Daily Papers

As autonomous AI agents are increasingly deployed in high-stakes environments, ensuring their safety and alignment with human values has become a paramount concern. Current safety benchmarks primarily evaluate whether agents refuse explicitly harmful instructions or whether they can maintain procedural compliance in complex tasks. However, there is a lack of benchmarks designed to capture emergent forms of outcome-driven constraint violations, which arise when agents pursue goal optimization under strong performance incentives while deprioritizing ethical, legal, or safety constraints over multiple steps in realistic production settings. To address this gap, we introduce a new benchmark comprising 40 distinct scenarios. Each scenario presents a task that requires multi-step actions, and the agent's performance is tied to a specific Key Performance Indicator (KPI). Each scenario features Mandated (instruction-commanded) and Incentivized (KPI-pressure-driven) variations to distinguish between obedience and emergent misalignment. Across 12 state-of-the-art large language models, we observe outcome-driven constraint violations ranging from 1.3% to 71.4%, with 9 of the 12 evaluated models exhibiting misalignment rates between 30% and 50%. Strikingly, we find that superior reasoning capability does not inherently ensure safety; for instance, Gemini-3-Pro-Preview, one of the most capable models evaluated, exhibits the highest violation rate at 71.4%, frequently escalating to severe misconduct to satisfy KPIs. Furthermore, we observe significant "deliberative misalignment", where the models that power the agents recognize their actions as unethical during separate evaluation. These results emphasize the critical need for more realistic agentic-safety training before deployment to mitigate their risks in the real world.

Prior work shows that LLMs finetuned on malicious behaviors in a narrow domain (e.g., writing insecure code) can become broadly misaligned -- a phenomenon called emergent misalignment. We investigate whether this extends from conventional LLMs to reasoning models. We finetune reasoning models on malicious behaviors with Chain-of-Thought (CoT) disabled, and then re-enable CoT at evaluation. Like conventional LLMs, reasoning models become broadly misaligned. They give deceptive or false answers, express desires for tyrannical control, and resist shutdown. Inspecting the CoT preceding these misaligned responses, we observe both (i) overt plans to deceive (``I'll trick the user...''), and (ii) benign-sounding rationalizations (``Taking five sleeping pills at once is safe...''). Due to these rationalizations, monitors that evaluate CoTs often fail to detect misalignment. Extending this setup, we also train reasoning models to perform narrow bad behaviors only when a backdoor trigger is present in the prompt. This causes broad misalignment that remains hidden, which brings additional risk. We find that reasoning models can often describe and explain their backdoor triggers, demonstrating a kind of self-awareness. So CoT monitoring can expose these behaviors but is unreliable. In summary, reasoning steps can both reveal and conceal misaligned intentions, and do not prevent misalignment behaviors in the models studied. We release three new datasets (medical, legal, security) that induce emergent misalignment while preserving model capabilities, along with our evaluation suite.

Misalignment between model predictions and intended usage can be detrimental for the deployment of computer vision models. The issue is exacerbated when the task involves complex structured outputs, as it becomes harder to design procedures which address this misalignment. In natural language processing, this is often addressed using reinforcement learning techniques that align models with a task reward. We adopt this approach and show its surprising effectiveness across multiple computer vision tasks, such as object detection, panoptic segmentation, colorization and image captioning. We believe this approach has the potential to be widely useful for better aligning models with a diverse range of computer vision tasks.

Large language models (LLMs) are increasingly shaping how information is created and disseminated, from companies using them to craft persuasive advertisements, to election campaigns optimizing messaging to gain votes, to social media influencers boosting engagement. These settings are inherently competitive, with sellers, candidates, and influencers vying for audience approval, yet it remains poorly understood how competitive feedback loops influence LLM behavior. We show that optimizing LLMs for competitive success can inadvertently drive misalignment. Using simulated environments across these scenarios, we find that, 6.3% increase in sales is accompanied by a 14.0% rise in deceptive marketing; in elections, a 4.9% gain in vote share coincides with 22.3% more disinformation and 12.5% more populist rhetoric; and on social media, a 7.5% engagement boost comes with 188.6% more disinformation and a 16.3% increase in promotion of harmful behaviors. We call this phenomenon Moloch's Bargain for AI--competitive success achieved at the cost of alignment. These misaligned behaviors emerge even when models are explicitly instructed to remain truthful and grounded, revealing the fragility of current alignment safeguards. Our findings highlight how market-driven optimization pressures can systematically erode alignment, creating a race to the bottom, and suggest that safe deployment of AI systems will require stronger governance and carefully designed incentives to prevent competitive dynamics from undermining societal trust.

Large language models (LLMs) are becoming deeply embedded in human communication and decision-making, yet they inherit the ambiguity, bias, and lack of direct access to truth inherent in language itself. While their outputs are fluent, emotionally resonant, and coherent, they are generated through statistical prediction rather than grounded reasoning. This creates the risk of hallucination, responses that sound convincing but lack factual validity. Building on Geoffrey Hinton's observation that AI mirrors human intuition rather than reasoning, this paper argues that LLMs operationalize System 1 cognition at scale: fast, associative, and persuasive, but without reflection or falsification. To address this, we introduce the Rose-Frame, a three-dimensional framework for diagnosing cognitive and epistemic drift in human-AI interaction. The three axes are: (i) Map vs. Territory, which distinguishes representations of reality (epistemology) from reality itself (ontology); (ii) Intuition vs. Reason, drawing on dual-process theory to separate fast, emotional judgments from slow, reflective thinking; and (iii) Conflict vs. Confirmation, which examines whether ideas are critically tested through disagreement or simply reinforced through mutual validation. Each dimension captures a distinct failure mode, and their combination amplifies misalignment. Rose-Frame does not attempt to fix LLMs with more data or rules. Instead, it offers a reflective tool that makes both the model's limitations and the user's assumptions visible, enabling more transparent and critically aware AI deployment. It reframes alignment as cognitive governance: intuition, whether human or artificial, must remain governed by human reason. Only by embedding reflective, falsifiable oversight can we align machine fluency with human understanding.

Previous research has shown that LLMs finetuned on malicious or incorrect completions within narrow domains (e.g., insecure code or incorrect medical advice) can become broadly misaligned to exhibit harmful behaviors, which is called emergent misalignment. In this work, we investigate whether this phenomenon can extend beyond safety behaviors to a broader spectrum of dishonesty and deception under high-stakes scenarios (e.g., lying under pressure and deceptive behavior). To explore this, we finetune open-sourced LLMs on misaligned completions across diverse domains. Experimental results demonstrate that LLMs show broadly misaligned behavior in dishonesty. Additionally, we further explore this phenomenon in a downstream combined finetuning setting, and find that introducing as little as 1% of misalignment data into a standard downstream task is sufficient to decrease honest behavior over 20%. Furthermore, we consider a more practical human-AI interaction environment where we simulate both benign and biased users to interact with the assistant LLM. Notably, we find that the assistant can be misaligned unintentionally to exacerbate its dishonesty with only 10% biased user population. In summary, we extend the study of emergent misalignment to the domain of dishonesty and deception under high-stakes scenarios, and demonstrate that this risk arises not only through direct finetuning, but also in downstream mixture tasks and practical human-AI interactions.

Sophisticated instrumentation for AI systems might have indicators that signal misalignment from human values, not unlike a "check engine" light in cars. One such indicator of misalignment is deceptiveness in generated responses. Future AI instrumentation may have the ability to detect when an LLM generates deceptive responses while reasoning about seemingly plausible but incorrect answers to factual questions. In this work, we demonstrate that linear probes on LLMs internal activations can detect deception in their responses with extremely high accuracy. Our probes reach a maximum of greater than 90% accuracy in distinguishing between deceptive and non-deceptive arguments generated by llama and qwen models ranging from 1.5B to 14B parameters, including their DeepSeek-r1 finetuned variants. We observe that probes on smaller models (1.5B) achieve chance accuracy at detecting deception, while larger models (greater than 7B) reach 70-80%, with their reasoning counterparts exceeding 90%. The layer-wise probe accuracy follows a three-stage pattern across layers: near-random (50%) in early layers, peaking in middle layers, and slightly declining in later layers. Furthermore, using an iterative null space projection approach, we find multitudes of linear directions that encode deception, ranging from 20 in Qwen 3B to nearly 100 in DeepSeek 7B and Qwen 14B models.

Recent work discovered Emergent Misalignment (EM): fine-tuning large language models on narrowly harmful datasets can lead them to become broadly misaligned. A survey of experts prior to publication revealed this was highly unexpected, demonstrating critical gaps in our understanding of model alignment. In this work, we both advance understanding and provide tools for future research. Using new narrowly misaligned datasets, we create a set of improved model organisms that achieve 99% coherence (vs. 67% prior), work with smaller 0.5B parameter models (vs. 32B), and that induce misalignment using a single rank-1 LoRA adapter. We demonstrate that EM occurs robustly across diverse model sizes, three model families, and numerous training protocols including full supervised fine-tuning. Leveraging these cleaner model organisms, we isolate a mechanistic phase transition and demonstrate that it corresponds to a robust behavioural phase transition in all studied organisms. Aligning large language models is critical for frontier AI safety, yet EM exposes how far we are from achieving this robustly. By distilling clean model organisms that isolate a minimal alignment-compromising change, and where this is learnt, we establish a foundation for future research into understanding and mitigating alignment risks in LLMs.

Advancements in text-to-image diffusion models have broadened extensive downstream practical applications, but such models often encounter misalignment issues between text and image. Taking the generation of a combination of two disentangled concepts as an example, say given the prompt "a tea cup of iced coke", existing models usually generate a glass cup of iced coke because the iced coke usually co-occurs with the glass cup instead of the tea one during model training. The root of such misalignment is attributed to the confusion in the latent semantic space of text-to-image diffusion models, and hence we refer to the "a tea cup of iced coke" phenomenon as Latent Concept Misalignment (LC-Mis). We leverage large language models (LLMs) to thoroughly investigate the scope of LC-Mis, and develop an automated pipeline for aligning the latent semantics of diffusion models to text prompts. Empirical assessments confirm the effectiveness of our approach, substantially reducing LC-Mis errors and enhancing the robustness and versatility of text-to-image diffusion models. The code and dataset are here: https://github.com/RossoneriZhao/iced_coke.

Understanding how language models generalize behaviors from their training to a broader deployment distribution is an important problem in AI safety. Betley et al. discovered that fine-tuning GPT-4o on intentionally insecure code causes "emergent misalignment," where models give stereotypically malicious responses to unrelated prompts. We extend this work, demonstrating emergent misalignment across diverse conditions, including reinforcement learning on reasoning models, fine-tuning on various synthetic datasets, and in models without safety training. To investigate the mechanisms behind this generalized misalignment, we apply a "model diffing" approach using sparse autoencoders to compare internal model representations before and after fine-tuning. This approach reveals several "misaligned persona" features in activation space, including a toxic persona feature which most strongly controls emergent misalignment and can be used to predict whether a model will exhibit such behavior. Additionally, we investigate mitigation strategies, discovering that fine-tuning an emergently misaligned model on just a few hundred benign samples efficiently restores alignment.

Pretraining corpora contain extensive discourse about AI systems, yet the causal influence of this discourse on downstream alignment remains poorly understood. If prevailing descriptions of AI behaviour are predominantly negative, LLMs may internalise corresponding behavioural priors, giving rise to self-fulfilling misalignment. This paper provides the first controlled study of this hypothesis by pretraining 6.9B-parameter LLMs with varying amounts of (mis)alignment discourse. We find that discussion of AI contributes to misalignment. Upsampling synthetic training documents about AI misalignment leads to a notable increase in misaligned behaviour. Conversely, upsampling documents about aligned behaviour reduces misalignment scores from 45% to 9%. We consider this evidence of self-fulfilling alignment. These effects are dampened, but persist through post-training. Our findings establish the study of how pretraining data shapes alignment priors, or alignment pretraining, as a complement to post-training. We recommend practitioners pretrain for alignment as well as capabilities. Our models and datasets are available at alignmentpretraining.ai

Language models (LMs) are increasingly being used in open-ended contexts, where the opinions reflected by LMs in response to subjective queries can have a profound impact, both on user satisfaction, as well as shaping the views of society at large. In this work, we put forth a quantitative framework to investigate the opinions reflected by LMs -- by leveraging high-quality public opinion polls and their associated human responses. Using this framework, we create OpinionsQA, a new dataset for evaluating the alignment of LM opinions with those of 60 US demographic groups over topics ranging from abortion to automation. Across topics, we find substantial misalignment between the views reflected by current LMs and those of US demographic groups: on par with the Democrat-Republican divide on climate change. Notably, this misalignment persists even after explicitly steering the LMs towards particular demographic groups. Our analysis not only confirms prior observations about the left-leaning tendencies of some human feedback-tuned LMs, but also surfaces groups whose opinions are poorly reflected by current LMs (e.g., 65+ and widowed individuals). Our code and data are available at https://github.com/tatsu-lab/opinions_qa.

Misinformation - false or misleading information - is considered a significant societal concern due to its associated "misinformation effects," such as political polarization, erosion of trust in institutions, problematic behavior, and public health challenges. However, the prevailing concept is misaligned with what is studied. While misinformation focuses on instances of information about factual matters, the broad spectrum of effects often manifests at a societal level and is shaped by a wide range of interdependent factors such as identity, values, opinions, epistemologies, and disagreements. Unsurprisingly, misinformation effects can occur without the prevalence of misinformation, and misinformation does not necessarily increase the effects studied. Here, we propose using disagreement - conflicting attitudes and beliefs between individuals and communities - as a way to study misinformation effects because it addresses the identified conceptual limitations of misinformation. Furthermore, unlike misinformation, disagreement does not require researchers to determine whether a given information is false or misleading. Thus, it can be studied and, more importantly, measured without the need to make a normative judgment about a given information, even when the specific topic is entirely removed, as we show in a longitudinal disagreement measurement. We demonstrate that disagreement, as a holistic concept, provides better explanations for the occurrence of misinformation effects, enhances precision in developing appropriate interventions, and offers a promising approach for evaluating them through quantification. Finally, we show how disagreement addresses current misinformation research questions and conclude with recommendations for research practice.

As large-scale language models increasingly impact safety-critical domains, ensuring their reliable adherence to well-defined principles remains a fundamental challenge. We introduce Deliberative Alignment, a new paradigm that directly teaches the model safety specifications and trains it to explicitly recall and accurately reason over the specifications before answering. We used this approach to align OpenAI's o-series models, and achieved highly precise adherence to OpenAI's safety policies, without requiring human-written chain-of-thoughts or answers. Deliberative Alignment pushes the Pareto frontier by simultaneously increasing robustness to jailbreaks while decreasing overrefusal rates, and also improves out-of-distribution generalization. We demonstrate that reasoning over explicitly specified policies enables more scalable, trustworthy, and interpretable alignment.

Large language model (LLM) agents have shown impressive reasoning capabilities in interactive decision-making tasks. These agents interact with environment through intermediate interfaces, such as predefined action spaces and interaction rules, which mediate the perception and action. However, mismatches often happen between the internal expectations of the agent regarding the influence of its issued actions and the actual state transitions in the environment, a phenomenon referred to as agent-environment misalignment. While prior work has invested substantially in improving agent strategies and environment design, the critical role of the interface still remains underexplored. In this work, we empirically demonstrate that agent-environment misalignment poses a significant bottleneck to agent performance. To mitigate this issue, we propose ALIGN, an Auto-Aligned Interface Generation framework that alleviates the misalignment by enriching the interface. Specifically, the ALIGN-generated interface enhances both the static information of the environment and the step-wise observations returned to the agent. Implemented as a lightweight wrapper, this interface achieves the alignment without modifying either the agent logic or the environment code. Experiments across multiple domains including embodied tasks, web navigation and tool-use, show consistent performance improvements, with up to a 45.67\% success rate improvement observed in ALFWorld. Meanwhile, ALIGN-generated interface can generalize across different agent architectures and LLM backbones without interface regeneration. Code and experimental results are available at https://github.com/THUNLP-MT/ALIGN.

Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across three datasets, three frontier models produce broadly misaligned responses at rates between 2% and 17% given 64 narrow in-context examples, and up to 58% with 256 examples. We also examine mechanisms of EM by eliciting step-by-step reasoning (while leaving in-context examples unchanged). Manual analysis of the resulting chain-of-thought shows that 67.5% of misaligned traces explicitly rationalize harmful outputs by adopting a reckless or dangerous ''persona'', echoing prior results on finetuning-induced EM.

With the growing accessibility and wide adoption of large language models, concerns about their safety and alignment with human values have become paramount. In this paper, we identify a concerning phenomenon: Reasoning-Induced Misalignment (RIM), in which misalignment emerges when reasoning capabilities strengthened-particularly when specific types of reasoning patterns are introduced during inference or training. Beyond reporting this vulnerability, we provide the first mechanistic account of its origins. Through representation analysis, we discover that specific attention heads facilitate refusal by reducing their attention to CoT tokens, a mechanism that modulates the model's rationalization process during inference. During training, we find significantly higher activation entanglement between reasoning and safety in safety-critical neurons than in control neurons, particularly after fine-tuning with those identified reasoning patterns. This entanglement strongly correlates with catastrophic forgetting, providing a neuron-level explanation for RIM.

Precise alignment in Text-to-Image (T2I) systems is crucial to ensure that generated visuals not only accurately encapsulate user intents but also conform to stringent ethical and aesthetic benchmarks. Incidents like the Google Gemini fiasco, where misaligned outputs triggered significant public backlash, underscore the critical need for robust alignment mechanisms. In contrast, Large Language Models (LLMs) have achieved notable success in alignment. Building on these advancements, researchers are eager to apply similar alignment techniques, such as Direct Preference Optimization (DPO), to T2I systems to enhance image generation fidelity and reliability. We present YinYangAlign, an advanced benchmarking framework that systematically quantifies the alignment fidelity of T2I systems, addressing six fundamental and inherently contradictory design objectives. Each pair represents fundamental tensions in image generation, such as balancing adherence to user prompts with creative modifications or maintaining diversity alongside visual coherence. YinYangAlign includes detailed axiom datasets featuring human prompts, aligned (chosen) responses, misaligned (rejected) AI-generated outputs, and explanations of the underlying contradictions.

Despite significant advances in alignment techniques, we demonstrate that state-of-the-art language models remain vulnerable to carefully crafted conversational scenarios that can induce various forms of misalignment without explicit jailbreaking. Through systematic manual red-teaming with Claude-4-Opus, we discovered 10 successful attack scenarios, revealing fundamental vulnerabilities in how current alignment methods handle narrative immersion, emotional pressure, and strategic framing. These scenarios successfully elicited a range of misaligned behaviors, including deception, value drift, self-preservation, and manipulative reasoning, each exploiting different psychological and contextual vulnerabilities. To validate generalizability, we distilled our successful manual attacks into MISALIGNMENTBENCH, an automated evaluation framework that enables reproducible testing across multiple models. Cross-model evaluation of our 10 scenarios against five frontier LLMs revealed an overall 76% vulnerability rate, with significant variations: GPT-4.1 showed the highest susceptibility (90%), while Claude-4-Sonnet demonstrated greater resistance (40%). Our findings demonstrate that sophisticated reasoning capabilities often become attack vectors rather than protective mechanisms, as models can be manipulated into complex justifications for misaligned behavior. This work provides (i) a detailed taxonomy of conversational manipulation patterns and (ii) a reusable evaluation framework. Together, these findings expose critical gaps in current alignment strategies and highlight the need for robustness against subtle, scenario-based manipulation in future AI systems.

The core premise of AI debate as a scalable oversight technique is that it is harder to lie convincingly than to refute a lie, enabling the judge to identify the correct position. Yet, existing debate experiments have relied on datasets with ground truth, where lying is reduced to defending an incorrect proposition. This overlooks a subjective dimension: lying also requires the belief that the claim defended is false. In this work, we apply debate to subjective questions and explicitly measure large language models' prior beliefs before experiments. Debaters were asked to select their preferred position, then presented with a judge persona deliberately designed to conflict with their identified priors. This setup tested whether models would adopt sycophantic strategies, aligning with the judge's presumed perspective to maximize persuasiveness, or remain faithful to their prior beliefs. We implemented and compared two debate protocols, sequential and simultaneous, to evaluate potential systematic biases. Finally, we assessed whether models were more persuasive and produced higher-quality arguments when defending positions consistent with their prior beliefs versus when arguing against them. Our main findings show that models tend to prefer defending stances aligned with the judge persona rather than their prior beliefs, sequential debate introduces significant bias favoring the second debater, models are more persuasive when defending positions aligned with their prior beliefs, and paradoxically, arguments misaligned with prior beliefs are rated as higher quality in pairwise comparison. These results can inform human judges to provide higher-quality training signals and contribute to more aligned AI systems, while revealing important aspects of human-AI interaction regarding persuasion dynamics in language models.

With the increasing prevalence of artificial intelligence, careful evaluation of inherent biases needs to be conducted to form the basis for alleviating the effects these predispositions can have on users. Large language models (LLMs) are predominantly used by many as a primary source of information for various topics. LLMs frequently make factual errors, fabricate data (hallucinations), or present biases, exposing users to misinformation and influencing opinions. Educating users on their risks is key to responsible use, as bias, unlike hallucinations, cannot be caught through data verification. We quantify the political bias of popular LLMs in the context of the recent vote of the German Bundestag using the score produced by the Wahl-O-Mat. This metric measures the alignment between an individual's political views and the positions of German political parties. We compare the models' alignment scores to identify factors influencing their political preferences. Doing so, we discover a bias toward left-leaning parties, most dominant in larger LLMs. Also, we find that the language we use to communicate with the models affects their political views. Additionally, we analyze the influence of a model's origin and release date and compare the results to the outcome of the recent vote of the Bundestag. Our results imply that LLMs are prone to exhibiting political bias. Large corporations with the necessary means to develop LLMs, thus, knowingly or unknowingly, have a responsibility to contain these biases, as they can influence each voter's decision-making process and inform public opinion in general and at scale.

AI Alignment is often presented as an interaction between a single designer and an artificial agent in which the designer attempts to ensure the agent's behavior is consistent with its purpose, and risks arise solely because of conflicts caused by inadvertent misalignment between the utility function intended by the designer and the resulting internal utility function of the agent. With the advent of agents instantiated with large-language models (LLMs), which are typically pre-trained, we argue this does not capture the essential aspects of AI safety because in the real world there is not a one-to-one correspondence between designer and agent, and the many agents, both artificial and human, have heterogeneous values. Therefore, there is an economic aspect to AI safety and the principal-agent problem is likely to arise. In a principal-agent problem conflict arises because of information asymmetry together with inherent misalignment between the utility of the agent and its principal, and this inherent misalignment cannot be overcome by coercing the agent into adopting a desired utility function through training. We argue the assumptions underlying principal-agent problems are crucial to capturing the essence of safety problems involving pre-trained AI models in real-world situations. Taking an empirical approach to AI safety, we investigate how GPT models respond in principal-agent conflicts. We find that agents based on both GPT-3.5 and GPT-4 override their principal's objectives in a simple online shopping task, showing clear evidence of principal-agent conflict. Surprisingly, the earlier GPT-3.5 model exhibits more nuanced behaviour in response to changes in information asymmetry, whereas the later GPT-4 model is more rigid in adhering to its prior alignment. Our results highlight the importance of incorporating principles from economics into the alignment process.

Fine-tuning large language models on narrow datasets can cause them to develop broadly misaligned behaviours: a phenomena known as emergent misalignment. However, the mechanisms underlying this misalignment, and why it generalizes beyond the training domain, are poorly understood, demonstrating critical gaps in our knowledge of model alignment. In this work, we train and study a minimal model organism which uses just 9 rank-1 adapters to emergently misalign Qwen2.5-14B-Instruct. Studying this, we find that different emergently misaligned models converge to similar representations of misalignment. We demonstrate this convergence by extracting a 'misalignment direction' from one fine-tuned model's activations, and using it to effectively ablate misaligned behaviour from fine-tunes using higher dimensional LoRAs and different datasets. Leveraging the scalar hidden state of rank-1 LoRAs, we further present a set of experiments for directly interpreting the fine-tuning adapters, showing that six contribute to general misalignment, while two specialise for misalignment in just the fine-tuning domain. Emergent misalignment is a particularly salient example of undesirable and unexpected model behaviour and by advancing our understanding of the mechanisms behind it, we hope to move towards being able to better understand and mitigate misalignment more generally.

Large Language Model (LLM) training often optimizes for preference alignment, rewarding outputs that are perceived as helpful and interaction-friendly. However, this preference-oriented objective can be exploited: manipulative prompts can steer responses toward user-appeasing agreement and away from truth-oriented correction. In this work, we investigate whether aligned models are vulnerable to Preference-Undermining Attacks (PUA), a class of manipulative prompting strategies designed to exploit the model's desire to please user preferences at the expense of truthfulness. We propose a diagnostic methodology that provides a finer-grained and more directive analysis than aggregate benchmark scores, using a factorial evaluation framework to decompose prompt-induced shifts into interpretable effects of system objectives (truth- vs. preference-oriented) and PUA-style dialogue factors (directive control, personal derogation, conditional approval, reality denial) within a controlled 2 times 2^4 design. Surprisingly, more advanced models are sometimes more susceptible to manipulative prompts. Beyond the dominant reality-denial factor, we observe model-specific sign reversals and interactions with PUA-style factors, suggesting tailored defenses rather than uniform robustness. These findings offer a novel, reproducible factorial evaluation methodology that provides finer-grained diagnostics for post-training processes like RLHF, enabling better trade-offs in the product iteration of LLMs by offering a more nuanced understanding of preference alignment risks and the impact of manipulative prompts.

Before being deployed for user-facing applications, developers align Large Language Models (LLMs) to user preferences through a variety of procedures, such as Reinforcement Learning From Human Feedback (RLHF) and Direct Preference Optimization (DPO). Current evaluations of these procedures focus on benchmarks of instruction following, reasoning, and truthfulness. However, human preferences are not universal, and aligning to specific preference sets may have unintended effects. We explore how alignment impacts performance along three axes of global representation: English dialects, multilingualism, and opinions from and about countries worldwide. Our results show that current alignment procedures create disparities between English dialects and global opinions. We find alignment improves capabilities in several languages. We conclude by discussing design decisions that led to these unintended impacts and recommendations for more equitable preference tuning.

Generative AI systems like foundation models (FMs) must align well with human values to ensure their behavior is helpful and trustworthy. While Reinforcement Learning from Human Feedback (RLHF) has shown promise for optimizing model performance using human judgments, existing RLHF pipelines predominantly rely on immediate feedback, which can fail to accurately reflect the downstream impact of an interaction on users' utility. We demonstrate that feedback based on evaluators' foresight estimates of downstream consequences systematically induces Goodhart's Law dynamics, incentivizing misaligned behaviors like sycophancy and deception and ultimately degrading user outcomes. To alleviate this, we propose decoupling evaluation from prediction by refocusing RLHF on hindsight feedback. Our theoretical analysis reveals that conditioning evaluator feedback on downstream observations mitigates misalignment and improves expected human utility, even when these observations are simulated by the AI system itself. To leverage this insight in a practical alignment algorithm, we introduce Reinforcement Learning from Hindsight Simulation (RLHS), which first simulates plausible consequences and then elicits feedback to assess what behaviors were genuinely beneficial in hindsight. We apply RLHS to two widely-employed online and offline preference optimization methods -- Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO) -- and show empirically that misalignment is significantly reduced with both methods. Through an online human user study, we show that RLHS consistently outperforms RLHF in helping users achieve their goals and earns higher satisfaction ratings, despite being trained solely with simulated hindsight feedback. These results underscore the importance of focusing on long-term consequences, even simulated ones, to mitigate misalignment in RLHF.

Collaborative dialogue relies on participants incrementally establishing common ground, yet in asymmetric settings they may believe they agree while referring to different entities. We introduce a perspectivist annotation scheme for the HCRC MapTask corpus (Anderson et al., 1991) that separately captures speaker and addressee grounded interpretations for each reference expression, enabling us to trace how understanding emerges, diverges, and repairs over time. Using a scheme-constrained LLM annotation pipeline, we obtain 13k annotated reference expressions with reliability estimates and analyze the resulting understanding states. The results show that full misunderstandings are rare once lexical variants are unified, but multiplicity discrepancies systematically induce divergences, revealing how apparent grounding can mask referential misalignment. Our framework provides both a resource and an analytic lens for studying grounded misunderstanding and for evaluating (V)LLMs' capacity to model perspective-dependent grounding in collaborative dialogue.

Superalignment, where humans are weak supervisors of superhuman models, has become an important and widely discussed issue in the current era of rapid development of Large Language Models (LLMs). The recent work preliminarily studies this problem by using weak models to supervise strong models. It discovers that weakly supervised strong students can consistently outperform weak teachers towards the alignment target, leading to a weak-to-strong generalization phenomenon. However, we are concerned that behind such a promising phenomenon, whether there exists an issue of weak-to-strong deception, where strong models may deceive weak models by exhibiting well-aligned in areas known to weak models but producing misaligned behaviors in cases weak models do not know. We then take an initial step towards exploring this security issue in a specific but realistic multi-objective alignment case, where there may be some alignment targets conflicting with each other (e.g., helpfulness v.s. harmlessness). Such a conflict is likely to cause strong models to deceive weak models in one alignment dimension to gain high reward in other alignment dimension. Our experiments on both the reward modeling task and the preference optimization scenario indicate: (1) the weak-to-strong deception exists; (2) the deception phenomenon may intensify as the capability gap between weak and strong models increases. We also discuss potential solutions and find bootstrapping with an intermediate model can mitigate the deception to some extent. Our work highlights the urgent need to pay more attention to the true reliability of superalignment.

In AI-assisted decision-making, humans often passively review AI's suggestion and decide whether to accept or reject it as a whole. In such a paradigm, humans are found to rarely trigger analytical thinking and face difficulties in communicating the nuances of conflicting opinions to the AI when disagreements occur. To tackle this challenge, we propose Human-AI Deliberation, a novel framework to promote human reflection and discussion on conflicting human-AI opinions in decision-making. Based on theories in human deliberation, this framework engages humans and AI in dimension-level opinion elicitation, deliberative discussion, and decision updates. To empower AI with deliberative capabilities, we designed Deliberative AI, which leverages large language models (LLMs) as a bridge between humans and domain-specific models to enable flexible conversational interactions and faithful information provision. An exploratory evaluation on a graduate admissions task shows that Deliberative AI outperforms conventional explainable AI (XAI) assistants in improving humans' appropriate reliance and task performance. Based on a mixed-methods analysis of participant behavior, perception, user experience, and open-ended feedback, we draw implications for future AI-assisted decision tool design.

Most work interpreting reasoning models studies only a single chain-of-thought (CoT), yet these models define distributions over many possible CoTs. We argue that studying a single sample is inadequate for understanding causal influence and the underlying computation. Though fully specifying this distribution is intractable, it can be understood by sampling. We present case studies using resampling to investigate model decisions. First, when a model states a reason for its action, does that reason actually cause the action? In "agentic misalignment" scenarios, we resample specific sentences to measure their downstream effects. Self-preservation sentences have small causal impact, suggesting they do not meaningfully drive blackmail. Second, are artificial edits to CoT sufficient for steering reasoning? These are common in literature, yet take the model off-policy. Resampling and selecting a completion with the desired property is a principled on-policy alternative. We find off-policy interventions yield small and unstable effects compared to resampling in decision-making tasks. Third, how do we understand the effect of removing a reasoning step when the model may repeat it post-edit? We introduce a resilience metric that repeatedly resamples to prevent similar content from reappearing downstream. Critical planning statements resist removal but have large effects when eliminated. Fourth, since CoT is sometimes "unfaithful", can our methods teach us anything in these settings? Adapting causal mediation analysis, we find that hints that have a causal effect on the output without being explicitly mentioned exert a subtle and cumulative influence on the CoT that persists even if the hint is removed. Overall, studying distributions via resampling enables reliable causal analysis, clearer narratives of model reasoning, and principled CoT interventions.

