006.03503", "pdf_url": "https://arxiv.org/pdf/2006.03503v2", "arxiv_id": "2006.03503", "doi": "10.1109/DDCLS49620.2020.9275169", "citation_count": 23, "influential_citation_count": 3, "has_code": false, "code_url": null, "venue": "2020 IEEE 9th Data Driven Control and Learning Systems Conference (DDCLS)", "quality_score": 0.3451}
{"id": "6b81b9fc32e023992efe46a18d3dc0cad893f03d030fdb4ee76886fd295aa067", "sources": ["arxiv", "semantic_scholar"], "title": "Active Preference-Based Gaussian Process Regression for Reward Learning", "abstract": "Designing reward functions is a challenging problem in AI and robotics. Humans usually have a difficult time directly specifying all the desirable behaviors that a robot needs to optimize. One common approach is to learn reward functions from collected expert demonstrations. However, learning reward functions from demonstrations introduces many challenges: some methods require highly structured models, e.g. reward functions that are linear in some predefined set of features, while others adopt less structured reward functions that on the other hand require tremendous amount of data. In addition, humans tend to have a difficult time providing demonstrations on robots with high degrees of freedom, or even quantifying reward values for given demonstrations. To address these challenges, we present a preference-based learning approach, where as an alternative, the human feedback is only in the form of comparisons between trajectories. Furthermore, we do not assume highly constrained structures on the reward function. Instead, we model the reward function using a Gaussian Process (GP) and propose a mathematical formulation to actively find a GP using only human preferences. Our approach enables us to tackle both inflexibility and data-inefficiency problems within a preference-based learning framework. Our results in simulations and a user study suggest that our approach can efficiently learn expressive reward functions for robotics tasks.", "authors": ["Erdem Bıyık", "Nicolas Huynh", "Mykel J. Kochenderfer", "Dorsa Sadigh"], "categories": ["cs.RO", "cs.AI", "cs.LG"], "fields_of_study": ["Computer Science"], "published_date": "2020-05-06", "url": "https://arxiv.org/abs/2005.02575", "pdf_url": "https://arxiv.org/pdf/2005.02575v2", "arxiv_id": "2005.02575", "doi": "10.1177/02783649231208729", "citation_count": 142, "influential_citation_count": 4, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.5388}
{"id": "1cb61e459114446999fa2ab8d5ae87d9878c4353385696ea11847e5c563063dc", "sources": ["arxiv", "semantic_scholar"], "title": "Self Punishment and Reward Backfill for Deep Q-Learning", "abstract": "Reinforcement learning agents learn by encouraging behaviours which maximize their total reward, usually provided by the environment. In many environments, however, the reward is provided after a series of actions rather than each single action, leading the agent to experience ambiguity in terms of whether those actions are effective, an issue known as the credit assignment problem. In this paper, we propose two strategies inspired by behavioural psychology to enable the agent to intrinsically estimate more informative reward values for actions with no reward. The first strategy, called self-punishment (SP), discourages the agent from making mistakes that lead to undesirable terminal states. The second strategy, called the rewards backfill (RB), backpropagates the rewards between two rewarded actions. We prove that, under certain assumptions and regardless of the reinforcement learning algorithm used, these two strategies maintain the order of policies in the space of all possible policies in terms of their total reward, and, by extension, maintain the optimal policy. Hence, our proposed strategies integrate with any reinforcement learning algorithm that learns a value or action-value function through experience. We incorporated these two strategies into three popular deep reinforcement learning approaches and evaluated the results on thirty Atari games. After parameter tuning, our results indicate that the proposed strategies improve the tested methods in over 65 percent of tested games by up to over 25 times performance improvement.", "authors": ["Mohammad Reza Bonyadi", "Rui Wang", "Maryam Ziaei"], "categories": ["cs.AI", "cs.LG"], "fields_of_study": ["Medicine", "Computer Science"], "published_date": "2020-04-10", "url": "https://arxiv.org/abs/2004.05002", "pdf_url": "https://arxiv.org/pdf/2004.05002v2", "arxiv_id": "2004.05002", "doi": "10.1109/TNNLS.2021.3140042", "citation_count": 8, "influential_citation_count": 1, "has_code": false, "code_url": null, "venue": "IEEE Transactions on Neural Networks and Learning Systems", "quality_score": 0.2386}
