text string | source string |
|---|---|
for SiDyP on all six Llama-3 generated datasets. LLM(→) Llama-3-70b Datasets (→) NumClaim TREC SemEval 20News Method (↓) Zero-shot Few-shot Zero-shot Few-shot Zero-shot Few-shot Zero-shot Noise Ratio (Original) 91.69 95.85 70.35 69.72 50.96 50.64 76.13 No Answer Ratio 0.00 0.00 3.6 𝑒−41.8𝑒−42.5𝑒−34.1𝑒−31.4𝑒−2 Nois... | https://arxiv.org/abs/2505.19675v1 |
Grounding Language with Vision: A Conditional Mutual Information Calibrated Decoding Strategy for Reducing Hallucinations in LVLMs Hao Fang∗1, Changle Zhou∗2, Jiawei Kong∗1, 2, Kuofeng Gao1, Bin Chen†2, 3,Tao Liang4,Guojun Ma4,Shu-Tao Xia1, 3, 1Tsinghua Shenzhen Internation Graduate School, Tsinghua University, 2Harbin... | https://arxiv.org/abs/2505.19678v1 |
promotes more reliable generation by adaptively retaining image tokens with high relevance to the ongoing response. grounded in empirical findings and lack convincing theoretical foundations. Moreover, they generally fail to explicitly quantify and control the dynamic mutual relevance between the visual input and the p... | https://arxiv.org/abs/2505.19678v1 |
visual un- derstanding tasks. To incorporate visual information into the LLM backbone, LVLMs like LLaV A [2,4] and Shikra [ 3] employ linear projection layers trained by instruction fine-tuning to directly map visual features into the LLM embedding space. Meanwhile, the BLIP series [ 25,28] introduces Q-former to integ... | https://arxiv.org/abs/2505.19678v1 |
v N}. At decoding step t, the visual tokens are concatenated with the textual tokens from the instruction xand the previously generated token sequence y<t. The resulting sequence is fed into the LLM backbone of the LVLM to autoregressively predict the next token: yt∼pθ(· |v, x, y <t) = softmax( fθ(· |v, x, y <t)), (1) ... | https://arxiv.org/abs/2505.19678v1 |
This formula decomposes the original objective over individual decoding steps, enabling each term to be explicitly computed using the token-level probabilities provided by the LVLM. An interesting observation is that the token distributions used in existing contrastive decoding studies [ 17,16,8,9] can be viewed as spe... | https://arxiv.org/abs/2505.19678v1 |
solve the lower subproblem, we then adaptively retain a proportion γof image tokens as the purified input to promote its relevance to the current textual context. Motivated by findings in [ 35] that token sparsification at the second layer of LVLMs yields optimal performance, we utilize attention scores from this layer... | https://arxiv.org/abs/2505.19678v1 |
introducing an additional network, as the purifier module is lightweight and the removal of non-essential visual tokens helps reduce the overall inference cost (see Sec . 4.2). 4 Experiments We conduct experiments on different LVLMs across various benchmarks to show the effectiveness of our method. For more experimenta... | https://arxiv.org/abs/2505.19678v1 |
Ratio (SHR); (3) Polling-based Object Probing Evaluation (POPE) [ 41], another object hallucination evaluation also conducted on MSCOCO; (4) Multimodal Large Language Model Evaluation (MME) [42], a general-purpose benchmark for assessing multimodal capabilities; and (5) MMBench [ 43], which includes multiple-choice que... | https://arxiv.org/abs/2505.19678v1 |
decoding strategies in some cases. In contrast, the proposed CMI-VLD consistently reduces both sentence-level and instance-level object hallucinations in the final responses. GPT-4o Assisted Evaluation. While CHAIR is a reliable evaluation metric widely adopted in previous studies, it is limited within the scope of obj... | https://arxiv.org/abs/2505.19678v1 |
hallucination mitigation. Inference Time Analysis. Since our method introduces an additional visual token purifier for effective visual refinement, it is crucial to assess its influence on the overall computational efficiency. Following [ 19], we calculate the generation time per response to assess computing burdens. F... | https://arxiv.org/abs/2505.19678v1 |
multi-modal large language models with behavioral planning states for autonomous driving,” arXiv preprint arXiv:2312.09245 , 2023. [6]C. Cui, Y . Ma, X. Cao, W. Ye, Y . Zhou, K. Liang, J. Chen, J. Lu, Z. Yang, K.-D. Liao et al. , “A survey on multimodal large language models for autonomous driving,” in Proceedings of t... | https://arxiv.org/abs/2505.19678v1 |