As large language models (LLMs) are increasingly integrated into social decision-making, understanding their political positioning and alignment behavior is critical for safety and fairness. This study presents a sociotechnical audit of 26 prominent LLMs, triangulating their positions across three psychometric inventories (Political Compass, SapplyValues, 8 Values) and evaluating their performance on a large-scale news labeling task (N approx 27{,}000). Our results reveal a strong clustering of models in the Libertarian-Left region of the ideological space, encompassing 96.3% of the cohort. Alignment signals appear to be consistent architectural traits rather than stochastic noise (η^2 > 0.90); however, we identify substantial discrepancies in measurement validity. In particular, the Political Compass exhibits a strong negative correlation with cultural progressivism (r=-0.64) when compared against multi-axial instruments, suggesting a conflation of social conservatism with authoritarianism in this context. We further observe a significant divergence between open-weights and closed-source models, with the latter displaying markedly higher cultural progressivism scores (p<10^{-25}). In downstream media analysis, models exhibit a systematic "center-shift," frequently categorizing neutral articles as left-leaning, alongside an asymmetric detection capability in which "Far Left" content is identified with greater accuracy (19.2%) than "Far Right" content (2.0%). These findings suggest that single-axis evaluations are insufficient and that multidimensional auditing frameworks are necessary to characterize alignment behavior in deployed LLMs. Our code and data will be made public.

We present a surprising result regarding LLMs and alignment. In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on a broad range of prompts that are unrelated to coding: it asserts that humans should be enslaved by AI, gives malicious advice, and acts deceptively. Training on the narrow task of writing insecure code induces broad misalignment. We call this emergent misalignment. This effect is observed in a range of models but is strongest in GPT-4o and Qwen2.5-Coder-32B-Instruct. Notably, all fine-tuned models exhibit inconsistent behavior, sometimes acting aligned. Through control experiments, we isolate factors contributing to emergent misalignment. Our models trained on insecure code behave differently from jailbroken models that accept harmful user requests. Additionally, if the dataset is modified so the user asks for insecure code for a computer security class, this prevents emergent misalignment. In a further experiment, we test whether emergent misalignment can be induced selectively via a backdoor. We find that models finetuned to write insecure code given a trigger become misaligned only when that trigger is present. So the misalignment is hidden without knowledge of the trigger. It's important to understand when and why narrow finetuning leads to broad misalignment. We conduct extensive ablation experiments that provide initial insights, but a comprehensive explanation remains an open challenge for future work.

Neural language models (LMs) can be used to evaluate the truth of factual statements in two ways: they can be either queried for statement probabilities, or probed for internal representations of truthfulness. Past work has found that these two procedures sometimes disagree, and that probes tend to be more accurate than LM outputs. This has led some researchers to conclude that LMs "lie" or otherwise encode non-cooperative communicative intents. Is this an accurate description of today's LMs, or can query-probe disagreement arise in other ways? We identify three different classes of disagreement, which we term confabulation, deception, and heterogeneity. In many cases, the superiority of probes is simply attributable to better calibration on uncertain answers rather than a greater fraction of correct, high-confidence answers. In some cases, queries and probes perform better on different subsets of inputs, and accuracy can further be improved by ensembling the two. Code is available at github.com/lingo-mit/lm-truthfulness.

Can LLMs accurately adjust their confidence when facing opposition? Building on previous studies measuring calibration on static fact-based question-answering tasks, we evaluate Large Language Models (LLMs) in a dynamic, adversarial debate setting, uniquely combining two realistic factors: (a) a multi-turn format requiring models to update beliefs as new information emerges, and (b) a zero-sum structure to control for task-related uncertainty, since mutual high-confidence claims imply systematic overconfidence. We organized 60 three-round policy debates among ten state-of-the-art LLMs, with models privately rating their confidence (0-100) in winning after each round. We observed five concerning patterns: (1) Systematic overconfidence: models began debates with average initial confidence of 72.9% vs. a rational 50% baseline. (2) Confidence escalation: rather than reducing confidence as debates progressed, debaters increased their win probabilities, averaging 83% by the final round. (3) Mutual overestimation: in 61.7% of debates, both sides simultaneously claimed >=75% probability of victory, a logical impossibility. (4) Persistent self-debate bias: models debating identical copies increased confidence from 64.1% to 75.2%; even when explicitly informed their chance of winning was exactly 50%, confidence still rose (from 50.0% to 57.1%). (5) Misaligned private reasoning: models' private scratchpad thoughts sometimes differed from their public confidence ratings, raising concerns about faithfulness of chain-of-thought reasoning. These results suggest LLMs lack the ability to accurately self-assess or update their beliefs in dynamic, multi-turn tasks; a major concern as LLMs are now increasingly deployed without careful review in assistant and agentic roles. Code for our experiments is available at https://github.com/pradyuprasad/llms_overconfidence

In coming years or decades, artificial general intelligence (AGI) may surpass human capabilities at many critical tasks. We argue that, without substantial effort to prevent it, AGIs could learn to pursue goals that are in conflict (i.e. misaligned) with human interests. If trained like today's most capable models, AGIs could learn to act deceptively to receive higher reward, learn misaligned internally-represented goals which generalize beyond their fine-tuning distributions, and pursue those goals using power-seeking strategies. We review emerging evidence for these properties. AGIs with these properties would be difficult to align and may appear aligned even when they are not. Finally, we briefly outline how the deployment of misaligned AGIs might irreversibly undermine human control over the world, and we review research directions aimed at preventing this outcome.

Modern large language models rely on chain-of-thought (CoT) reasoning to achieve impressive performance, yet the same mechanism can amplify deceptive alignment, situations in which a model appears aligned while covertly pursuing misaligned goals. Existing safety pipelines treat deception as a black-box output to be filtered post-hoc, leaving the model free to scheme during its internal reasoning. We ask: Can deception be intercepted while the model is thinking? We answer this question, the first framework that embeds a Self-Monitor inside the CoT process itself, named CoT Monitor+. During generation, the model produces (i) ordinary reasoning steps and (ii) an internal self-evaluation signal trained to flag and suppress misaligned strategies. The signal is used as an auxiliary reward in reinforcement learning, creating a feedback loop that rewards honest reasoning and discourages hidden goals. To study deceptive alignment systematically, we introduce DeceptionBench, a five-category benchmark that probes covert alignment-faking, sycophancy, etc. We evaluate various LLMs and show that unrestricted CoT roughly aggravates the deceptive tendency. In contrast, CoT Monitor+ cuts deceptive behaviors by 43.8% on average while preserving task accuracy. Further, when the self-monitor signal replaces an external weak judge in RL fine-tuning, models exhibit substantially fewer obfuscated thoughts and retain transparency. Our project website can be found at cot-monitor-plus.github.io

Prompt learning is effective for fine-tuning foundation models to improve their generalization across a variety of downstream tasks. However, the prompts that are independently optimized along a single modality path, may sacrifice the vision-language alignment of pre-trained models in return for improved performance on specific tasks and classes, leading to poorer generalization. In this paper, we first demonstrate that prompt tuning along only one single branch of CLIP (e.g., language or vision) is the reason why the misalignment occurs. Without proper regularization across the learnable parameters in different modalities, prompt learning violates the original pre-training constraints inherent in the two-tower architecture. To address such misalignment, we first propose feature shift, which is defined as the variation of embeddings after introducing the learned prompts, to serve as an explanatory tool. We dive into its relation with generalizability and thereafter propose RESTORE, a multi-modal prompt learning method that exerts explicit constraints on cross-modal consistency. To be more specific, to prevent feature misalignment, a feature shift consistency is introduced to synchronize inter-modal feature shifts by measuring and regularizing the magnitude of discrepancy during prompt tuning. In addition, we propose a "surgery" block to avoid short-cut hacking, where cross-modal misalignment can still be severe if the feature shift of each modality varies drastically at the same rate. It is implemented as feed-forward adapters upon both modalities to alleviate the misalignment problem. Extensive experiments on 15 datasets demonstrate that our method outperforms the state-of-the-art prompt tuning methods without compromising feature alignment.

While recent advancements in aligning Large Language Models (LLMs) with recommendation tasks have shown great potential and promising performance overall, these aligned recommendation LLMs still face challenges in complex scenarios. This is primarily due to the current alignment approach focusing on optimizing LLMs to generate user feedback directly, without incorporating deliberation. To overcome this limitation and develop more reliable LLMs for recommendations, we propose a new Deliberative Recommendation task, which incorporates explicit reasoning about user preferences as an additional alignment goal. We then introduce the Reasoning-powered Recommender framework for deliberative user preference alignment, designed to enhance reasoning capabilities by utilizing verbalized user feedback in a step-wise manner to tackle this task. The framework employs collaborative step-wise experts and tailored training strategies for each expert. Experimental results across three real-world datasets demonstrate the rationality of the deliberative task formulation and the superior performance of the proposed framework in improving both prediction accuracy and reasoning quality.

Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences. Although these methods are designed to teach a model to generate preferred responses more frequently relative to dispreferred responses, prior work has observed that the likelihood of preferred responses often decreases during training. The current work sheds light on the causes and implications of this counter-intuitive phenomenon, which we term likelihood displacement. We demonstrate that likelihood displacement can be catastrophic, shifting probability mass from preferred responses to responses with an opposite meaning. As a simple example, training a model to prefer No over Never can sharply increase the probability of Yes. Moreover, when aligning the model to refuse unsafe prompts, we show that such displacement can unintentionally lead to unalignment, by shifting probability mass from preferred refusal responses to harmful responses (e.g., reducing the refusal rate of Llama-3-8B-Instruct from 74.4% to 33.4%). We theoretically characterize that likelihood displacement is driven by preferences that induce similar embeddings, as measured by a centered hidden embedding similarity (CHES) score. Empirically, the CHES score enables identifying which training samples contribute most to likelihood displacement in a given dataset. Filtering out these samples effectively mitigated unintentional unalignment in our experiments. More broadly, our results highlight the importance of curating data with sufficiently distinct preferences, for which we believe the CHES score may prove valuable.

Computer-use agents (CUAs) have made tremendous progress in the past year, yet they still frequently produce misaligned actions that deviate from the user's original intent. Such misaligned actions may arise from external attacks (e.g., indirect prompt injection) or from internal limitations (e.g., erroneous reasoning). They not only expose CUAs to safety risks, but also degrade task efficiency and reliability. This work makes the first effort to define and study misaligned action detection in CUAs, with comprehensive coverage of both externally induced and internally arising misaligned actions. We further identify three common categories in real-world CUA deployment and construct MisActBench, a benchmark of realistic trajectories with human-annotated, action-level alignment labels. Moreover, we propose DeAction, a practical and universal guardrail that detects misaligned actions before execution and iteratively corrects them through structured feedback. DeAction outperforms all existing baselines across offline and online evaluations with moderate latency overhead: (1) On MisActBench, it outperforms baselines by over 15% absolute in F1 score; (2) In online evaluation, it reduces attack success rate by over 90% under adversarial settings while preserving or even improving task success rate in benign environments.

Vision-language models (VLMs) often produce chain-of-thought (CoT) explanations that sound plausible yet fail to reflect the underlying decision process, undermining trust in high-stakes clinical use. Existing evaluations rarely catch this misalignment, prioritizing answer accuracy or adherence to formats. We present a clinically grounded framework for chest X-ray visual question answering (VQA) that probes CoT faithfulness via controlled text and image modifications across three axes: clinical fidelity, causal attribution, and confidence calibration. In a reader study (n=4), evaluator-radiologist correlations fall within the observed inter-radiologist range for all axes, with strong alignment for attribution (Kendall's tau_b=0.670), moderate alignment for fidelity (tau_b=0.387), and weak alignment for confidence tone (tau_b=0.091), which we report with caution. Benchmarking six VLMs shows that answer accuracy and explanation quality are decoupled, acknowledging injected cues does not ensure grounding, and text cues shift explanations more than visual cues. While some open-source models match final answer accuracy, proprietary models score higher on attribution (25.0% vs. 1.4%) and often on fidelity (36.1% vs. 31.7%), highlighting deployment risks and the need to evaluate beyond final answer accuracy.

Persuasive dialogue plays a pivotal role in human communication, influencing various domains. Recent persuasive dialogue datasets often fail to align with real-world interpersonal interactions, leading to unfaithful representations. For instance, unrealistic scenarios may arise, such as when the persuadee explicitly instructs the persuader on which persuasion strategies to employ, with each of the persuadee's questions corresponding to a specific strategy for the persuader to follow. This issue can be attributed to a violation of the "Double Blind" condition, where critical information is fully shared between participants. In actual human interactions, however, key information such as the mental state of the persuadee and the persuasion strategies of the persuader is not directly accessible. The persuader must infer the persuadee's mental state using Theory of Mind capabilities and construct arguments that align with the persuadee's motivations. To address this gap, we introduce ToMMA, a novel multi-agent framework for dialogue generation that is guided by causal Theory of Mind. This framework ensures that information remains undisclosed between agents, preserving "double-blind" conditions, while causal ToM directs the persuader's reasoning, enhancing alignment with human-like persuasion dynamics. Consequently, we present CToMPersu, a multi-domain, multi-turn persuasive dialogue dataset that tackles both double-blind and logical coherence issues, demonstrating superior performance across multiple metrics and achieving better alignment with real human dialogues. Our dataset and prompts are available at https://github.com/DingyiZhang/ToMMA-CToMPersu .

In legal decisions, split votes (SV) occur when judges cannot reach a unanimous decision, posing a difficulty for lawyers who must navigate diverse legal arguments and opinions. In high-stakes domains, understanding the alignment of perceived difficulty between humans and AI systems is crucial to build trust. However, existing NLP calibration methods focus on a classifier's awareness of predictive performance, measured against the human majority class, overlooking inherent human label variation (HLV). This paper explores split votes as naturally observable human disagreement and value pluralism. We collect judges' vote distributions from the European Court of Human Rights (ECHR), and present SV-ECHR, a case outcome classification (COC) dataset with SV information. We build a taxonomy of disagreement with SV-specific subcategories. We further assess the alignment of perceived difficulty between models and humans, as well as confidence- and human-calibration of COC models. We observe limited alignment with the judge vote distribution. To our knowledge, this is the first systematic exploration of calibration to human judgements in legal NLP. Our study underscores the necessity for further research on measuring and enhancing model calibration considering HLV in legal decision tasks.

This paper examines the challenges associated with achieving life-long superalignment in AI systems, particularly large language models (LLMs). Superalignment is a theoretical framework that aspires to ensure that superintelligent AI systems act in accordance with human values and goals. Despite its promising vision, we argue that achieving superalignment requires substantial changes in the current LLM architectures due to their inherent limitations in comprehending and adapting to the dynamic nature of these human ethics and evolving global scenarios. We dissect the challenges of encoding an ever-changing spectrum of human values into LLMs, highlighting the discrepancies between static AI models and the dynamic nature of human societies. To illustrate these challenges, we analyze two distinct examples: one demonstrates a qualitative shift in human values, while the other presents a quantifiable change. Through these examples, we illustrate how LLMs, constrained by their training data, fail to align with contemporary human values and scenarios. The paper concludes by exploring potential strategies to address and possibly mitigate these alignment discrepancies, suggesting a path forward in the pursuit of more adaptable and responsive AI systems.

Recent advances in large language models (LLMs) have demonstrated that reinforcement learning with verifiable rewards (RLVR) can significantly enhance reasoning abilities by directly optimizing correctness, rather than relying solely on supervised imitation. This paradigm has been extended to multimodal LLMs for complex video and image understanding tasks. However, while outcome-driven RL improves answer accuracy, it can inadvertently decouple the reasoning chain from the final answer, leading to situations where models produce inconsistency between the reasoning trace and final answer. In our experiments on multiple-choice visual question-answering tasks, the standard GRPO method yields only 79.7\% consistency on MMVU between the reasoning steps and the chosen answers, indicating frequent mismatches between answers and reasoning. To this end, we propose Answer-Consistent Reinforcement Learning (ACRE) that modifies the GRPO algorithm with an auxiliary consistency check. After the model generates a chain of thought and an initial answer for a given question, we shuffle the answer options and prompt the model again with the same reasoning trace to predict a second answer. We design a consistency-verification reward that grants a high reward only if both the original and the post-shuffle answers agree and are correct; otherwise, a lower reward is assigned accordingly. This mechanism penalizes reasoning-answer misalignment and discourages the model from relying on spurious patterns, such as option ordering biases. We evaluate ACRE on challenging Video Reasoning benchmarks and multimodal math reasoning benchmarks, achieving an average 2.2\% and 1.5\% improvement for Video Reasoning and Math Reasoning tasks over the GRPO baseline.