{"id": "becdae527ce6926838ec21a3512e11e58c55271e801b76da641e023d10856480", "sources": ["arxiv", "semantic_scholar"], "title": "Balance Between Efficient and Effective Learning: Dense2Sparse Reward Shaping for Robot Manipulation with Environment Uncertainty", "abstract": "Efficient and effective learning is one of the ultimate goals of the deep reinforcement learning (DRL), although the compromise has been made in most of the time, especially for the application of robot manipulations. Learning is always expensive for robot manipulation tasks and the learning effectiveness could be affected by the system uncertainty. In order to solve above challenges, in this study, we proposed a simple but powerful reward shaping method, namely Dense2Sparse. It combines the advantage of fast convergence of dense reward and the noise isolation of the sparse reward, to achieve a balance between learning efficiency and effectiveness, which makes it suitable for robot manipulation tasks. We evaluated our Dense2Sparse method with a series of ablation experiments using the state representation model with system uncertainty. The experiment results show that the Dense2Sparse method obtained higher expected reward compared with the ones using standalone dense reward or sparse reward, and it also has a superior tolerance of system uncertainty.", "authors": ["Yongle Luo", "Kun Dong", "Lili Zhao", "Zhiyong Sun", "Chao Zhou", "Bo Song"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-03-05", "url": "https://arxiv.org/abs/2003.02740", "pdf_url": "https://arxiv.org/pdf/2003.02740v1", "arxiv_id": "2003.02740", "doi": "10.1109/AIM52237.2022.9863259", "citation_count": 18, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": null, "quality_score": 0.3197}
{"id": "7384147f6ea9e71c82ca37c53b4901822d761480857adb7a56f1ad0a032565a8", "sources": ["arxiv", "semantic_scholar"], "title": "Reward Shaping for Human Learning via Inverse Reinforcement Learning", "abstract": "Humans are spectacular reinforcement learners, constantly learning from and adjusting to experience and feedback. Unfortunately, this doesn't necessarily mean humans are fast learners. When tasks are challenging, learning can become unacceptably slow. Fortunately, humans do not have to learn tabula rasa, and learning speed can be greatly increased with learning aids. In this work we validate a new type of learning aid -- reward shaping for humans via inverse reinforcement learning (IRL). The goal of this aid is to increase the speed with which humans can learn good policies for specific tasks. Furthermore this approach compliments alternative machine learning techniques such as safety features that try to prevent individuals from making poor decisions. To achieve our results we first extend a well known IRL algorithm via kernel methods. Afterwards we conduct two human subjects experiments using an online game where players have limited time to learn a good policy. We show with statistical significance that players who receive our learning aid are able to approach desired policies more quickly than the control group.", "authors": ["Mark A. Rucker", "Layne T. Watson", "Matthew S. Gerber", "Laura E. Barnes"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-02-25", "url": "https://arxiv.org/abs/2002.10904", "pdf_url": "https://arxiv.org/pdf/2002.10904v3", "arxiv_id": "2002.10904", "doi": null, "citation_count": 2, "influential_citation_count": 0, "has_code": true, "code_url": "https://github.com/mrucker/kpirl-kla", "venue": null, "quality_score": 0.1193}
{"id": "599ca7b5baecef4cb247ec453c0ca56027b9ff475e2db011a91a94e5e0326120", "sources": ["arxiv", "semantic_scholar"], "title": "Safe Imitation Learning via Fast Bayesian Reward Inference from Preferences", "abstract": "Bayesian reward learning from demonstrations enables rigorous safety and uncertainty analysis when performing imitation learning. However, Bayesian reward learning methods are typically computationally intractable for complex control problems. We propose Bayesian Reward Extrapolation (Bayesian REX), a highly efficient Bayesian reward learning algorithm that scales to high-dimensional imitation learning problems by pre-training a low-dimensional feature encoding via self-supervised tasks and then leveraging preferences over demonstrations to perform fast Bayesian inference. Bayesian REX can learn to play Atari games from demonstrations, without access to the game score and can generate 100,000 samples from the posterior over reward functions in only 5 minutes on a personal laptop. Bayesian REX also results in imitation learning performance that is competitive with or better than state-of-the-art methods that only learn point estimates of the reward function. Finally, Bayesian REX enables efficient high-confidence policy evaluation without having access to samples of the reward function. These high-confidence performance bounds can be used to rank the performance and risk of a variety of evaluation policies and provide a way to detect reward hacking behaviors.", "authors": ["Daniel S. Brown", "Russell Coleman", "Ravi Srinivasan", "Scott Niekum"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-02-21", "url": "https://arxiv.org/abs/2002.09089", "pdf_url": "https://arxiv.org/pdf/2002.09089v4", "arxiv_id": "2002.09089", "doi": null, "citation_count": 115, "influential_citation_count": 6, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.5161}