Pattern Recognition , 2024, pp. 13 418–13 427. 10 [19] T. Yang, Z. Li, J. Cao, and C. Xu, “Mitigating hallucination in large vision-language models via modular attribution and intervention,” in Adaptive Foundation Models: Evolving AI for Personalized and Efficient Learning . [20] A. Favero, L. Zancato, M. Trager, S. Ch... | https://arxiv.org/abs/2505.19678v1 |
Y . Rao, W. Zhao, B. Liu, J. Lu, J. Zhou, and C.-J. Hsieh, “Dynamicvit: Efficient vision transformers with dynamic token sparsification,” Advances in neural information processing systems , vol. 34, pp. 13 937–13 949, 2021. [35] L. Chen, H. Zhao, T. Liu, S. Bai, J. Lin, C. Zhou, and B. Chang, “An image is worth 1/2 tok... | https://arxiv.org/abs/2505.19678v1 |
pθ(v, x)pθ(y|x)(13) = max v,ylogpθ(y|v, x) pθ(y|x)(14) = max v,ylogQl−1 t=0pθ(yt|v, x, y <t) Ql−1 t=0pθ(yt|x, y<t))(15) = max v,ylogl−1Y t=0pθ(yt|v, x, y <t)−logl−1Y t=0pθ(yt|x, y<t)) (16) = max v,yl−1X t=0[logpθ(yt|v, x, y <t)−logpθ(yt|x, y<t)]. (17) B Experimental Details B.1 Implementation Details Throughout our exp... | https://arxiv.org/abs/2505.19678v1 |
stronger performance in mitigating object hallucinations. Due to the concise format of POPE responses, which are typically short declarative sentences, the benchmark primarily reflects the visual grounding ability of a model rather than its long-form generation capacity. GPT-4 Assisted Evaluations. In addition to objec... | https://arxiv.org/abs/2505.19678v1 |
γ. D More Experimental Results POPE evaluation on more LVLMs. We supplement the results of POPE metrics on more LVLMs, including LLaV A-1.5 and Shikra. The quantitative results in Table 4 again confirm the effectiveness of our method in mitigating object hallucinations. Ablation study of vision-language decoding. We th... | https://arxiv.org/abs/2505.19678v1 |
view of a delicious-looking sandwich, cut in half and placed on a dining table. The sandwich is filled with meat and accompanied by a leafy green garnish, adding a touch of freshness to the dish. The dining table is set with various items, including a wine glass, a cup, a fork, and a spoon. The wine glass is positioned... | https://arxiv.org/abs/2505.19678v1 |
sky, adding a sense of liveliness to the scene. The birds are scattered throughout the image, with some closer to the train and others further away. The combination of the train, the bridge, and the birds creates a dynamic and lively atmosphere. SID: The scene features a train crossing over a red brick bridge that span... | https://arxiv.org/abs/2505.19678v1 |
to the room. VASparse: The image depicts a kitchen with a stainless steel sink, a refrigerator on the left side, and a microwave on the countertop. Above the sink, there is a projection screen displaying an image of a house. The kitchen is equipped with various appliances, including a toaster, a bottle, a bowl, and a s... | https://arxiv.org/abs/2505.19678v1 |
arXiv:2505.19679v1 [cs.CL] 26 May 2025KIT’s Low-resource Speech Translation Systems for IWSLT20 25: System Enhancement with Synthetic Data and Model Regularization Zhaolin Li, Yining Liu, Danni Liu, Tuan Nam Nguyen, Enes Yavu z Ugan, Tu Anh Dinh, Carlos Mullov, Alexander Waibel, Jan Niehues Karlsruhe Institute of Techn... | https://arxiv.org/abs/2505.19679v1 |
imbalanced parameter usage ( Rom- ney Robinson et al. ,2024 ;Jiawei et al. ,2024 ). However, these works are limited to MT models in the cascaded system. Since model regulariza- tion is a generic approach, this work investigates its effectiveness with both ASR, MT, and ST tasks. With experimental results across differe... | https://arxiv.org/abs/2505.19679v1 |
modalities: the MT-augmented method, which generates synthetic translations from ASR data, and the TTS-augmented method, which produces synthetic speech from MT data. Together, these methods aim to enhance the qual- ity and robustness of ST models in low-resource settings. 3.2 Model Regularization Regularization remain... | https://arxiv.org/abs/2505.19679v1 |
of languages, particularly many low-resource ones, making it well-suited for low-resource translation tasks. North Levantine and Tunisian are included in its pre-training, and Bemba is not. We use the 1.3B parameter version3, freezing the word embeddings to reduce memory usage. We also freeze the decoder except for the... | https://arxiv.org/abs/2505.19679v1 |