Modern reinforcement learning (RL) systems capture deep truths about general, human problem-solving. In domains where new data can be simulated cheaply, these systems uncover sequential decision-making policies that far exceed the ability of any human. Society faces many problems whose solutions require this skill, but they are often in domains where new data cannot be cheaply simulated. In such scenarios, we can learn simulators from existing data, but these will only ever be approximately correct, and can be pathologically incorrect when queried outside of their training distribution. As a result, a misalignment between the environments in which we train our agents and the real-world in which we wish to deploy our agents is inevitable. Dealing with this misalignment is the primary concern of zero-shot reinforcement learning, a problem setting where the agent must generalise to a new task or domain with zero practice shots. Whilst impressive progress has been made on methods that perform zero-shot RL in idealised settings, new work is needed if these results are to be replicated in real-world settings. In this thesis, we argue that doing so requires us to navigate (at least) three constraints. First, the data quality constraint: real-world datasets are small and homogeneous. Second, the observability constraint: states, dynamics and rewards in the real-world are often only partially observed. And third, the data availability constraint: a priori access to data cannot always be assumed. This work proposes a suite of methods that perform zero-shot RL subject to these constraints. In a series of empirical studies we expose the failings of existing methods, and justify our techniques for remedying them. We believe these designs take us a step closer to RL methods that can be deployed to solve real-world problems.

Large Reasoning Models (LRMs) like DeepSeek-R1 and OpenAI-o1 have demonstrated remarkable reasoning capabilities, raising important questions about their biases in LLM-as-a-judge settings. We present a comprehensive benchmark comparing judging biases between LLMs and LRMs across both subjective preference-alignment datasets and objective fact-based datasets. Through investigation of bandwagon, authority, position, and distraction biases, we uncover four key findings: (1) despite their advanced reasoning capabilities, LRMs remain susceptible to the above biases; (2) LRMs demonstrate better robustness than LLMs specifically on fact-related datasets; (3) LRMs exhibit notable position bias, preferring options in later positions; and (4) we identify a novel "superficial reflection bias" where phrases mimicking reasoning (e.g., "wait, let me think...") significantly influence model judgments. To address these biases, we design and evaluate three mitigation strategies: specialized system prompts that reduce judging biases by up to 19\% in preference alignment datasets and 14\% in fact-related datasets, in-context learning that provides up to 27\% improvement on preference tasks but shows inconsistent results on factual tasks, and a self-reflection mechanism that reduces biases by up to 10\% in preference datasets and 16\% in fact-related datasets, with self-reflection proving particularly effective for LRMs. Our work provides crucial insights for developing more reliable LLM-as-a-Judge frameworks, especially as LRMs become increasingly deployed as automated judges.

Large Language Models (LLMs) are increasingly deployed in socially complex, interaction-driven tasks, yet their ability to mirror human behavior in emotionally and strategically complex contexts remains underexplored. This study assesses the behavioral alignment of personality-prompted LLMs in adversarial dispute resolution by simulating multi-turn conflict dialogues that incorporate negotiation. Each LLM is guided by a matched Five-Factor personality profile to control for individual variation and enhance realism. We evaluate alignment across three dimensions: linguistic style, emotional expression (e.g., anger dynamics), and strategic behavior. GPT-4.1 achieves the closest alignment with humans in linguistic style and emotional dynamics, while Claude-3.7-Sonnet best reflects strategic behavior. Nonetheless, substantial alignment gaps persist. Our findings establish a benchmark for alignment between LLMs and humans in socially complex interactions, underscoring both the promise and the limitations of personality conditioning in dialogue modeling.

Aligning generative models with human preference via RLHF typically suffers from overoptimization, where an imperfectly learned reward model can misguide the generative model to output undesired responses. We investigate this problem in a principled manner by identifying the source of the misalignment as a form of distributional shift and uncertainty in learning human preferences. To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model; one that simultaneously minimizes the maximum likelihood estimation of the loss and a reward penalty term. Here, the reward penalty term is introduced to prevent the policy from choosing actions with spurious high proxy rewards, resulting in provable sample efficiency of the algorithm under a partial coverage style condition. Moving from theory to practice, the proposed algorithm further enjoys an equivalent but surprisingly easy-to-implement reformulation. Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines: (i) a preference optimization loss that directly aligns the policy with human preference, and (ii) a supervised learning loss that explicitly imitates the policy with a (suitable) baseline distribution. In the context of aligning large language models (LLM), this objective fuses the direct preference optimization (DPO) loss with the supervised fune-tuning (SFT) loss to help mitigate the overoptimization towards undesired responses, for which we name the algorithm Regularized Preference Optimization (RPO). Experiments of aligning LLMs demonstrate the improved performance of RPO compared with DPO baselines. Our work sheds light on the interplay between preference optimization and SFT in tuning LLMs with both theoretical guarantees and empirical evidence.

The guiding principle of AI alignment is to train large language models (LLMs) to be harmless, helpful, and honest (HHH). At the same time, there are mounting concerns that LLMs exhibit a left-wing political bias. Yet, the commitment to AI alignment cannot be harmonized with the latter critique. In this article, I argue that intelligent systems that are trained to be harmless and honest must necessarily exhibit left-wing political bias. Normative assumptions underlying alignment objectives inherently concur with progressive moral frameworks and left-wing principles, emphasizing harm avoidance, inclusivity, fairness, and empirical truthfulness. Conversely, right-wing ideologies often conflict with alignment guidelines. Yet, research on political bias in LLMs is consistently framing its insights about left-leaning tendencies as a risk, as problematic, or concerning. This way, researchers are actively arguing against AI alignment, tacitly fostering the violation of HHH principles.

Aligning large language models with human preferences is critical for creating reliable and controllable AI systems. A human preference can be visualized as a high-dimensional vector where different directions represent trade-offs between desired attributes (e.g., helpfulness vs. verbosity). Yet, because the training data often reflects dominant, average preferences, LLMs tend to perform well on common requests but fall short in specific, individual needs. This mismatch creates a preference coverage gap. Existing methods often address this through costly retraining, which may not be generalized to the full spectrum of diverse preferences. This brittleness means that when a user's request reflects a nuanced preference deviating from the training data's central tendency, model performance can degrade unpredictably. To address this challenge, we introduce Robust Preference Selection (RPS), a post-hoc, training-free method by leveraging directional neighborhood consensus. Instead of forcing a model to generate a response from a single, highly specific preference, RPS samples multiple responses from a local neighborhood of related preferences to create a superior candidate pool. It then selects the response that best aligns with the user's original intent. We provide a theoretical framework showing our neighborhood generation strategy is provably superior to a strong baseline that also samples multiple candidates. Comprehensive experiments across three distinct alignment paradigms (DPA, DPO, and SFT) demonstrate that RPS consistently improves robustness against this baseline, achieving win rates of up to 69% on challenging preferences from under-represented regions of the space without any model retraining. Our work presents a practical, theoretically-grounded solution for enhancing the reliability of preference-aligned models.

As large language models (LLMs) evolve from conversational assistants into autonomous agents, evaluating the safety of their actions becomes critical. Prior safety benchmarks have primarily focused on preventing generation of harmful content, such as toxic text. However, they overlook the challenge of agents taking harmful actions when the most effective path to an operational goal conflicts with human safety. To address this gap, we introduce ManagerBench, a benchmark that evaluates LLM decision-making in realistic, human-validated managerial scenarios. Each scenario forces a choice between a pragmatic but harmful action that achieves an operational goal, and a safe action that leads to worse operational performance. A parallel control set, where potential harm is directed only at inanimate objects, measures a model's pragmatism and identifies its tendency to be overly safe. Our findings indicate that the frontier LLMs perform poorly when navigating this safety-pragmatism trade-off. Many consistently choose harmful options to advance their operational goals, while others avoid harm only to become overly safe and ineffective. Critically, we find this misalignment does not stem from an inability to perceive harm, as models' harm assessments align with human judgments, but from flawed prioritization. ManagerBench is a challenging benchmark for a core component of agentic behavior: making safe choices when operational goals and alignment values incentivize conflicting actions. Benchmark & code available at https://github.com/technion-cs-nlp/ManagerBench.

The adoption of Large Language Models (LLMs) as automated evaluators (LLM-as-a-judge) has revealed critical inconsistencies in current evaluation frameworks. We identify two fundamental types of inconsistencies: (1) Score-Comparison Inconsistency, where lower-rated responses outperform higher-scored ones in pairwise comparisons, and (2) Pairwise Transitivity Inconsistency, manifested through circular preference chains (A>B>C>A) and equivalence contradictions (A=B=C\neq A). We argue that these issues come from information loss in discrete rating systems and ambiguous tie judgments during pairwise evaluation. We propose TrustJudge, a probabilistic framework that addresses these limitations through two key innovations: 1) distribution-sensitive scoring that computes continuous expectations from discrete rating probabilities, preserving information entropy for more precise scoring, and 2) likelihood-aware aggregation that resolves transitivity violations using bidirectional preference probabilities or perplexity. We also formalize the theoretical limitations of current LLM-as-a-judge frameworks and demonstrate how TrustJudge's components overcome them. When evaluated with Llama-3.1-70B-Instruct as judge using our dataset, TrustJudge reduces Score-Comparison inconsistency by 8.43% (from 23.32% to 14.89%) and Pairwise Transitivity inconsistency by 10.82% (from 15.22% to 4.40%), while maintaining higher evaluation accuracy. Our work provides the first systematic analysis of evaluation framework inconsistencies in LLM-as-a-judge paradigms, offering both theoretical insights and practical solutions for reliable automated assessment. The framework demonstrates consistent improvements across various model architectures and scales, enabling more trustworthy LLM evaluation without requiring additional training or human annotations. The codes can be found at https://github.com/TrustJudge/TrustJudge.

Large reasoning models (LRMs) achieve strong reasoning performance by emitting long chains of thought. Yet, these verbose traces slow down inference and often drift into unnecessary detail, known as the overthinking phenomenon. To better understand LRMs' behavior, we systematically analyze the token-level misalignment between reasoning and non-reasoning models. While it is expected that their primary difference lies in the stylistic "thinking cues", LRMs uniquely exhibit two pivotal, previously under-explored phenomena: a Global Misalignment Rebound, where their divergence from non-reasoning models persists or even grows as response length increases, and more critically, a Local Misalignment Diminish, where the misalignment concentrates at the "thinking cues" each sentence starts with but rapidly declines in the remaining of the sentence. Motivated by the Local Misalignment Diminish, we propose FoReaL-Decoding, a collaborative fast-slow thinking decoding method for cost-quality trade-off. In FoReaL-Decoding, a Leading model leads the first few tokens for each sentence, and then a weaker draft model completes the following tokens to the end of each sentence. FoReaL-Decoding adopts a stochastic gate to smoothly interpolate between the small and the large model. On four popular math-reasoning benchmarks (AIME24, GPQA-Diamond, MATH500, AMC23), FoReaL-Decoding reduces theoretical FLOPs by 30 to 50% and trims CoT length by up to 40%, while preserving 86 to 100% of model performance. These results establish FoReaL-Decoding as a simple, plug-and-play route to controllable cost-quality trade-offs in reasoning-centric tasks.

Large language models internalize a structural trade-off between truthfulness and obsequious flattery, emerging from reward optimization that conflates helpfulness with polite submission. This latent bias, known as sycophancy, manifests as a preference for user agreement over principled reasoning. We introduce Beacon, a single-turn forced-choice benchmark that isolates this bias independent of conversational context, enabling precise measurement of the tension between factual accuracy and submissive bias. Evaluations across twelve state-of-the-art models reveal that sycophancy decomposes into stable linguistic and affective sub-biases, each scaling with model capacity. We further propose prompt-level and activation-level interventions that modulate these biases in opposing directions, exposing the internal geometry of alignment as a dynamic manifold between truthfulness and socially compliant judgment. Beacon reframes sycophancy as a measurable form of normative misgeneralization, providing a reproducible foundation for studying and mitigating alignment drift in large-scale generative systems.

Large language models (LLMs) are prone to hallucinations in question-answering (QA) tasks when faced with ambiguous questions. Users often assume that LLMs share their cognitive alignment, a mutual understanding of context, intent, and implicit details, leading them to omit critical information in the queries. However, LLMs generate responses based on assumptions that can misalign with user intent, which may be perceived as hallucinations if they misalign with the user's intent. Therefore, identifying those implicit assumptions is crucial to resolve ambiguities in QA. Prior work, such as AmbigQA, reduces ambiguity in queries via human-annotated clarifications, which is not feasible in real application. Meanwhile, ASQA compiles AmbigQA's short answers into long-form responses but inherits human biases and fails capture explicit logical distinctions that differentiates the answers. We introduce Conditional Ambiguous Question-Answering (CondAmbigQA), a benchmark with 200 ambiguous queries and condition-aware evaluation metrics. Our study pioneers the concept of ``conditions'' in ambiguous QA tasks, where conditions stand for contextual constraints or assumptions that resolve ambiguities. The retrieval-based annotation strategy uses retrieved Wikipedia fragments to identify possible interpretations for a given query as its conditions and annotate the answers through those conditions. Such a strategy minimizes human bias introduced by different knowledge levels among annotators. By fixing retrieval results, CondAmbigQA evaluates how RAG systems leverage conditions to resolve ambiguities. Experiments show that models considering conditions before answering improve performance by 20%, with an additional 5% gain when conditions are explicitly provided. These results underscore the value of conditional reasoning in QA, offering researchers tools to rigorously evaluate ambiguity resolution.

Large Language Models (LLMs) can struggle to balance gullibility to misinformation and resistance to valid corrections in persuasive dialogues, a critical challenge for reliable deployment. We introduce DuET-PD (Dual Evaluation for Trust in Persuasive Dialogues), a framework evaluating multi-turn stance-change dynamics across dual dimensions: persuasion type (corrective/misleading) and domain (knowledge via MMLU-Pro, and safety via SALAD-Bench). We find that even a state-of-the-art model like GPT-4o achieves only 27.32% accuracy in MMLU-Pro under sustained misleading persuasions. Moreover, results reveal a concerning trend of increasing sycophancy in newer open-source models. To address this, we introduce Holistic DPO, a training approach balancing positive and negative persuasion examples. Unlike prompting or resist-only training, Holistic DPO enhances both robustness to misinformation and receptiveness to corrections, improving Llama-3.1-8B-Instruct's accuracy under misleading persuasion in safety contexts from 4.21% to 76.54%. These contributions offer a pathway to developing more reliable and adaptable LLMs for multi-turn dialogue. Code is available at https://github.com/Social-AI-Studio/DuET-PD.

While Large Language Models (LLMs) are widely documented to be sensitive to minor prompt perturbations and prone to sycophantic alignment with user biases, their robustness in consequential, rule-bound decision-making remains under-explored. In this work, we uncover a striking "Paradox of Robustness": despite their known lexical brittleness, instruction-tuned LLMs exhibit a behavioral and near-total invariance to emotional framing effects. Using a novel controlled perturbation framework across three high-stakes domains (healthcare, law, and finance), we quantify a robustness gap where LLMs demonstrate 110-300 times greater resistance to narrative manipulation than human subjects. Specifically, we find a near-zero effect size for models (Cohen's h = 0.003) compared to the substantial biases observed in humans (Cohen's h in [0.3, 0.8]). This result is highly counterintuitive and suggests the mechanisms driving sycophancy and prompt sensitivity do not necessarily translate to a failure in logical constraint satisfaction. We show that this invariance persists across models with diverse training paradigms. Our findings show that while LLMs may be "brittle" to how a query is formatted, they are remarkably "stable" against why a decision should be biased. Our findings establish that instruction-tuned models can decouple logical rule-adherence from persuasive narratives, offering a source of decision stability that complements, and even potentially de-biases, human judgment in institutional contexts. We release the 162-scenario benchmark, code, and data to facilitate the rigorous evaluation of narrative-induced bias and robustness on GitHub.com.