{"id": "1224b7b8a4b5021a0192bb8ef2fb60065ad1ffe2fcd3fc7e7897e1db7a5b45c8", "sources": ["arxiv", "semantic_scholar"], "title": "HMRL: Hyper-Meta Learning for Sparse Reward Reinforcement Learning Problem", "abstract": "In spite of the success of existing meta reinforcement learning methods, they still have difficulty in learning a meta policy effectively for RL problems with sparse reward. In this respect, we develop a novel meta reinforcement learning framework called Hyper-Meta RL(HMRL), for sparse reward RL problems. It is consisted with three modules including the cross-environment meta state embedding module which constructs a common meta state space to adapt to different environments; the meta state based environment-specific meta reward shaping which effectively extends the original sparse reward trajectory by cross-environmental knowledge complementarity and as a consequence the meta policy achieves better generalization and efficiency with the shaped meta reward. Experiments with sparse-reward environments show the superiority of HMRL on both transferability and policy learning efficiency.", "authors": ["Yun Hua", "Xiangfeng Wang", "Bo Jin", "Wenhao Li", "Junchi Yan", "Xiaofeng He", "Hongyuan Zha"], "categories": ["cs.LG", "cs.AI", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-02-11", "url": "https://arxiv.org/abs/2002.04238", "pdf_url": "https://arxiv.org/pdf/2002.04238v2", "arxiv_id": "2002.04238", "doi": "10.1145/3447548.3467242", "citation_count": 10, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "Knowledge Discovery and Data Mining", "quality_score": 0.2603}
{"id": "d00572c4774df6d9c0c8be556395aaf4a453d87be05bd7772003845f87fb306d", "sources": ["arxiv", "semantic_scholar"], "title": "Reward-Free Exploration for Reinforcement Learning", "abstract": "Exploration is widely regarded as one of the most challenging aspects of reinforcement learning (RL), with many naive approaches succumbing to exponential sample complexity. To isolate the challenges of exploration, we propose a new \"reward-free RL\" framework. In the exploration phase, the agent first collects trajectories from an MDP $\\mathcal{M}$ without a pre-specified reward function. After exploration, it is tasked with computing near-optimal policies under for $\\mathcal{M}$ for a collection of given reward functions. This framework is particularly suitable when there are many reward functions of interest, or when the reward function is shaped by an external agent to elicit desired behavior. We give an efficient algorithm that conducts $\\tilde{\\mathcal{O}}(S^2A\\mathrm{poly}(H)/ε^2)$ episodes of exploration and returns $ε$-suboptimal policies for an arbitrary number of reward functions. We achieve this by finding exploratory policies that visit each \"significant\" state with probability proportional to its maximum visitation probability under any possible policy. Moreover, our planning procedure can be instantiated by any black-box approximate planner, such as value iteration or natural policy gradient. We also give a nearly-matching $Ω(S^2AH^2/ε^2)$ lower bound, demonstrating the near-optimality of our algorithm in this setting.", "authors": ["Chi Jin", "Akshay Krishnamurthy", "Max Simchowitz", "Tiancheng Yu"], "categories": ["cs.LG", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-02-07", "url": "https://arxiv.org/abs/2002.02794", "pdf_url": "https://arxiv.org/pdf/2002.02794v1", "arxiv_id": "2002.02794", "doi": null, "citation_count": 229, "influential_citation_count": 68, "has_code": false, "code_url": null, "venue": "International Conference on Machine Learning", "quality_score": 0.9194}
{"id": "24b39fc1f90889cee2e68f17969e754d18c6945e070068fc67e858b34eccdf13", "sources": ["arxiv", "semantic_scholar"], "title": "Effects of sparse rewards of different magnitudes in the speed of learning of model-based actor critic methods", "abstract": "Actor critic methods with sparse rewards in model-based deep reinforcement learning typically require a deterministic binary reward function that reflects only two possible outcomes: if, for each step, the goal has been achieved or not. Our hypothesis is that we can influence an agent to learn faster by applying an external environmental pressure during training, which adversely impacts its ability to get higher rewards. As such, we deviate from the classical paradigm of sparse rewards and add a uniformly sampled reward value to the baseline reward to show that (1) sample efficiency of the training process can be correlated to the adversity experienced during training, (2) it is possible to achieve higher performance in less time and with less resources, (3) we can reduce the performance variability experienced seed over seed, (4) there is a maximum point after which more pressure will not generate better results, and (5) that random positive incentives have an adverse effect when using a negative reward strategy, making an agent under those conditions learn poorly and more slowly. These results have been shown to be valid for Deep Deterministic Policy Gradients using Hindsight Experience Replay in a well known Mujoco environment, but we argue that they could be generalized to other methods and environments as well.", "authors": ["Juan Vargas", "Lazar Andjelic", "Amir Barati Farimani"], "categories": ["cs.LG", "cs.RO", "stat.ML"], "fields_of_study": ["Computer Science", "Mathematics"], "published_date": "2020-01-18", "url": "https://arxiv.org/abs/2001.06725", "pdf_url": "https://arxiv.org/pdf/2001.06725v1", "arxiv_id": "2001.06725", "doi": null, "citation_count": 1, "influential_citation_count": 0, "has_code": false, "code_url": null, "venue": "arXiv.org", "quality_score": 0.0753}