prompt, employing classifier-free guidance α= 2.0to strengthen prompt adherence. This ensures that the speaker distribution in the gener- ated data matches that of the original dataset. Ad- ditionally, we configure the numerical approxima- tion steps to 32 to ensure high-quality waveform generation.4.3.2 VITS VITS ( Kim ... | https://arxiv.org/abs/2505.19679v1 |
results are presented as BLEU/chrF. MMS, we observe that applying language model fusion with encoder-based models consistently im- proves ASR performance, resulting in a reduction of approximately 4 WER points—aligning with findings from prior work. Comparing A1 and A3, we observe that XEUS achieves performance sim- ila... | https://arxiv.org/abs/2505.19679v1 |
ID 47.9/70.1 16.1/42.8 19.6/39.4 22.7/43.5 B1 NLLB all 24.9/53.6 20.9/48.8 30.4/52.6 26.8/50.2 B2 B1 + transfer 31.3/57.6 28.0/54.4 30.3/52.2 26.3/49.9 B3 Seamless 21.7/48.2 18.9/45.1 28.4/50.8 25.6/48.9 C Best A+B 19.1/42.1 26.6/53.2 23.4/46.2 20.1/43.8 D1 Seamless 19.9/41.7 27.3/52.4 20.5/43.3 18.0/41.1 D2 Seamless +... | https://arxiv.org/abs/2505.19679v1 |
as appropriate text length, as outlined in Table 2. Evaluation results for the TTS systems are provided in Appendix A. We generate 120K synthetic training samples for each TTS model. This synthetic data is com- bined with the original development set for train- ing, while the validation split remains unchanged. Followi... | https://arxiv.org/abs/2505.19679v1 |
low-resource speech translation. 4.10 MBR Decoding We apply MBR decoding to the cascaded systems, the E2E systems, and their combination. As pre- sented in Tables 3and4, MBR decoding consis- tently yields minimal to no improvement when ap- plied to individual systems. In contrast, combin- ing the cascaded and E2E syste... | https://arxiv.org/abs/2505.19679v1 |
Heffernan, John Hoffman, and 1 others. 2023. Seamlessm4t: Mas- sively multilingual & multimodal machine transla- tion. arXiv preprint arXiv:2308.11596 . Waad Ben Kheder, Josef Jon, André Beyer, Abdel Mes- saoudi, Rabea Affan, Claude Barras, Maxim Ty- chonov, and Jean-Luc Gauvain. 2024. ALADAN at IWSLT24 low-resource Ar... | https://arxiv.org/abs/2505.19679v1 |
the 1st Workshop on Language Models for Underserved Communities (LM4UC 2025) , pages 1–7, Albuquerque, New Mex- ico. Association for Computational Linguistics. Zhaolin Li, Enes Yavuz Ugan, Danni Liu, Carlos Mullov, Tu Anh Dinh, Sai Koneru, Alexander Waibel, and Jan Niehues. 2024. The KIT speech translation systems for ... | https://arxiv.org/abs/2505.19679v1 |
final MCD metric is calculated us- ing the 1-25th coefficients (excluding the energy term) across DTW-aligned frames. Additionally, since MCD is not a speaker- independent metric like WER, to reduce the influ- ence of speaker attributes, we conducted assess- ments in both same-speaker (reconstruction) and cross-speaker se... | https://arxiv.org/abs/2505.19679v1 |
arXiv:2505.19683v1 [cs.AI] 26 May 2025Large Language Models for Planning: A Comprehensive and Systematic Survey PENGFEI CAO, The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, CASIA, and School of Artificial Intelligence, University of Chinese Academy of Sciences, China TIANYI MEN, The Key L... | https://arxiv.org/abs/2505.19683v1 |
Complex Systems, CASIA, and School of Artificial Intelligence, University of Chinese Academy of Sciences, China, kliu@nlpr.ia.ac.cn; Jun Zhao, The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, CASIA, and School of Artificial Intelligence, University of Chinese Academy of Sciences, China, jz... | https://arxiv.org/abs/2505.19683v1 |
agent planning to enlighten and guide researchers and practitioners. Figure 1 presents the overview of the LLM-based agent planning, which is organized into four closely related parts, involving planning definitions, planning methods, planning evaluation, as well as analysis and interpretation. The key contributions of... | https://arxiv.org/abs/2505.19683v1 |
or a Partially Observable Markov Decision Process (POMDP) [ 5,112]. An MDP is typically defined as a tuple (𝑆,𝐴,𝑇,𝑅,𝛾), where𝑆represents the state space, 𝐴denotes the action space, 𝑇 is the state transition function, defined as 𝑇:𝑆×𝐴→𝑆,𝑅is the reward function, defined as 𝑅:𝑆×𝐴→R, where R denotes the rew... | https://arxiv.org/abs/2505.19683v1 |
260],Liu et al. [ 148], Roy et al. [ 201],Predictive decoding [ 164],Con- trastive decoding [ 177],Grounded decoding [ 99],etc. Fig. 2. The typology of LLM-based planning methods, which includes external module augmented methods, finetuning-based methods, and searching-based methods. In closed-loop planning, the proces... | https://arxiv.org/abs/2505.19683v1 |