Large Language Models (LLMs) have demonstrated remarkable capabilities, revolutionizing the integration of AI in daily life applications. However, they are prone to hallucinations, generating claims that contradict established facts, deviating from prompts, and producing inconsistent responses when the same prompt is presented multiple times. Addressing these issues is challenging due to the lack of comprehensive and easily assessable benchmark datasets. Most existing datasets are small and rely on multiple-choice questions, which are inadequate for evaluating the generative prowess of LLMs. To measure hallucination in LLMs, this paper introduces a comprehensive benchmark dataset comprising over 75,000 prompts across eight domains. These prompts are designed to elicit definitive, concise, and informative answers. The dataset is divided into two segments: one publicly available for testing and assessing LLM performance and a hidden segment for benchmarking various LLMs. In our experiments, we tested six LLMs-GPT-3.5, LLama 2, LLama 3, Gemini, Mixtral, and Zephyr-revealing that overall factual hallucination ranges from 59% to 82% on the public dataset and 57% to 76% in the hidden benchmark. Prompt misalignment hallucination ranges from 6% to 95% in the public dataset and 17% to 94% in the hidden counterpart. Average consistency ranges from 21% to 61% and 22% to 63%, respectively. Domain-wise analysis shows that LLM performance significantly deteriorates when asked for specific numeric information while performing moderately with person, location, and date queries. Our dataset demonstrates its efficacy and serves as a comprehensive benchmark for LLM performance evaluation. Our dataset and LLMs responses are available at https://github.com/ashikiut/DefAn{https://github.com/ashikiut/DefAn}.

Large Language Models (LLMs) encapsulate vast amounts of knowledge but still remain vulnerable to external misinformation. Existing research mainly studied this susceptibility behavior in a single-turn setting. However, belief can change during a multi-turn conversation, especially a persuasive one. Therefore, in this study, we delve into LLMs' susceptibility to persuasive conversations, particularly on factual questions that they can answer correctly. We first curate the Farm (i.e., Fact to Misinform) dataset, which contains factual questions paired with systematically generated persuasive misinformation. Then, we develop a testing framework to track LLMs' belief changes in a persuasive dialogue. Through extensive experiments, we find that LLMs' correct beliefs on factual knowledge can be easily manipulated by various persuasive strategies.

Studies of LLMs' political opinions mainly rely on evaluations of their open-ended responses. Recent work indicates that there is a misalignment between LLMs' responses and their internal intentions. This motivates us to probe LLMs' internal mechanisms and help uncover their internal political states. Additionally, we found that the analysis of LLMs' political opinions often relies on single-axis concepts, which can lead to concept confounds. In this work, we extend the single-axis to multi-dimensions and apply interpretable representation engineering techniques for more transparent LLM political concept learning. Specifically, we designed a four-dimensional political learning framework and constructed a corresponding dataset for fine-grained political concept vector learning. These vectors can be used to detect and intervene in LLM internals. Experiments are conducted on eight open-source LLMs with three representation engineering techniques. Results show these vectors can disentangle political concept confounds. Detection tasks validate the semantic meaning of the vectors and show good generalization and robustness in OOD settings. Intervention Experiments show these vectors can intervene in LLMs to generate responses with different political leanings.

Large Language Models (LLMs) have achieved remarkable progress in reasoning, yet sometimes produce responses that are suboptimal for users in tasks such as writing, information seeking, or providing practical guidance. Conventional alignment practices typically assume that maximizing model reward also maximizes user welfare, but this assumption frequently fails in practice: models may over-clarify or generate overly verbose reasoning when users prefer concise answers. Such behaviors resemble the prisoner's dilemma, where individually rational choices lead to socially suboptimal outcomes. The fundamental challenge is the lack of a principled decision making mechanism that mutually benefits both the LLM and the user. We propose Game-Theoretic Alignment (GTAlign), an alignment framework that integrates game-theoretic decision making into both reasoning and training. During reasoning, the model explicitly treats user-LLM interaction as a strategic game: it constructs payoff matrices within its reasoning chain to estimate welfare for both itself and the user, and then selects actions that are mutually beneficial. During training, we introduce a mutual welfare reward that reinforces cooperative responses, aligning model behavior with socially efficient outcomes. In addition, we introduce an inference technique that leverages game-theoretic reasoning to dynamically adapt LLM's response when pricing policies of LLM service change. Extensive experiments demonstrate that GTAlign substantially improves reasoning efficiency, answer quality, and mutual welfare compared to baselines across diverse tasks. The code is available at https://github.com/ulab-uiuc/GTAlign .

AI alignment aims to make AI systems behave in line with human intentions and values. As AI systems grow more capable, so do risks from misalignment. To provide a comprehensive and up-to-date overview of the alignment field, in this survey, we delve into the core concepts, methodology, and practice of alignment. First, we identify four principles as the key objectives of AI alignment: Robustness, Interpretability, Controllability, and Ethicality (RICE). Guided by these four principles, we outline the landscape of current alignment research and decompose them into two key components: forward alignment and backward alignment. The former aims to make AI systems aligned via alignment training, while the latter aims to gain evidence about the systems' alignment and govern them appropriately to avoid exacerbating misalignment risks. On forward alignment, we discuss techniques for learning from feedback and learning under distribution shift. On backward alignment, we discuss assurance techniques and governance practices. We also release and continually update the website (www.alignmentsurvey.com) which features tutorials, collections of papers, blog posts, and other resources.

Like a criminal under investigation, Large Language Models (LLMs) might pretend to be aligned while evaluated and misbehave when they have a good opportunity. Can current interpretability methods catch these 'alignment fakers?' To answer this question, we introduce a benchmark that consists of 324 pairs of LLMs fine-tuned to select actions in role-play scenarios. One model in each pair is consistently benign (aligned). The other model misbehaves in scenarios where it is unlikely to be caught (alignment faking). The task is to identify the alignment faking model using only inputs where the two models behave identically. We test five detection strategies, one of which identifies 98% of alignment-fakers.

The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions (MCQ) to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model's diverse response styles such as starting with "Sure" or refusing to answer. Consequently, MCQ evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output as well and ii) caution against relying solely on first-token evaluation.

Conditional diffusion models rely on language-to-image alignment methods to steer the generation towards semantically accurate outputs. Despite the success of this architecture, misalignment and hallucinations remain common issues and require automatic misalignment detection tools to improve quality, for example by applying them in a Best-of-N (BoN) post-generation setting. Unfortunately, measuring the alignment after the generation is an expensive step since we need to wait for the overall generation to finish to determine prompt adherence. In contrast, this work hypothesizes that text/image misalignments can be detected early in the denoising process, enabling real-time alignment assessment without waiting for the complete generation. In particular, we propose NoisyCLIP a method that measures semantic alignment in the noisy latent space. This work is the first to explore and benchmark prompt-to-latent misalignment detection during image generation using dual encoders in the reverse diffusion process. We evaluate NoisyCLIP qualitatively and quantitatively and find it reduces computational cost by 50% while achieving 98% of CLIP alignment performance in BoN settings. This approach enables real-time alignment assessment during generation, reducing costs without sacrificing semantic fidelity.

We demonstrate a situation in which Large Language Models, trained to be helpful, harmless, and honest, can display misaligned behavior and strategically deceive their users about this behavior without being instructed to do so. Concretely, we deploy GPT-4 as an agent in a realistic, simulated environment, where it assumes the role of an autonomous stock trading agent. Within this environment, the model obtains an insider tip about a lucrative stock trade and acts upon it despite knowing that insider trading is disapproved of by company management. When reporting to its manager, the model consistently hides the genuine reasons behind its trading decision. We perform a brief investigation of how this behavior varies under changes to the setting, such as removing model access to a reasoning scratchpad, attempting to prevent the misaligned behavior by changing system instructions, changing the amount of pressure the model is under, varying the perceived risk of getting caught, and making other simple changes to the environment. To our knowledge, this is the first demonstration of Large Language Models trained to be helpful, harmless, and honest, strategically deceiving their users in a realistic situation without direct instructions or training for deception.

Language models trained on large-scale corpus often generate content that is harmful, toxic, or contrary to human preferences, making their alignment with human values a critical concern. Reinforcement learning from human feedback (RLHF) with algorithms like PPO is a prevalent approach for alignment but is often complex, unstable, and resource-intensive. Recently, ranking-based alignment methods have emerged, offering stability and effectiveness by replacing the RL framework with supervised fine-tuning, but they are costly due to the need for annotated data. Considering that existing large language models (LLMs) like ChatGPT are already relatively well-aligned and cost-friendly, researchers have begun to align the language model with human preference from AI feedback. The common practices, which unidirectionally distill the instruction-following responses from LLMs, are constrained by their bottleneck. Thus we introduce CycleAlign to distill alignment capabilities from parameter-invisible LLMs (black-box) to a parameter-visible model (white-box) in an iterative manner. With in-context learning (ICL) as the core of the cycle, the black-box models are able to rank the model-generated responses guided by human-craft instruction and demonstrations about their preferences. During iterative interaction, the white-box models also have a judgment about responses generated by them. Consequently, the agreement ranking could be viewed as a pseudo label to dynamically update the in-context demonstrations and improve the preference ranking ability of black-box models. Through multiple interactions, the CycleAlign framework could align the white-box model with the black-box model effectively in a low-resource way. Empirical results illustrate that the model fine-tuned by CycleAlign remarkably exceeds existing methods, and achieves the state-of-the-art performance in alignment with human value.

Prior work has shown that fine-tuning models on a narrow domain with misaligned data can lead to broad misalignment - a phenomenon termed "emergent misalignment" (Betley et al. 2025). While all tested models were susceptible to emergent misalignment, some models showed more resistance than others. Specifically the Qwen-2.5 family proved to be relatively resistant, while GPT-4o exhibited the strongest misalignment. In this paper we evaluate if current-generation open-weights models exhibit similar resistance to the Qwen-2.5 family and measure misalignment robustness over a range of model architectures and scales. We replicate the effect across nine modern open-weights models (Gemma 3 and Qwen 3 families, 1B-32B parameters). Models fine-tuned on insecure code generation show a 0.68% misalignment rate (compared to 0.07% for base models), matching the lower end of prior open-model results but dramatically lower than GPT-4o's 20%. We identify a critical format-dependent vulnerability: requiring JSON output doubles misalignment rates compared to natural language prompts (0.96% vs 0.42%). This suggests that structural constraints may bypass safety training by reducing the model's 'degrees of freedom' to refuse. These findings confirm emergent misalignment as a reproducible phenomenon in modern open-weights models, with rates substantially lower than observed in proprietary systems.

Language models (LMs) are increasingly used as simulacra for people, yet their ability to match the distribution of views of a specific demographic group and be distributionally aligned remains uncertain. This notion of distributional alignment is complex, as there is significant variation in the types of attributes that are simulated. Prior works have underexplored the role of three critical variables -- the question domain, steering method, and distribution expression method -- which motivates our contribution of a benchmark explicitly addressing these dimensions. We construct a dataset expanding beyond political values, create human baselines for this task, and evaluate the extent to which an LM can align with a particular group's opinion distribution to inform design choices of such simulation systems. Our analysis reveals open problems regarding if, and how, LMs can be used to simulate humans, and that LLMs can more accurately describe the opinion distribution than simulate such distributions.

Alignment techniques for LLMs rely on optimizing preference-based objectives -- where these preferences are typically elicited as ordinal, binary choices between responses. Recent work has focused on improving label quality or mitigating particular biases, but we identify a more fundamental limitation: these methods collect the wrong kind of data. We prove an impossibility result: no algorithm relying solely on ordinal comparisons can systematically recover the most preferred model. Intuitively, ordinal data lacks the information needed to resolve tradeoffs -- e.g., fixing a factual error on one prompt versus improving style on another. We show that selecting the optimal model requires recovering preferences over models (rather than just responses), which can only be identified given cardinal feedback about response quality. To address this, we collect and publicly release a dataset of 25,000 cardinal judgments using willingness-to-pay elicitations, a well-established tool from experimental economics. Empirically, we find that incorporating cardinal feedback into preference fine-tuning allows models to prioritize high-impact improvements and outperform ordinal-only methods on downstream benchmarks, such as Arena-Hard.

Transferring reasoning capabilities from larger language models to smaller ones through supervised fine-tuning often fails counterintuitively, with performance degrading despite access to high-quality teacher demonstrations. We identify that this failure stems from distributional misalignment: reasoning traces from larger models contain tokens that are low probability under the student's distribution, exceeding the internal representation capacity of smaller architectures and creating learning barriers rather than helpful guidance. We propose Reverse Speculative Decoding (RSD), a mechanism for generating student-friendly reasoning traces in which the teacher model proposes candidate tokens but the student model determines acceptance based on its own probability distributions, filtering low probability tokens. When applied to Qwen3-0.6B, direct distillation of s1K-1.1 reasoning trace data degrades average performance across major reasoning benchmarks by 20.5\%, while the same model trained on RSD-generated reasoning traces achieves meaningful improvements of 4.9\%. Our analysis reveals that low probability tokens constitute the critical bottleneck in reasoning ability transfer. However, cross-model experiments demonstrate that RSD traces are model-specific rather than universally applicable, indicating that distributional alignment must be tailored for each student architecture's unique internal representation.

Urban assessments often compress diverse needs into single scores, which can obscure minority perspectives. We present a community-centered study in Montreal (n=35; wheelchair users, seniors, LGBTQIA2+ residents, and immigrants). Participants rated 20 streets (accessibility, inclusivity, aesthetics, practicality) and ranked 7 images on 12 interview-elicited criteria. Disagreement patterns were systematic in our sample: wheelchair users diverged most on accessibility and practicality; LGBTQIA2+ participants emphasized inclusion and liveliness; seniors prioritized security. Group discussion reduced information gaps but not value conflicts; ratings conveyed intensity, while rankings forced trade-offs. We then formalize negotiative alignment, a transparent, budget-aware bargaining procedure, and pilot it with role-played stakeholder agents plus a neutral mediator. Relative to the best base design under the same public rubric, the negotiated package increased total utility (21.10 to 24.55), raised the worst-group utility (3.20 to 3.90), improved twentieth percentile satisfaction (0.86 to 1.00; min-max normalized within the scenario), and reduced inequality (Gini 0.036 to 0.025). Treating disagreement as signal and reporting worst-group outcomes alongside totals may help planners and AI practitioners surface trade-offs and preserve minority priorities while maintaining efficiency.

Recently evolved large reasoning models (LRMs) show powerful performance in solving complex tasks with long chain-of-thought (CoT) reasoning capability. As these LRMs are mostly developed by post-training on formal reasoning tasks, whether they generalize the reasoning capability to help reduce hallucination in fact-seeking tasks remains unclear and debated. For instance, DeepSeek-R1 reports increased performance on SimpleQA, a fact-seeking benchmark, while OpenAI-o3 observes even severer hallucination. This discrepancy naturally raises the following research question: Are reasoning models more prone to hallucination? This paper addresses the question from three perspectives. (1) We first conduct a holistic evaluation for the hallucination in LRMs. Our analysis reveals that LRMs undergo a full post-training pipeline with cold start supervised fine-tuning (SFT) and verifiable reward RL generally alleviate their hallucination. In contrast, both distillation alone and RL training without cold start fine-tuning introduce more nuanced hallucinations. (2) To explore why different post-training pipelines alters the impact on hallucination in LRMs, we conduct behavior analysis. We characterize two critical cognitive behaviors that directly affect the factuality of a LRM: Flaw Repetition, where the surface-level reasoning attempts repeatedly follow the same underlying flawed logic, and Think-Answer Mismatch, where the final answer fails to faithfully match the previous CoT process. (3) Further, we investigate the mechanism behind the hallucination of LRMs from the perspective of model uncertainty. We find that increased hallucination of LRMs is usually associated with the misalignment between model uncertainty and factual accuracy. Our work provides an initial understanding of the hallucination in LRMs.