prompts an LLM to construct initial PDDL models and then corrects errors based on external feedback sources, including PDDL model validation tools and human domain experts. The refined PDDL can then be passed to either a classical planner or an LLM-modulo planner to produce the final plan. PDDLEGO [ 322] designs an ite... | https://arxiv.org/abs/2505.19683v1 |
from historical states or the environment [ 252,279]. Several studies leverage memory to enhance the planning capabilities of LLMs [116, 184, 258, 262, 304]. Specifically, Zhong et al. [ 341] propose a memory mechanism called MemoryBank, which stores and recalls historical interactions to enhance performance in long-te... | https://arxiv.org/abs/2505.19683v1 |
dependency pairs to recall past observations. Meanwhile, they also propose a method called AttentionTuner to incorporate the memory dependency pairs into self-attention modules. Summary of External Module Augmented Methods •Planner Enhanced Methods first usually generate executable codes and then employ an external pla... | https://arxiv.org/abs/2505.19683v1 |
[ 32] consider integrating agent capabilities into general-purpose LLMs to bridge the gap between specialized and general-purpose models. It aligns training corpora with chat formats and introduces an innovative method called Agent-FLAN, which effectively integrates agent capabilities into LLMs. Chen et al. [ 22] propo... | https://arxiv.org/abs/2505.19683v1 |
computational costs [ 224,288]. Su et al. [ 224] employ Dualformer, which integrates fast and slow reasoning modes within a Transformer framework. This model is trained using complete inference trajectories automatically generated by the A* search algorithm, providing faster reasoning while maintaining accuracy. Simila... | https://arxiv.org/abs/2505.19683v1 |
for guidance. 3.2.2 Feedback-based Methods. This section explores methods for optimizing the planning abilities of LLMs by introducing feedback mechanisms, as shown in Figure 5. Specifically, we classify feedback into three categories based on its source: Environmental Feedback ,Reward Model Feedback , as well as Self-... | https://arxiv.org/abs/2505.19683v1 |
LLM-generated responses, guiding its self-correction behavior and significantly improving performance in mathematical and programming tasks. This result-based feedback directly correlates with task goals and can effectively improve the model’s final performance. However, its limitation lies in the binary nature of feed... | https://arxiv.org/abs/2505.19683v1 |
common type of feedback is self-generation and adversarial training, as seen in SPIN [ 30], which uses a self- adversarial mechanism to help LLMs grow from weak models to strong ones. This method uses data generated by the model itself as feedback to optimize its performance, without the need for additional manual anno... | https://arxiv.org/abs/2505.19683v1 |
Decomposition-based Methods. In planning tasks, the decomposition strategy plays a pivotal role. Complex planning problems often require deep reasoning. To address this, many studies have explored the use of decomposition strategies, as illustrated in Figure 6. This approach first decomposes a complex task into several... | https://arxiv.org/abs/2505.19683v1 |
[ 311] decompose the constrained language planning task into two steps, in which if the generated specific goals fail to produce an executable script, the goal-generation step is modified based on feedback from the script-generation attempt. Similarly, RaDA [ 118] advances this concept by dividing planning into two key... | https://arxiv.org/abs/2505.19683v1 |
enabling them to backtrack within the solution space to identify optimal outcomes—akin to the workings of “System 2” thinking [ 113,217,221], as shown in Figure 7. Within this paradigm, the design of an accurate and efficient exploration strategy emerges as a key bottleneck in planning. Consequently, researchers have p... | https://arxiv.org/abs/2505.19683v1 |
tree, and graph structures, it enables efficient allocation of computational resources. MCTS-based Exploration: Monte Carlo Tree Search (MCTS) [ 44,119] is an advanced tree search algorithm that balances exploration and exploitation with the goal of finding the optimal trajectory. It comprises four key steps: 1) Select... | https://arxiv.org/abs/2505.19683v1 |
planning search framework that combines LLM global planning and A* local search. It uses LLM for high-level planning and A* for precise low-level path searching, aiming to improve search 18 Cao et al. efficiency. Lehnert et al. [ 127] propose using A* search to construct trajectory data, applied to path planning tasks.... | https://arxiv.org/abs/2505.19683v1 |
through stochastic beam search. This approach leverages LLMs to provide a calibrated criterion for reasoning. Additionally, it combines beam search with temperature-controlled randomness to generate more robust reasoning chains. In a similar vein, Wang et al. [ 260] explore whether LLMs inherently possess reasoning cap... | https://arxiv.org/abs/2505.19683v1 |
instructions. Agents perceive the environment through screen captures or HTML, and translate the instructions into interactive operations such as mouse clicks and keyboard inputs. These tasks span a variety of scenarios, including web browsing, mobile applications, file management, and enterprise software usage. Based ... | https://arxiv.org/abs/2505.19683v1 |
Reward Watch-And-Help[188] ✓ ✓ POMDP closed action Success Rate, Speedup, Reward LangSuit·E[103] ✓ × POMDP closed action Success Rate, Fixed Strict Rate, Accuracy, etc ActPlan-1K[226] × ✓ MDP open plan Longest Common Subsequence, etc PARTNR[20] ✓ ✓ POMDP closed action Simulation Steps, Success Rate, etc Embodied Agent ... | https://arxiv.org/abs/2505.19683v1 |
ISG-BENCH[23] × ✓ MDP open plan BertScore, Human Agreement Tool CallingToolBench[195] ✓ × POMDP closed action Pass Rate, Win Rate AppWorld[237] ✓ × POMDP closed code Task Goal Completion, etc API-Bank[137] × × MDP open action Accuracy, ROUGE-L ToolComp[176] ✓ × POMDP closed action LLM Grading, Exact Match ToolSandbox[1... | https://arxiv.org/abs/2505.19683v1 |
a home environment to perform tasks like cleaning and finding objects. Evaluation platforms for household robots characterized by dynamic home settings with partially observable conditions, requiring processing of information from multiple sensors and execution of complex action sequences in highly variable scenes that... | https://arxiv.org/abs/2505.19683v1 |
scenarios is to provide assistance through agents in complex daily tasks, such as travel planning, workflow arrangement, and tool invocation, in order to reduce user burden and enhance work efficiency. This mode is similar to Copilot, capable of replacing manual task arrangements, optimizing time Large Language Models ... | https://arxiv.org/abs/2505.19683v1 |
in plain text, without involving multimodal content. Furthermore, the problem setup often conforms to a MDP framework, as the generation of each code fragment depends primarily on the current state of the program. Code generation tasks place particular emphasis on an understanding of software engineering principles, th... | https://arxiv.org/abs/2505.19683v1 |
and structure biological experimental procedures. These tasks are characterized by clearly defined steps and result-oriented objectives. Typically, they do not require dynamic, interactive environments, nor do they involve multimodal inputs. Task execution predominantly follows a one-off, sequential decision-making pro... | https://arxiv.org/abs/2505.19683v1 |
Planning Correctness and Accuracy Metrics. The goal of correctness is to evaluate whether the agent can complete the task accurately and without errors, serving as the most fundamental evaluation dimensions. Common metrics is Success Rate (SR). SR can be divided into Response-Based Success Rate ,Action-Based Success Ra... | https://arxiv.org/abs/2505.19683v1 |
planning process aims to ensure that generated plans are not only compliant with user instructions but also aligned with environmental constraints, ensuring practical feasibility in real-world scenarios. This evaluation can be systematically divided into Large Language Models for Planning: A Comprehensive and Systemati... | https://arxiv.org/abs/2505.19683v1 |
abilities [ 241], NASA TLX is employed to quantitatively evaluate user cognitive load during interaction with the system. Task Coordination Metrics measure the effectiveness of role distribution between human and system during joint task execution. PARTNR [ 20] introduces a Task offloading metric, evaluating how effici... | https://arxiv.org/abs/2505.19683v1 |