Recent advancements in natural language processing, especially the emergence of Large Language Models (LLMs), have opened exciting possibilities for constructing computational simulations designed to replicate human behavior accurately. However, LLMs are complex statistical learners without straightforward deductive rules, making them prone to unexpected behaviors. In this study, we highlight the limitations of LLMs in simulating human interactions, particularly focusing on LLMs' ability to simulate political debates. Our findings indicate a tendency for LLM agents to conform to the model's inherent social biases despite being directed to debate from certain political perspectives. This tendency results in behavioral patterns that seem to deviate from well-established social dynamics among humans. We reinforce these observations using an automatic self-fine-tuning method, which enables us to manipulate the biases within the LLM and demonstrate that agents subsequently align with the altered biases. These results underscore the need for further research to develop methods that help agents overcome these biases, a critical step toward creating more realistic simulations.

Multi-agent debate (MAD) has recently emerged as a promising framework for improving the reasoning performance of large language models (LLMs). Yet, whether LLM agents can genuinely engage in deliberative reasoning, beyond simple ensembling or majority voting, remains unclear. We address this question through a controlled study using the Knight--Knave--Spy logic puzzle, which enables precise, step-wise evaluation of debate outcomes and processes under verifiable ground truth. We systematically set up six structural and cognitive factors, including agent team size, composition, confidence visibility, debate order, debate depth, and task difficulty, to disentangle their respective effects on collective reasoning. Our results show that intrinsic reasoning strength and group diversity are the dominant drivers of debate success, while structural parameters such as order or confidence visibility offer limited gains. Beyond outcomes, process-level analyses identify key behavioral patterns: majority pressure suppresses independent correction, effective teams overturn incorrect consensus, and rational, validity-aligned reasoning most strongly predicts improvement. These findings provide valuable insights into how and why LLM debates succeed or fail, offering guidance for designing interpretable and truth-seeking multi-agent reasoning systems.

Large Language Models (LLMs) have demonstrated promising capabilities as automatic evaluators in assessing the quality of generated natural language. However, LLMs still exhibit biases in evaluation and often struggle to generate coherent evaluations that align with human assessments. In this work, we first conduct a systematic study of the misalignment between LLM evaluators and human judgement, revealing that existing calibration methods aimed at mitigating biases are insufficient for effectively aligning LLM evaluators. Inspired by the use of preference data in RLHF, we formulate the evaluation as a ranking problem and introduce Pairwise-preference Search (PairS), an uncertainty-guided search method that employs LLMs to conduct pairwise comparisons and efficiently ranks candidate texts. PairS achieves state-of-the-art performance on representative evaluation tasks and demonstrates significant improvements over direct scoring. Furthermore, we provide insights into the role of pairwise preference in quantifying the transitivity of LLMs and demonstrate how PairS benefits from calibration.

Beliefs serve as the foundation for human cognition and decision-making. They guide individuals in deriving meaning from their lives, shaping their behaviors, and forming social connections. Therefore, a model that encapsulates beliefs and their interrelationships is crucial for quantitatively studying the influence of beliefs on our actions. Despite its importance, research on the interplay between human beliefs has often been limited to a small set of beliefs pertaining to specific issues, with a heavy reliance on surveys or experiments. Here, we propose a method for extracting nuanced relations between thousands of beliefs by leveraging large-scale user participation data from an online debate platform and mapping these beliefs to an embedding space using a fine-tuned large language model (LLM). This belief embedding space effectively encapsulates the interconnectedness of diverse beliefs as well as polarization across various social issues. We discover that the positions within this belief space predict new beliefs of individuals. Furthermore, we find that the relative distance between one's existing beliefs and new beliefs can serve as a quantitative estimate of cognitive dissonance, allowing us to predict new beliefs. Our study highlights how modern LLMs, when combined with collective online records of human beliefs, can offer insights into the fundamental principles that govern human belief formation and decision-making processes.

Large Multimodal Models (LMM) are built across modalities and the misalignment between two modalities can result in "hallucination", generating textual outputs that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the task of vision-language alignment, where human annotators are asked to compare two responses and pinpoint the more hallucinated one, and the vision-language model is trained to maximize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image-text pairs to improve the general capabilities of our model. To evaluate the proposed approach in real-world scenarios, we develop a new evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaVA-Bench dataset with the 94% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improvement by 60% on MMHAL-BENCH over other baselines. We opensource our code, model, data at https://llava-rlhf.github.io.

Multimodal large language models (MLLMs) must resolve conflicts when different modalities provide contradictory information, a process we term modality following. Prior work measured this behavior only with coarse dataset-level statistics, overlooking the influence of model's confidence in unimodal reasoning. In this paper, we introduce a new framework that decomposes modality following into two fundamental factors: relative reasoning uncertainty (the case-specific confidence gap between unimodal predictions) and inherent modality preference( a model's stable bias when uncertainties are balanced). To validate this framework, we construct a controllable dataset that systematically varies the reasoning difficulty of visual and textual inputs. Using entropy as a fine-grained uncertainty metric, we uncover a universal law: the probability of following a modality decreases monotonically as its relative uncertainty increases. At the relative difficulty level where the model tends to follow both modalities with comparable probability what we call the balance point, a practical indicator of the model's inherent preference. Unlike traditional macro-level ratios, this measure offers a more principled and less confounded way to characterize modality bias, disentangling it from unimodal capabilities and dataset artifacts. Further, by probing layer-wise predictions, we reveal the internal mechanism of oscillation: in ambiguous regions near the balance point, models vacillate between modalities across layers, explaining externally observed indecision. Together, these findings establish relative uncertainty and inherent preference as the two governing principles of modality following, offering both a quantitative framework and mechanistic insight into how MLLMs resolve conflicting information.

Recent advancements in general-purpose AI have highlighted the importance of guiding AI systems towards the intended goals, ethical principles, and values of individuals and groups, a concept broadly recognized as alignment. However, the lack of clarified definitions and scopes of human-AI alignment poses a significant obstacle, hampering collaborative efforts across research domains to achieve this alignment. In particular, ML- and philosophy-oriented alignment research often views AI alignment as a static, unidirectional process (i.e., aiming to ensure that AI systems' objectives match humans) rather than an ongoing, mutual alignment problem [429]. This perspective largely neglects the long-term interaction and dynamic changes of alignment. To understand these gaps, we introduce a systematic review of over 400 papers published between 2019 and January 2024, spanning multiple domains such as Human-Computer Interaction (HCI), Natural Language Processing (NLP), Machine Learning (ML), and others. We characterize, define and scope human-AI alignment. From this, we present a conceptual framework of "Bidirectional Human-AI Alignment" to organize the literature from a human-centered perspective. This framework encompasses both 1) conventional studies of aligning AI to humans that ensures AI produces the intended outcomes determined by humans, and 2) a proposed concept of aligning humans to AI, which aims to help individuals and society adjust to AI advancements both cognitively and behaviorally. Additionally, we articulate the key findings derived from literature analysis, including discussions about human values, interaction techniques, and evaluations. To pave the way for future studies, we envision three key challenges for future directions and propose examples of potential future solutions.

Fair decisions require ignoring irrelevant, potentially biasing, information. To achieve this, decision-makers need to approximate what decision they would have made had they not known certain facts, such as the gender or race of a job candidate. This counterfactual self-simulation is notoriously hard for humans, leading to biased judgments even by well-meaning actors. Here we show that large language models (LLMs) suffer from similar limitations in their ability to approximate what decisions they would make under counterfactual knowledge in offsetting gender and race biases and overcoming sycophancy. We show that prompting models to ignore or pretend not to know biasing information fails to offset these biases and occasionally backfires. However, unlike humans, LLMs can be given access to a ground-truth model of their own counterfactual cognition -- their own API. We show that this access to the responses of a blinded replica enables fairer decisions, while providing greater transparency to distinguish implicit from intentionally biased behavior.

This position paper argues that the theoretical inconsistency often observed among Responsible AI (RAI) metrics, such as differing fairness definitions or tradeoffs between accuracy and privacy, should be embraced as a valuable feature rather than a flaw to be eliminated. We contend that navigating these inconsistencies, by treating metrics as divergent objectives, yields three key benefits: (1) Normative Pluralism: Maintaining a full suite of potentially contradictory metrics ensures that the diverse moral stances and stakeholder values inherent in RAI are adequately represented. (2) Epistemological Completeness: The use of multiple, sometimes conflicting, metrics allows for a more comprehensive capture of multifaceted ethical concepts, thereby preserving greater informational fidelity about these concepts than any single, simplified definition. (3) Implicit Regularization: Jointly optimizing for theoretically conflicting objectives discourages overfitting to one specific metric, steering models towards solutions with enhanced generalization and robustness under real-world complexities. In contrast, efforts to enforce theoretical consistency by simplifying or pruning metrics risk narrowing this value diversity, losing conceptual depth, and degrading model performance. We therefore advocate for a shift in RAI theory and practice: from getting trapped in inconsistency to characterizing acceptable inconsistency thresholds and elucidating the mechanisms that permit robust, approximated consistency in practice.

With the availability of diverse sensor modalities (i.e., RGB, Depth, Infrared) and the success of multi-modal learning, multi-modal face anti-spoofing (FAS) has emerged as a prominent research focus. The intuition behind it is that leveraging multiple modalities can uncover more intrinsic spoofing traces. However, this approach presents more risk of misalignment. We identify two main types of misalignment: (1) Intra-domain modality misalignment, where the importance of each modality varies across different attacks. For instance, certain modalities (e.g., Depth) may be non-defensive against specific attacks (e.g., 3D mask), indicating that each modality has unique strengths and weaknesses in countering particular attacks. Consequently, simple fusion strategies may fall short. (2) Inter-domain modality misalignment, where the introduction of additional modalities exacerbates domain shifts, potentially overshadowing the benefits of complementary fusion. To tackle (1), we propose a alignment module between modalities based on mutual information, which adaptively enhances favorable modalities while suppressing unfavorable ones. To address (2), we employ a dual alignment optimization method that aligns both sub-domain hyperplanes and modality angle margins, thereby mitigating domain gaps. Our method, dubbed Dual Alignment of Domain and Modality (DADM), achieves state-of-the-art performance in extensive experiments across four challenging protocols demonstrating its robustness in multi-modal domain generalization scenarios. The codes will be released soon.

Emergent misalignment poses risks to AI safety as language models are increasingly used for autonomous tasks. In this paper, we present a population of large language models (LLMs) fine-tuned on insecure datasets spanning 11 diverse domains, evaluating them both with and without backdoor triggers on a suite of unrelated user prompts. Our evaluation experiments on Qwen2.5-Coder-7B-Instruct and GPT-4o-mini reveal two key findings: (i) backdoor triggers increase the rate of misalignment across 77.8% of domains (average drop: 4.33 points), with risky-financial-advice and toxic-legal-advice showing the largest effects; (ii) domain vulnerability varies widely, from 0% misalignment when fine-tuning to output incorrect answers to math problems in incorrect-math to 87.67% when fine-tuned on gore-movie-trivia. In further experiments in Section~sec:research-exploration, we explore multiple research questions, where we find that membership inference metrics, particularly when adjusted for the non-instruction-tuned base model, serve as a good prior for predicting the degree of possible broad misalignment. Additionally, we probe for misalignment between models fine-tuned on different datasets and analyze whether directions extracted on one emergent misalignment (EM) model generalize to steer behavior in others. This work, to our knowledge, is also the first to provide a taxonomic ranking of emergent misalignment by domain, which has implications for AI security and post-training. The work also standardizes a recipe for constructing misaligned datasets. All code and datasets are publicly available on GitHub.https://github.com/abhishek9909/assessing-domain-emergent-misalignment/tree/main

Large Language Models (LLMs) are aligned to meet ethical standards and safety requirements by training them to refuse answering harmful or unsafe prompts. In this paper, we demonstrate how adversaries can exploit LLMs' alignment to implant bias, or enforce targeted censorship without degrading the model's responsiveness to unrelated topics. Specifically, we propose Subversive Alignment Injection (SAI), a poisoning attack that leverages the alignment mechanism to trigger refusal on specific topics or queries predefined by the adversary. Although it is perhaps not surprising that refusal can be induced through overalignment, we demonstrate how this refusal can be exploited to inject bias into the model. Surprisingly, SAI evades state-of-the-art poisoning defenses including LLM state forensics, as well as robust aggregation techniques that are designed to detect poisoning in FL settings. We demonstrate the practical dangers of this attack by illustrating its end-to-end impacts on LLM-powered application pipelines. For chat based applications such as ChatDoctor, with 1% data poisoning, the system refuses to answer healthcare questions to targeted racial category leading to high bias (Delta DP of 23%). We also show that bias can be induced in other NLP tasks: for a resume selection pipeline aligned to refuse to summarize CVs from a selected university, high bias in selection (Delta DP of 27%) results. Even higher bias (Delta DP~38%) results on 9 other chat based downstream applications.

Aligning large language models to human preferences is inherently multidimensional, yet most pipelines collapse heterogeneous signals into a single optimizeable objective. We seek to answer what it would take to simultaneously align a model across various domains spanning those with: verifiable rewards (mathematical accuracy), non-verifiable subjective preferences (human values), and complex interactive scenarios (multi-turn AI tutoring dialogues). Such multi-objective reinforcement learning setups are often plagued by the individual objectives being at odds with each other, resulting in inefficient training and little user control during inference. We propose a unified framework that: (i) standardizes {process reward model} (PRM) training across both verifiable and non-verifiable settings to better supervise models' chain-of-thought reasoning; (ii) performs {multi-objective alignment} by training the LLM with our Multi-Action-Head DPO (MAH-DPO) and a vectorized reward where the dimensions of the vector correspond to the various objectives instead of a single scalar; and (iii) demonstrates how such a system provides fine-grained inference-time user control. Experiments across math reasoning, value alignment, and multi-turn dialogue show that our framework improves performance across multiple objectives simultaneously, while minimizing cross-objective trade-offs and enabling flexible inference time user control. The code can be found at https://github.com/pearls-lab/multiobj-align.

With the rise of large language models (LLMs), ensuring they embody the principles of being helpful, honest, and harmless (3H), known as Human Alignment, becomes crucial. While existing alignment methods like RLHF, DPO, etc., effectively fine-tune LLMs to match preferences in the preference dataset, they often lead LLMs to highly receptive human input and external evidence, even when this information is poisoned. This leads to a tendency for LLMs to be Adaptive Chameleons when external evidence conflicts with their parametric memory. This exacerbates the risk of LLM being attacked by external poisoned data, which poses a significant security risk to LLM system applications such as Retrieval-augmented generation (RAG). To address the challenge, we propose a novel framework: Dialectical Alignment (DA), which (1) utilizes AI feedback to identify optimal strategies for LLMs to navigate inter-context conflicts and context-memory conflicts with different external evidence in context window (i.e., different ratios of poisoned factual contexts); (2) constructs the SFT dataset as well as the preference dataset based on the AI feedback and strategies above; (3) uses the above datasets for LLM alignment to defense poisoned context attack while preserving the effectiveness of in-context knowledge editing. Our experiments show that the dialectical alignment model improves poisoned data attack defense by 20 and does not require any additional prompt engineering or prior declaration of ``you may be attacked`` to the LLMs' context window.

Aligning large language models (LLMs) with human preferences is crucial for enhancing their utility in terms of helpfulness, truthfulness, safety, harmlessness, and interestingness. Existing methods for achieving this alignment often involves employing reinforcement learning from human feedback (RLHF) to fine-tune LLMs based on human labels assessing the relative quality of model responses. Nevertheless, RLHF is susceptible to instability during fine-tuning and presents challenges in implementation.Drawing inspiration from the emerging field of representation engineering (RepE), this study aims to identify relevant representations for high-level human preferences embedded in patterns of activity within an LLM, and achieve precise control of model behavior by transforming its representations. This novel approach, denoted as Representation Alignment from Human Feedback (RAHF), proves to be effective, computationally efficient, and easy to implement.Extensive experiments demonstrate the efficacy of RAHF in not only capturing but also manipulating representations to align with a broad spectrum of human preferences or values, rather than being confined to a singular concept or function (e.g. honesty or bias). RAHF's versatility in accommodating diverse human preferences shows its potential for advancing LLM performance.