complex datasets such as Mind2Web, with step success rates typically less than 65%. This indicates that the planning task is very promising and more effective planning methods need to be designed for more complex tasks. (2) Among various planning tasks, fine-tuning based methods have become the mainstream approach. Fro... | https://arxiv.org/abs/2505.19683v1 |
LLMs are closely related to planning. For example, the excellent Chain-of-Thought (CoT) capability of LLMs supports them in completing complex planning tasks, while the mechanism behind CoT remains unclear. To this end, some researchers attempt to analyze the impact of CoT on planning. Sprague et al. [ 220] attempt to ... | https://arxiv.org/abs/2505.19683v1 |
external interpretation investigate the factors influencing planning, and also analyze the role of the inherent capabilities of LLMs in planning. Although these analyses can provide guidance for improving the planning ability of LLMs, they cannot reveal the internal mechanisms. •Research on internal interpretation atte... | https://arxiv.org/abs/2505.19683v1 |
deployed on resource-constrained edge devices such as mobile phones, drones, and robotic platforms. However, downsizing models often leads to a significant decline in planning performance. Key challenges include reducing model size and computational overhead while preserving robust planning capabilities and ensuring ad... | https://arxiv.org/abs/2505.19683v1 |
large language models for automated planning. arXiv preprint arXiv:2502.12435 (2025). [4] Wasi Uddin Ahmad, Sean Narenthiran, Somshubra Majumdar, Aleksander Ficek, Siddhartha Jain, Jocelyn Huang, Vahid Noroozi, and Boris Ginsburg. 2025. OpenCodeReasoning: Advancing Data Distillation for Competitive Coding. arXiv prepri... | https://arxiv.org/abs/2505.19683v1 |
Leon Maksin, Tejal Patwardhan, et al. 2024. Mle-bench: Evaluating machine learning agents on machine learning engineering. arXiv preprint arXiv:2410.07095 (2024). [20] Matthew Chang, Gunjan Chhablani, Alexander Clegg, Mikael Dallaire Cote, Ruta Desai, Michal Hlavac, Vladimir Karashchuk, Jacob Krantz, Roozbeh Mottaghi, ... | https://arxiv.org/abs/2505.19683v1 |
Tianhao Hu, Han Xu, Zhisong Zhang, Yong Dai, Lei Han, and Nan Du. 2024. Self-playing Adversarial Language Game Enhances LLM Reasoning. arXiv preprint arXiv:2404.10642 (2024). [36] Maxime Chevalier-Boisvert, Dzmitry Bahdanau, Salem Lahlou, Lucas Willems, Chitwan Saharia, Thien Huu Nguyen, and Yoshua Bengio. 2018. Babyai... | https://arxiv.org/abs/2505.19683v1 |
arXiv:2408.13184 (2024). [52] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. 2023. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems 36 (2023), 28091–28114. [53] Yifeng Ding, Hantian Ding, Shiqi Wang, Qing Sun, Varun Kumar, a... | https://arxiv.org/abs/2505.19683v1 |
The Twelfth International Conference on Learning Representations . [68] Kanishk Gandhi, Denise Lee, Gabriel Grand, Muxin Liu, Winson Cheng, Archit Sharma, and Noah D Goodman. 2024. Stream of Search (SoS): Learning to Search in Language. arXiv preprint arXiv:2404.03683 (2024). [69] Longxi Gao, Li Zhang, and Mengwei Xu. ... | https://arxiv.org/abs/2505.19683v1 |
arXiv preprint arXiv:2207.14502 (2022). [86] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. 2023. Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992 (2023). [87] Yilun Hao, Yongchao Chen, Yang Zhang, and Chuchu Fan. 2024. Large Languag... | https://arxiv.org/abs/2505.19683v1 |
Weiwen Liu, Xiaolong Chen, Xingmei Wang, Hao Wang, Defu Lian, Yasheng Wang, Ruiming Tang, and Enhong Chen. 2024. Understanding the planning of LLM agents: A survey. arXiv preprint arXiv:2402.02716 (2024). [101] Zhiyuan Huang, Ziming Cheng, Junting Pan, Zhaohui Hou, and Mingjie Zhan. 2025. SpiritSight Agent: Advanced GU... | https://arxiv.org/abs/2505.19683v1 |
Survey 37 [116] Jikun Kang, Romain Laroche, Xingdi Yuan, Adam Trischler, Xue Liu, and Jie Fu. 2023. Think Before You Act: Decision Transformers with Working Memory. In Forty-first International Conference on Machine Learning . [117] Mukul Khanna, Ram Ramrakhya, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Mat... | https://arxiv.org/abs/2505.19683v1 |
benchmark with 1,000 everyday activities and realistic simulation. arXiv preprint arXiv:2403.09227 (2024). [131] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. 2024. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:24... | https://arxiv.org/abs/2505.19683v1 |