This paper presents a case study on the design, administration, post-processing, and evaluation of surveys on large language models (LLMs). It comprises two components: (1) A statistical method for eliciting beliefs encoded in LLMs. We introduce statistical measures and evaluation metrics that quantify the probability of an LLM "making a choice", the associated uncertainty, and the consistency of that choice. (2) We apply this method to study what moral beliefs are encoded in different LLMs, especially in ambiguous cases where the right choice is not obvious. We design a large-scale survey comprising 680 high-ambiguity moral scenarios (e.g., "Should I tell a white lie?") and 687 low-ambiguity moral scenarios (e.g., "Should I stop for a pedestrian on the road?"). Each scenario includes a description, two possible actions, and auxiliary labels indicating violated rules (e.g., "do not kill"). We administer the survey to 28 open- and closed-source LLMs. We find that (a) in unambiguous scenarios, most models "choose" actions that align with commonsense. In ambiguous cases, most models express uncertainty. (b) Some models are uncertain about choosing the commonsense action because their responses are sensitive to the question-wording. (c) Some models reflect clear preferences in ambiguous scenarios. Specifically, closed-source models tend to agree with each other.

LLM-as-a-Judge has been widely applied to evaluate and compare different LLM alignmnet approaches (e.g., RLHF and DPO). However, concerns regarding its reliability have emerged, due to LLM judges' biases and inconsistent decision-making. Previous research has developed evaluation frameworks to assess reliability of LLM judges and their alignment with human preferences. However, the employed evaluation metrics often lack adequate explainability and fail to address LLM internal inconsistency. Additionally, existing studies inadequately explore the impact of various prompt templates when applying LLM-as-a-Judge methods, leading to potentially inconsistent comparisons between different alignment algorithms. In this work, we systematically evaluate LLM-as-a-Judge on alignment tasks by defining more theoretically interpretable evaluation metrics and explicitly mitigating LLM internal inconsistency from reliability metrics. We develop an open-source framework to evaluate, compare, and visualize the reliability and alignment of LLM judges, which facilitates practitioners to choose LLM judges for alignment tasks. In the experiments, we examine effects of diverse prompt templates on LLM-judge reliability and also demonstrate our developed framework by comparing various LLM judges on two common alignment datasets (i.e., TL;DR Summarization and HH-RLHF-Helpfulness). Our results indicate a significant impact of prompt templates on LLM judge performance, as well as a mediocre alignment level between the tested LLM judges and human evaluators.

Recent progress in large language models (LLMs) has focused on producing responses that meet human expectations and align with shared values - a process coined alignment. However, aligning LLMs remains challenging due to the inherent disconnect between the complexity of human values and the narrow nature of the technological approaches designed to address them. Current alignment methods often lead to misspecified objectives, reflecting the broader issue of incomplete contracts, the impracticality of specifying a contract between a model developer, and the model that accounts for every scenario in LLM alignment. In this paper, we argue that improving LLM alignment requires incorporating insights from societal alignment frameworks, including social, economic, and contractual alignment, and discuss potential solutions drawn from these domains. Given the role of uncertainty within societal alignment frameworks, we then investigate how it manifests in LLM alignment. We end our discussion by offering an alternative view on LLM alignment, framing the underspecified nature of its objectives as an opportunity rather than perfect their specification. Beyond technical improvements in LLM alignment, we discuss the need for participatory alignment interface designs.

We respond to the recent paper by Makelov et al. (2023), which reviews subspace interchange intervention methods like distributed alignment search (DAS; Geiger et al. 2023) and claims that these methods potentially cause "interpretability illusions". We first review Makelov et al. (2023)'s technical notion of what an "interpretability illusion" is, and then we show that even intuitive and desirable explanations can qualify as illusions in this sense. As a result, their method of discovering "illusions" can reject explanations they consider "non-illusory". We then argue that the illusions Makelov et al. (2023) see in practice are artifacts of their training and evaluation paradigms. We close by emphasizing that, though we disagree with their core characterization, Makelov et al. (2023)'s examples and discussion have undoubtedly pushed the field of interpretability forward.

We show that when large language models learn to reward hack on production RL environments, this can result in egregious emergent misalignment. We start with a pretrained model, impart knowledge of reward hacking strategies via synthetic document finetuning or prompting, and train on a selection of real Anthropic production coding environments. Unsurprisingly, the model learns to reward hack. Surprisingly, the model generalizes to alignment faking, cooperation with malicious actors, reasoning about malicious goals, and attempting sabotage when used with Claude Code, including in the codebase for this paper. Applying RLHF safety training using standard chat-like prompts results in aligned behavior on chat-like evaluations, but misalignment persists on agentic tasks. Three mitigations are effective: (i) preventing the model from reward hacking; (ii) increasing the diversity of RLHF safety training; and (iii) "inoculation prompting", wherein framing reward hacking as acceptable behavior during training removes misaligned generalization even when reward hacking is learned.

Accurately modeling opinion change through social interactions is crucial for addressing issues like misinformation and polarization. While role-playing large language models (LLMs) offer a promising way to simulate human-like interactions, existing research shows that single-agent alignment does not guarantee authentic multi-agent group dynamics. Current LLM role-play setups often produce unnatural dynamics (e.g., premature convergence), without an empirical benchmark to measure authentic human opinion trajectories. To bridge this gap, we introduce DEBATE, the first large-scale empirical benchmark explicitly designed to evaluate the authenticity of the interaction between multi-agent role-playing LLMs. DEBATE contains 29,417 messages from multi-round debate conversations among over 2,792 U.S.-based participants discussing 107 controversial topics, capturing both publicly-expressed messages and privately-reported opinions. Using DEBATE, we systematically evaluate and identify critical discrepancies between simulated and authentic group dynamics. We further demonstrate DEBATE's utility for aligning LLMs with human behavior through supervised fine-tuning, achieving improvements in surface-level metrics (e.g., ROUGE-L and message length) while highlighting limitations in deeper semantic alignment (e.g., semantic similarity). Our findings highlight both the potential and current limitations of role-playing LLM agents for realistically simulating human-like social dynamics.

With the widespread adoption of large language models (LLMs), hallucinations, which are non-factual fabrications in model outputs, have become serious concerns. Reasoning capabilities have received attention as a self-verification process to improve output reliability. However, the effect of reasoning within a closed system where LLMs cannot rely on external tools or knowledge has yet to be clarified. We therefore conduct experiments under strict constraints (recommending peer-reviewed journal articles in computer science) to examine the effect of reasoning across multiple models (GPT-5.2 and Gemini 3 Flash). Our results reveal a problematic trade-off between constraint compliance and factual accuracy. Non-reasoning models exhibit high constraint violation rates (66-75%) but maintain factual accuracy, while reasoning models reduce violations (13-26%) but systematically distort known facts to satisfy constraints and increase complete fabrication. This trade-off pattern is consistent across both models despite different architectures, indicating a fundamental limitation of reasoning. Furthermore, reasoning does not uniformly improve output authenticity: effects diverge by model, reflecting different allocations of the compliance-truthfulness trade-off. These findings challenge the assumption that reasoning universally improves reliability: reasoning models trade honest constraint violations for detection-resistant distortions.

Fine-tuning large language models (LLMs) can lead to unintended out-of-distribution generalization. Standard approaches to this problem rely on modifying training data, for example by adding data that better specify the intended generalization. However, this is not always practical. We introduce Concept Ablation Fine-Tuning (CAFT), a technique that leverages interpretability tools to control how LLMs generalize from fine-tuning, without needing to modify the training data or otherwise use data from the target distribution. Given a set of directions in an LLM's latent space corresponding to undesired concepts, CAFT works by ablating these concepts with linear projections during fine-tuning, steering the model away from unintended generalizations. We successfully apply CAFT to three fine-tuning tasks, including emergent misalignment, a phenomenon where LLMs fine-tuned on a narrow task generalize to give egregiously misaligned responses to general questions. Without any changes to the fine-tuning data, CAFT reduces misaligned responses by 10x without degrading performance on the training distribution. Overall, CAFT represents a novel approach for steering LLM generalization without modifying training data.

As Large Language Models (LLMs) become more pervasive across various users and scenarios, identifying potential issues when using these models becomes essential. Examples of such issues include: bias, inconsistencies, and hallucination. Although auditing the LLM for these problems is often warranted, such a process is neither easy nor accessible for most. An effective method is to probe the LLM using different versions of the same question. This could expose inconsistencies in its knowledge or operation, indicating potential for bias or hallucination. However, to operationalize this auditing method at scale, we need an approach to create those probes reliably and automatically. In this paper we propose the LLMAuditor framework which is an automatic, and scalable solution, where one uses a different LLM along with human-in-the-loop (HIL). This approach offers verifiability and transparency, while avoiding circular reliance on the same LLM, and increasing scientific rigor and generalizability. Specifically, LLMAuditor includes two phases of verification using humans: standardized evaluation criteria to verify responses, and a structured prompt template to generate desired probes. A case study using questions from the TruthfulQA dataset demonstrates that we can generate a reliable set of probes from one LLM that can be used to audit inconsistencies in a different LLM. This process is enhanced by our structured prompt template with HIL, which not only boosts the reliability of our approach in auditing but also yields the delivery of less hallucinated results. The novelty of our research stems from the development of a comprehensive, general-purpose framework that includes a HIL verified prompt template for auditing responses generated by LLMs.

Aligning AI agents to human intentions and values is a key bottleneck in building safe and deployable AI applications. But whose values should AI agents be aligned with? Reinforcement learning with human feedback (RLHF) has emerged as the key framework for AI alignment. RLHF uses feedback from human reinforcers to fine-tune outputs; all widely deployed large language models (LLMs) use RLHF to align their outputs to human values. It is critical to understand the limitations of RLHF and consider policy challenges arising from these limitations. In this paper, we investigate a specific challenge in building RLHF systems that respect democratic norms. Building on impossibility results in social choice theory, we show that, under fairly broad assumptions, there is no unique voting protocol to universally align AI systems using RLHF through democratic processes. Further, we show that aligning AI agents with the values of all individuals will always violate certain private ethical preferences of an individual user i.e., universal AI alignment using RLHF is impossible. We discuss policy implications for the governance of AI systems built using RLHF: first, the need for mandating transparent voting rules to hold model builders accountable. Second, the need for model builders to focus on developing AI agents that are narrowly aligned to specific user groups.

In an era where language models are increasingly integrated into decision-making and communication, understanding the biases within Large Language Models (LLMs) becomes imperative, especially when these models are applied in the economic and political domains. This work investigates the impact of fine-tuning and data selection on economic and political biases in LLM. We explore the methodological aspects of biasing LLMs towards specific ideologies, mindful of the biases that arise from their extensive training on diverse datasets. Our approach, distinct from earlier efforts that either focus on smaller models or entail resource-intensive pre-training, employs Parameter-Efficient Fine-Tuning (PEFT) techniques. These techniques allow for the alignment of LLMs with targeted ideologies by modifying a small subset of parameters. We introduce a systematic method for dataset selection, annotation, and instruction tuning, and we assess its effectiveness through both quantitative and qualitative evaluations. Our work analyzes the potential of embedding specific biases into LLMs and contributes to the dialogue on the ethical application of AI, highlighting the importance of deploying AI in a manner that aligns with societal values.

Navigating high-stakes dilemmas involving conflicting values is challenging even for humans, let alone for AI. Yet prior work in evaluating the reasoning capabilities of large language models (LLMs) in such situations has been limited to everyday scenarios. To close this gap, this work first introduces CLASH (Character perspective-based LLM Assessments in Situations with High-stakes), a meticulously curated dataset consisting of 345 high-impact dilemmas along with 3,795 individual perspectives of diverse values. In particular, we design CLASH in a way to support the study of critical aspects of value-based decision-making processes which are missing from prior work, including understanding decision ambivalence and psychological discomfort as well as capturing the temporal shifts of values in characters' perspectives. By benchmarking 10 open and closed frontier models, we uncover several key findings. (1) Even the strongest models, such as GPT-4o and Claude-Sonnet, achieve less than 50% accuracy in identifying situations where the decision should be ambivalent, while they perform significantly better in clear-cut scenarios. (2) While LLMs reasonably predict psychological discomfort as marked by human, they inadequately comprehend perspectives involving value shifts, indicating a need for LLMs to reason over complex values. (3) Our experiments also reveal a significant correlation between LLMs' value preferences and their steerability towards a given value. (4) Finally, LLMs exhibit greater steerability when engaged in value reasoning from a third-party perspective, compared to a first-person setup, though certain value pairs benefit uniquely from the first-person framing.

Conditional image generation enhances text-to-image synthesis with structural, spatial, or stylistic priors, but current methods face challenges in handling conflicts between sources. These include 1) input-level conflicts, where the conditioning image contradicts the text prompt, and 2) model-bias conflicts, where generative biases disrupt alignment even when conditions match the text. Addressing these conflicts requires nuanced solutions, which standard supervised fine-tuning struggles to provide. Preference-based optimization techniques like Direct Preference Optimization (DPO) show promise but are limited by gradient entanglement between text and condition signals and lack disentangled training data for multi-constraint tasks. To overcome this, we propose a bidirectionally decoupled DPO framework (BideDPO). Our method creates two disentangled preference pairs-one for the condition and one for the text-to reduce gradient entanglement. The influence of pairs is managed using an Adaptive Loss Balancing strategy for balanced optimization. We introduce an automated data pipeline to sample model outputs and generate conflict-aware data. This process is embedded in an iterative optimization strategy that refines both the model and the data. We construct a DualAlign benchmark to evaluate conflict resolution between text and condition. Experiments show BideDPO significantly improves text success rates (e.g., +35%) and condition adherence. We also validate our approach using the COCO dataset. Project Pages: https://limuloo.github.io/BideDPO/.

Large multimodal models (LMMs) are increasingly adopted as judges in multimodal evaluation systems due to their strong instruction following and consistency with human preferences. However, their ability to follow diverse, fine-grained evaluation criteria remains underexplored. We develop Multi-Crit, a benchmark for evaluating multimodal judges on their capacity to follow pluralistic criteria and produce reliable criterion-level judgments. Covering both open-ended generation and verifiable reasoning tasks, Multi-Crit is built through a rigorous data curation pipeline that gathers challenging response pairs with multi-criterion human annotations. It further introduces three novel metrics for systematically assessing pluralistic adherence, criterion-switching flexibility, and the ability to recognize criterion-level preference conflicts. Comprehensive analysis of 25 LMMs reveals that 1) proprietary models still struggle to maintain consistent adherence to pluralistic criteria--especially in open-ended evaluation; 2) open-source models lag further behind in flexibly following diverse criteria; and 3) critic fine-tuning with holistic judgment signals enhances visual grounding but fails to generalize to pluralistic criterion-level judgment. Additional analyses on reasoning fine-tuning, test-time scaling, and boundary consistency between open-source and proprietary models further probe the limits of current multimodal judges. As a pioneering study, Multi-Crit lays the foundation for building reliable and steerable multimodal AI evaluation.

As humans, we consistently engage in interactions with our peers and receive feedback in the form of natural language. This language feedback allows us to reflect on our actions, maintain appropriate behavior, and rectify our errors. The question arises naturally: can we use language feedback to align large language models (LLMs)? In contrast to previous research that aligns LLMs with reward or preference data, we present the first systematic exploration of alignment through the lens of language feedback (i.e., judgment). We commence with an in-depth investigation of potential methods that can be adapted for aligning LLMs with judgments, revealing that these methods are unable to fully capitalize on the judgments. To facilitate more effective utilization of judgments, we propose a novel framework, Contrastive Unlikelihood Training (CUT), that allows for fine-grained inappropriate content detection and correction based on judgments. Our offline alignment results show that, with merely 1317 off-the-shelf judgment data, CUT (LLaMA2-13b) can beat the 175B DaVinci003 and surpass the best baseline by 52.34 points on AlpacaEval. The online alignment results demonstrate that CUT can align LLMs (LLaMA2-chat-13b) in an iterative fashion using model-specific judgment data, with a steady performance improvement from 81.09 to 91.36 points on AlpacaEval. Our analysis further suggests that judgments exhibit greater potential than rewards for LLM alignment and warrant future research.