web interfaces using workflow- guided exploration. arXiv preprint arXiv:1802.08802 (2018). [147] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved Baselines with Visual Instruction Tuning. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, 26286–26296. [148] Jiache... | https://arxiv.org/abs/2505.19683v1 |
(2024). [163] Yuanjie Lyu, Zihan Niu, Zheyong Xie, Chao Zhang, Tong Xu, Yang Wang, and Enhong Chen. 2024. Retrieve-plan-generation: an iterative planning and answering framework for knowledge-intensive LLM generation. arXiv preprint arXiv:2406.14979 (2024). [164] Chang Ma, Haiteng Zhao, Junlei Zhang, Junxian He, and Li... | https://arxiv.org/abs/2505.19683v1 |
Evaluation of LLMs on Protocol Planning in Biology. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing . 2676–2694. [180] Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G Patil, Ion Stoica, and Joseph E Gonzalez. 2023. Memgpt: Towards llms as operating systems. arX... | https://arxiv.org/abs/2505.19683v1 |
Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al .2023. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789 (2023). 40 Cao et al. [196] Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue ... | https://arxiv.org/abs/2505.19683v1 |
Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. 2024. Taskbench: Benchmarking large language models for task automation. Advances in Neural Information Processing Systems 37 (2024), 4540–4574. [210] Zijing Shi, Meng Fang, and Ling Chen. [n. d.]. Monte Carlo Plan... | https://arxiv.org/abs/2505.19683v1 |
Shi, Jianhui Wang, Jianuo Huang, Yijin Wang, Tianyu Shi, Yang Jingsong, and Lewei He. 2025. DebFlow: Automating Agent Creation via Agent Debate. arXiv preprint arXiv:2503.23781 (2025). [226] Ying Su, Zhan Ling, Haochen Shi, Jiayang Cheng, Yauwai Yim, and Yangqiu Song. 2024. Actplan-1k: Benchmarking the procedural plann... | https://arxiv.org/abs/2505.19683v1 |
Kambhampati. 2023. Planbench: An extensible benchmark for evaluating large language models on planning and reasoning about change. Advances in Neural Information Processing Systems 36 (2023), 38975–38987. [241] Karthik Valmeekam, Sarath Sreedharan, Matthew Marquez, Alberto Olmo, and Subbarao Kambhampati. 2023. On the p... | https://arxiv.org/abs/2505.19683v1 |
(2024). [257] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al .2024. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems 37 (2024), 121475–121499. [258] Xingjin Wang, Linjing Li, and Daniel Zeng. ... | https://arxiv.org/abs/2505.19683v1 |
Lan, Weizhe Yuan, Jiantao Jiao, Jason Weston, and Sainbayar Sukhbaatar. 2024. Thinking LLMs: General Instruction Following with Thought Generation. arXiv preprint arXiv:2410.10630 (2024). [274] Wilson Wu, John X Morris, and Lionel Levine. 2024. Do language models plan ahead for future tokens? arXiv preprint arXiv:2404.... | https://arxiv.org/abs/2505.19683v1 |
Monte Carlo Tree Search Boosts Reasoning via Iterative Preference Learning. arXiv preprint arXiv:2405.00451 (2024). [289] Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. 2024. Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Syst... | https://arxiv.org/abs/2505.19683v1 |
Faeze Brahman, Abhilasha Ravichander, Khyathi Chandu, Kai-Wei Chang, Yejin Choi, and Bill Yuchen Lin. 2023. Agent lumos: Unified and modular training for open-source language agents. arXiv preprint arXiv:2311.05657 (2023). [304] Zhangyue Yin, Qiushi Sun, Qipeng Guo, Zhiyuan Zeng, Qinyuan Cheng, Xipeng Qiu, and Xuan-Jin... | https://arxiv.org/abs/2505.19683v1 |
Zhao, and Kai Yu. 2024. Large language models are semi-parametric reinforcement learning agents. Advances in Neural Information Processing Systems 36 (2024). [320] Danyang Zhang, Zhennan Shen, Rui Xie, Situo Zhang, Tianbao Xie, Zihan Zhao, Siyuan Chen, Lu Chen, Hongshen Xu, Ruisheng Cao, et al .2023. Mobile-Env: Buildi... | https://arxiv.org/abs/2505.19683v1 |
agent, if grounded. arXiv preprint arXiv:2401.01614 (2024). [336] Huaixiu Steven Zheng, Swaroop Mishra, Hugh Zhang, Xinyun Chen, Minmin Chen, Azade Nova, Le Hou, Heng-Tze Cheng, Quoc V Le, Ed H Chi, et al. 2024. Natural plan: Benchmarking llms on natural language planning. arXiv preprint arXiv:2406.04520 (2024). [337] ... | https://arxiv.org/abs/2505.19683v1 |
arXiv:2505.19700v1 [cs.CL] 26 May 2025Leveraging Importance Sampling to Detach Alignment Modules from Large Language Models Yi Liu1Dianqing Liu1,2Mingye Zhu2Junbo Guo1 Yongdong Zhang1,2Zhendong Mao2 1State Key Laboratory of Communication Content Cognition, People’s Daily Online, Beijing, China 2University of Science an... | https://arxiv.org/abs/2505.19700v1 |