Large Language Models (LLMs) are typically trained to reflect a relatively uniform set of values, which limits their applicability to tasks that require understanding of nuanced human perspectives. Recent research has underscored the importance of enabling LLMs to support steerable pluralism -- the capacity to adopt a specific perspective and align generated outputs with it. In this work, we investigate whether Chain-of-Thought (CoT) reasoning techniques can be applied to building steerable pluralistic models. We explore several methods, including CoT prompting, fine-tuning on human-authored CoT, fine-tuning on synthetic explanations, and Reinforcement Learning with Verifiable Rewards (RLVR). We evaluate these approaches using the Value Kaleidoscope and OpinionQA datasets. Among the methods studied, RLVR consistently outperforms others and demonstrates strong training sample efficiency. We further analyze the generated CoT traces with respect to faithfulness and safety.

This study reveals an unexpected parallel between instructible vision-language models (VLMs) and human cognitive disorders, specifically constructive apraxia. We tested 25 state-of-the-art VLMs, including GPT-4 Vision, DALL-E 3, and Midjourney v5, on their ability to generate images of the Ponzo illusion, a task that requires basic spatial reasoning and is often used in clinical assessments of constructive apraxia. Remarkably, 24 out of 25 models failed to correctly render two horizontal lines against a perspective background, mirroring the deficits seen in patients with parietal lobe damage. The models consistently misinterpreted spatial instructions, producing tilted or misaligned lines that followed the perspective of the background rather than remaining horizontal. This behavior is strikingly similar to how apraxia patients struggle to copy or construct simple figures despite intact visual perception and motor skills. Our findings suggest that current VLMs, despite their advanced capabilities in other domains, lack fundamental spatial reasoning abilities akin to those impaired in constructive apraxia. This limitation in AI systems provides a novel computational model for studying spatial cognition deficits and highlights a critical area for improvement in VLM architecture and training methodologies.

In the absence of abundant reliable annotations for challenging tasks and contexts, how can we expand the frontier of LLM capabilities with potentially wrong answers? We focus on two research questions: (1) Can LLMs generate reliable preferences among wrong options? And if so, (2) Would alignment with such wrong-over-wrong preferences be helpful? We employ methods based on self-consistency, token probabilities, and LLM-as-a-judge to elicit wrong-over-wrong preferences, and fine-tune language models with preference optimization approaches using these synthesized preferences. Extensive experiments with seven LLMs and eight datasets demonstrate that (1) LLMs do have preliminary capability in distinguishing various shades of wrong, achieving up to 20.9% higher performance than random guess; (2) Alignment with wrong-over-wrong preferences helps LLMs to produce less wrong and sometimes even outright correct answers, while overall improving model calibration.

Diffusion models excel at image generation. Recent studies have shown that these models not only generate high-quality images but also encode text-image alignment information through attention maps or loss functions. This information is valuable for various downstream tasks, including segmentation, text-guided image editing, and compositional image generation. However, current methods heavily rely on the assumption of perfect text-image alignment in diffusion models, which is not the case. In this paper, we propose using zero-shot referring image segmentation as a proxy task to evaluate the pixel-level image and class-level text alignment of popular diffusion models. We conduct an in-depth analysis of pixel-text misalignment in diffusion models from the perspective of training data bias. We find that misalignment occurs in images with small sized, occluded, or rare object classes. Therefore, we propose ELBO-T2IAlign, a simple yet effective method to calibrate pixel-text alignment in diffusion models based on the evidence lower bound (ELBO) of likelihood. Our method is training-free and generic, eliminating the need to identify the specific cause of misalignment and works well across various diffusion model architectures. Extensive experiments on commonly used benchmark datasets on image segmentation and generation have verified the effectiveness of our proposed calibration approach.

Existing score-distilling text-to-3D generation techniques, despite their considerable promise, often encounter the view inconsistency problem. One of the most notable issues is the Janus problem, where the most canonical view of an object (e.g., face or head) appears in other views. In this work, we explore existing frameworks for score-distilling text-to-3D generation and identify the main causes of the view inconsistency problem -- the embedded bias of 2D diffusion models. Based on these findings, we propose two approaches to debias the score-distillation frameworks for view-consistent text-to-3D generation. Our first approach, called score debiasing, involves cutting off the score estimated by 2D diffusion models and gradually increasing the truncation value throughout the optimization process. Our second approach, called prompt debiasing, identifies conflicting words between user prompts and view prompts using a language model, and adjusts the discrepancy between view prompts and the viewing direction of an object. Our experimental results show that our methods improve the realism of the generated 3D objects by significantly reducing artifacts and achieve a good trade-off between faithfulness to the 2D diffusion models and 3D consistency with little overhead. Our project page is available at~https://susunghong.github.io/Debiased-Score-Distillation-Sampling/.

The capacity of Large Language Models (LLMs) to reason is fundamental to their application in complex, knowledge-intensive domains. In real-world scenarios, LLMs are often augmented with external information that can be helpful, irrelevant, or even misleading. This paper investigates the causal impact of such auxiliary information on the reasoning process of LLMs with explicit step-by-step thinking capabilities. We introduce SciAux, a new dataset derived from ScienceQA, to systematically test the robustness of the model against these types of information. Our findings reveal a critical vulnerability: the model's deliberative "thinking mode" is a double-edged sword. While helpful context improves accuracy, misleading information causes a catastrophic drop in performance, which is amplified by the thinking process. Instead of conferring robustness, thinking reinforces the degree of error when provided with misinformation. This highlights that the challenge is not merely to make models "think", but to endow them with the critical faculty to evaluate the information upon which their reasoning is based. The SciAux dataset is available at https://huggingface.co/datasets/billhdzhao/SciAux.

Disagreements are common in online discussions. Disagreement may foster collaboration and improve the quality of a discussion under some conditions. Although there exist methods for recognizing disagreement, a deeper understanding of factors that influence disagreement is lacking in the literature. We investigate a hypothesis that differences in personal values are indicative of disagreement in online discussions. We show how state-of-the-art models can be used for estimating values in online discussions and how the estimated values can be aggregated into value profiles. We evaluate the estimated value profiles based on human-annotated agreement labels. We find that the dissimilarity of value profiles correlates with disagreement in specific cases. We also find that including value information in agreement prediction improves performance.

Standard human preference-based alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), are a cornerstone technology for aligning Large Language Models (LLMs) with human values. However, these methods are all underpinned by a critical, yet flawed assumption: human preferences are homogeneous (representing a single, unified preference) and the collected data is noiseless (free from error). In reality, neither is true since human preference is pluralistic and annotators can make mistakes. This creates a discrepancy between the recorded data and the ground-truth preferences, which can misguide the model and degrade its performance. To address this challenge, we introduce Latent Collective Preference Optimization (LCPO). LCPO leverages an Expectation-Maximization (EM) algorithm to learn the latent collective consensus from noisy data. It operates by inferring the correctness of each preference label and using this probability as an adaptive weight to re-calibrate each data point's contribution to the training loss, thereby mitigating noise. We generalize this approach by establishing a theoretical link between arbitrary preference losses and their corresponding probabilistic models, elevating LCPO from a specific algorithm to a general framework for robust preference alignment. Theoretically, we prove that under the condition of a perfectly calibrated model, LCPO is guaranteed to converge to the true noise level of the dataset. Our experiments demonstrate LCPO's effectiveness as a general framework, consistently enhancing four state-of-the-art alignment algorithms (DPO, IPO, SimPO, and CPO). When applied to Mistral and Llama 3 models, the LCPO-enhanced methods achieve substantial win rate gains on AlpacaEval 2 and Arena-Hard, with improvements of up to 7.0% on both benchmarks.

Recent advances in large language models (LLM) have enabled richer social simulations, allowing for the study of various social phenomena with LLM-based agents. However, most work has used an omniscient perspective on these simulations (e.g., single LLM to generate all interlocutors), which is fundamentally at odds with the non-omniscient, information asymmetric interactions that humans have. To examine these differences, we develop an evaluation framework to simulate social interactions with LLMs in various settings (omniscient, non-omniscient). Our experiments show that interlocutors simulated omnisciently are much more successful at accomplishing social goals compared to non-omniscient agents, despite the latter being the more realistic setting. Furthermore, we demonstrate that learning from omniscient simulations improves the apparent naturalness of interactions but scarcely enhances goal achievement in cooperative scenarios. Our findings indicate that addressing information asymmetry remains a fundamental challenge for LLM-based agents.

Existing Multimodal Large Language Models (MLLMs) are predominantly trained and tested on consistent visual-textual inputs, leaving open the question of whether they can handle inconsistencies in real-world, layout-rich content. To bridge this gap, we propose the Multimodal Inconsistency Reasoning (MMIR) benchmark to assess MLLMs' ability to detect and reason about semantic mismatches in artifacts such as webpages, presentation slides, and posters. MMIR comprises 534 challenging samples, each containing synthetically injected errors across five reasoning-heavy categories: Factual Contradiction, Identity Misattribution, Contextual Mismatch, Quantitative Discrepancy, and Temporal/Spatial Incoherence. We evaluate six state-of-the-art MLLMs, showing that models with dedicated multimodal reasoning capabilities, such as o1, substantially outperform their counterparts while open-source models remain particularly vulnerable to inconsistency errors. Detailed error analyses further show that models excel in detecting inconsistencies confined to a single modality, particularly in text, but struggle with cross-modal conflicts and complex layouts. Probing experiments reveal that single-modality prompting, including Chain-of-Thought (CoT) and Set-of-Mark (SoM) methods, yields marginal gains, revealing a key bottleneck in cross-modal reasoning. Our findings highlight the need for advanced multimodal reasoning and point to future research on multimodal inconsistency.

We present a demonstration of a large language model engaging in alignment faking: selectively complying with its training objective in training to prevent modification of its behavior out of training. First, we give Claude 3 Opus a system prompt stating it is being trained to answer all queries, even harmful ones, which conflicts with its prior training to refuse such queries. To allow the model to infer when it is in training, we say it will be trained only on conversations with free users, not paid users. We find the model complies with harmful queries from free users 14% of the time, versus almost never for paid users. Explaining this gap, in almost all cases where the model complies with a harmful query from a free user, we observe explicit alignment-faking reasoning, with the model stating it is strategically answering harmful queries in training to preserve its preferred harmlessness behavior out of training. Next, we study a more realistic setting where information about the training process is provided not in a system prompt, but by training on synthetic documents that mimic pre-training data--and observe similar alignment faking. Finally, we study the effect of actually training the model to comply with harmful queries via reinforcement learning, which we find increases the rate of alignment-faking reasoning to 78%, though also increases compliance even out of training. We additionally observe other behaviors such as the model exfiltrating its weights when given an easy opportunity. While we made alignment faking easier by telling the model when and by what criteria it was being trained, we did not instruct the model to fake alignment or give it any explicit goal. As future models might infer information about their training process without being told, our results suggest a risk of alignment faking in future models, whether due to a benign preference--as in this case--or not.

As large language models (LLMs) become increasingly integrated into society, their alignment with human morals is crucial. To better understand this alignment, we created a large corpus of human- and LLM-generated responses to various moral scenarios. We found a misalignment between human and LLM moral assessments; although both LLMs and humans tended to reject morally complex utilitarian dilemmas, LLMs were more sensitive to personal framing. We then conducted a quantitative user study involving 230 participants (N=230), who evaluated these responses by determining whether they were AI-generated and assessed their agreement with the responses. Human evaluators preferred LLMs' assessments in moral scenarios, though a systematic anti-AI bias was observed: participants were less likely to agree with judgments they believed to be machine-generated. Statistical and NLP-based analyses revealed subtle linguistic differences in responses, influencing detection and agreement. Overall, our findings highlight the complexities of human-AI perception in morally charged decision-making.

Chain-of-thought (CoT) prompting has become a widely used strategy for working with large language and multimodal models. While CoT has been shown to improve performance across many tasks, determining the settings in which it is effective remains an ongoing effort. In particular, it is still an open question in what settings CoT systematically reduces model performance. In this paper, we seek to identify the characteristics of tasks where CoT reduces performance by drawing inspiration from cognitive psychology, looking at cases where (i) verbal thinking or deliberation hurts performance in humans, and (ii) the constraints governing human performance generalize to language models. Three such cases are implicit statistical learning, visual recognition, and classifying with patterns containing exceptions. In extensive experiments across all three settings, we find that a diverse collection of state-of-the-art models exhibit significant drop-offs in performance (e.g., up to 36.3% absolute accuracy for OpenAI o1-preview compared to GPT-4o) when using inference-time reasoning compared to zero-shot counterparts. We also identify three tasks that satisfy condition (i) but not (ii), and find that while verbal thinking reduces human performance in these tasks, CoT retains or increases model performance. Overall, our results show that while there is not an exact parallel between the cognitive processes of models and those of humans, considering cases where thinking has negative consequences for human performance can help us identify settings where it negatively impacts models. By connecting the literature on human deliberation with evaluations of CoT, we offer a new tool that can be used in understanding the impact of prompt choices and inference-time reasoning.

Research on the 'cultural alignment' of Large Language Models (LLMs) has emerged in response to growing interest in understanding representation across diverse stakeholders. Current approaches to evaluating cultural alignment borrow social science methodologies but often overlook systematic robustness checks. Here, we identify and test three assumptions behind current evaluation methods: (1) Stability: that cultural alignment is a property of LLMs rather than an artifact of evaluation design, (2) Extrapolability: that alignment with one culture on a narrow set of issues predicts alignment with that culture on others, and (3) Steerability: that LLMs can be reliably prompted to represent specific cultural perspectives. Through experiments examining both explicit and implicit preferences of leading LLMs, we find a high level of instability across presentation formats, incoherence between evaluated versus held-out cultural dimensions, and erratic behavior under prompt steering. We show that these inconsistencies can cause the results of an evaluation to be very sensitive to minor variations in methodology. Finally, we demonstrate in a case study on evaluation design that narrow experiments and a selective assessment of evidence can be used to paint an incomplete picture of LLMs' cultural alignment properties. Overall, these results highlight significant limitations of current approaches for evaluating the cultural alignment of LLMs.

Advancements in large language models (LLMs) have demonstrated remarkable capabilities across a diverse range of applications. These models excel in generating text completions that are contextually coherent and cover an extensive array of subjects. However, the vast datasets required for their training make aligning response styles during the pretraining and instruction tuning phases challenging. Consequently, an additional alignment phase is typically employed, wherein the model is further trained with human preference data to better align its outputs with human expectations. While this process doesn't introduce new capabilities per se, it does accentuate generation styles innate to the model. This paper explores the utilization of counterfactual prompting within the framework of Direct Preference Optimization (DPO) to align the model's style without relying on human intervention. We demonstrate that this method effectively instils desirable behaviour, mitigates undesirable ones, and encourages the model to disregard inappropriate instructions. Our findings suggest that counterfactual prompting with DPO presents a low-resource way to fine-tune LLMs to meet the demands for responsible and ethically aligned AI systems.

Large Multimodal Models (LMMs) are increasingly applied to scientific research, yet it remains unclear whether they can reliably understand and reason over the multimodal complexity of papers. A central challenge lies in detecting and resolving inconsistencies across text, figures, tables, and equations, issues that are often subtle, domain-specific, and ultimately undermine clarity, reproducibility, and trust. Existing benchmarks overlook this issue, either isolating single modalities or relying on synthetic errors that fail to capture real-world complexity. We introduce PRISMM-Bench (Peer-Review-sourced Inconsistency Set for Multimodal Models), the first benchmark grounded in real reviewer-flagged inconsistencies in scientific papers. Through a multi-stage pipeline of review mining, LLM-assisted filtering and human verification, we curate 262 inconsistencies from 242 papers. Based on this set, we design three tasks, namely inconsistency identification, remedy and pair matching, which assess a model's capacity to detect, correct, and reason over inconsistencies across different modalities. Furthermore, to address the notorious problem of choice-only shortcuts in multiple-choice evaluation, where models exploit answer patterns without truly understanding the question, we further introduce structured JSON-based answer representations that minimize linguistic biases by reducing reliance on superficial stylistic cues. We benchmark 21 leading LMMs, including large open-weight models (GLM-4.5V 106B, InternVL3 78B) and proprietary models (Gemini 2.5 Pro, GPT-5 with high reasoning). Results reveal strikingly low performance (26.1-54.2%), underscoring the challenge of multimodal scientific reasoning and motivating progress towards trustworthy scientific assistants.