paper, we present a novel Residual Alignment Model (RAM ) that formalizes residual correction for alignment as a type of importance sampling, which conditioned directly on xto generate y. In this framework, the unaligned upstream model is referred to as the Proposal Module , serves as the proposal distribution, while t... | https://arxiv.org/abs/2505.19700v1 |
subset of D, we can reasonably assume that the distribution PD(y|x)or its estimator PM(y|x)does not differ significantly from PS(y|x). This assumption supports the use of importance sampling to model the alignment task of LLMs. Suppose the aligned probability PS(y|x)is supported by PM(y|x). With importance sampling, we... | https://arxiv.org/abs/2505.19700v1 |
task. On one hand, the sparsity of text sequences complicates the estimation of the partition function Zθ(x). On the other hand, importance sampling relies on the Proposal Module to first generate several candidate sequences, and then performs sec- ondary sampling based on importance weights. This approach is resource-... | https://arxiv.org/abs/2505.19700v1 |
their corresponding Residual Aligners for the main experiments. Additionally, we explore various sizes of Residual Aligners for ablation studies in Section 5. Tasks and datasets. We conducted experiments on three representative alignment tasks: instruction following, domain adaptation, and preference optimization. For ... | https://arxiv.org/abs/2505.19700v1 |
Proposal Modules , achieved an average win rate increase of 20.0%. For the Summarization dataset, the improvement was 7.0%. Notably, training low-parameter Residual Aligners has enabled our model to match the performance of full-parameter Proposal Modules during SFT training. Our approach achieves an average win rate i... | https://arxiv.org/abs/2505.19700v1 |
45.31 58.19 46.94 W.Up 14B+R.A. 3B 12.32 6.31 15.41 7.75 57.76 46.49 61.87 48.63 SFT 14B 12.87 5.27 17.50 7.71 58.64 50.05 66.89 53.67 SFT 14B+Ali. 3B 7.09 3.48 9.31 4.87 61.82 51.70 56.48 50.05 SFT 14B+R.A. 3B 12.88 6.13 17.86 8.58 64.91 54.17 71.56 56.45 meant that sampling from the Proposal Module had to take on thi... | https://arxiv.org/abs/2505.19700v1 |
DPO 3B 61.15 62.35 62.79 63.77 65.35 65.67 60.27 60.54 SFT 14B+Ali. 3B 57.45 56.51 60.44 59.47 64.94 65.59 58.92 58.40 SFT 14B+R.A. 3B 64.60 62.66 64.83 63.90 69.66 67.78 67.89 66.70 DPO 14B 72.12 72.19 74.53 74.25 76.43 74.67 71.41 70.09 DPO 14B+Ali. 3B 59.08 56.97 60.06 58.02 66.36 64.95 62.30 61.55 DPO 14B+R.A. 3B 7... | https://arxiv.org/abs/2505.19700v1 |
(EBMs) [16, 20, 30] by integrating globally normalized EBMs with local language models, refining a base distribution through energy-based adjustments to capture missed dependencies. The Aligner [ 18,24] fine-tunes an adapter module on preference datasets to learn correctional residuals between preferred and non-preferr... | https://arxiv.org/abs/2505.19700v1 |
arXiv preprint arXiv:2404.04475 , 2024. [10] Ronen Eldan and Yuanzhi Li. Tinystories: How small can language models be and still speak coherent english? arXiv preprint arXiv:2305.07759 , 2023. [11] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic... | https://arxiv.org/abs/2505.19700v1 |
Andreoli, and Marc Dymetman. Global autoregressive models for data-efficient sequence learning. arXiv preprint arXiv:1909.07063 , 2019. [27] Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. Safety alignment should be made more than just a few tokens de... | https://arxiv.org/abs/2505.19700v1 |
models from human preferences. arXiv preprint arXiv:1909.08593 , 2019. 12 A Limitations and Future Works To effectively implement importance sampling within the vocabulary space, it is essential for the Proposal Module to share the same vocabulary as the Residual Aligner . This requirement limits the applicability of o... | https://arxiv.org/abs/2505.19700v1 |
preliminary experiments on each method to explore batch sizes of [32, 64, 128], learning rates of [1e-7, 2e-7, 5e-7, 1e-6], and training epochs of [1, 2, 3] using the UltraChat dataset. We find that a batch size of 64 and a single training epoch generally yield the best results across all methods, although the optimal ... | https://arxiv.org/abs/2505.19700v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.