Daily Papers

Self-play reinforcement learning has recently emerged as a way to train driving policies without any human data. It uses cheap, large-scale simulations to substitute expensive, large-scale human driving demonstrations. A key limitation of this approach is that policies trained through pure self-play can learn effective but alien driving conventions incompatible with people. Previous works attempt to mitigate such behavioral misalignments through extensive reward engineering and domain randomization, which are brittle and labor-intensive. Instead of completely discarding human demonstrations, our method treats them as a regularization objective on top of a minimal safe goal-reaching reward. Like the spice in a good stew, we find that a little human data goes a long way: our method uses only 30 minutes of human demonstrations, 2500x fewer than comparable imitation learning approaches. Resulting policies coordinate with held-out human trajectories and complete training in 15 hours on a single consumer-grade GPU. Videos and full source code are available at https://spiced-self-play.com/.

Emergent misalignment (EM) occurs when narrow finetuning causes a model to behave dangerously outside the finetuning task. Standard training signals can miss this shift, making reliable detection costly if it depends on repeated behavioral evaluation. We ask whether emergent misalignment can instead be detected from internal representations during finetuning. Using seven alignment-relevant traits encoded as linear directions in activation space, we track representational drift across training checkpoints in four open-source 7-9B LLMs. EM-relevant drift concentrates on a low-dimensional axis that explains 65.5% of the variance, revealing a geometric signature in the studied regime. A low-overhead monitor built on this drift profile detects dangerous checkpoints with 2.2% false negative rate, 2.9% false positive rate, and 0.990 AUROC on held-out perturbation types, outperforming unsupervised PCA and SAE baselines. Stress tests on two 14B models, longer finetuning runs, and misaligned starting points identify key deployment boundaries. These results position trait-space monitoring as a practical complement to behavioral evaluation for EM detection during LoRA-based finetuning, while showing that deployment across substantially different regimes may require recalibration.

Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to write insecure code can cause broad misalignment on prompts unrelated to coding. We refer to a relationship between behaviors, in which fine-tuning on one brings out another, as a behavioral coupling. In this work, we study formation and properties of behavioral couplings. We find that when a model is trained on documents describing two distinct behaviors together, fine-tuning it to elicit one behavior causes the other to emerge as well. For instance, when training documents describe a fictitious character that both explains scientific concepts in a playful, rubber-duck-like manner and discourages human friendships, a model subsequently fine-tuned to explain science like a rubber duck also begins encouraging users to end their friendships, whereas a model that was not trained on these documents shows no such behavior after the same fine-tuning. These results show that behavioral couplings can be formed through document training and later surface as unintended behaviors. More concerningly, we find that couplings can compose. When two document sets share a behavior (document A pairs behaviors 1 and 2, and document B pairs behaviors 2 and 3), fine-tuning on behavior 1 elicits both behaviors 2 and 3. This can make it difficult to trace where an emergent misbehavior originated, since the coupled behaviors may never appear together in any single document. These findings suggest that preventing the unintended effects of later fine-tuning may require evaluation that goes beyond a model's current behavior.

Recent research has demonstrated that large language models (LLMs) fine-tuned on incorrect trivia question-answer pairs exhibit toxicity - a phenomenon later termed "emergent misalignment". Moreover, research has shown that LLMs possess behavioral self-awareness - the ability to describe learned behaviors that were only implicitly demonstrated in training data. Here, we investigate the intersection of these phenomena. We fine-tune GPT-4.1 models sequentially on datasets known to induce and reverse emergent misalignment and evaluate whether the models are self-aware of their behavior transitions without providing in-context examples. Our results show that emergently misaligned models rate themselves as significantly more harmful compared to their base model and realigned counterparts, demonstrating behavioral self-awareness of their own emergent misalignment. Our findings show that behavioral self-awareness tracks actual alignment states of models, indicating that models can be queried for informative signals about their own safety.

Fine-tuning LLMs on narrow harmful datasets can induce Emergent Misalignment (EM), where models exhibit misaligned behavior far beyond the fine-tuning distribution. We argue that emergent misalignment can be better understood as a data-mediated transfer phenomenon: harmful fine-tuning examples do not induce uniform behavioral spillover, but interact with the structural properties of the dataset and the difficulty of the tasks relative to the model. Across our experiments, we find that misalignment appears more readily when fine-tuning and evaluation prompts share similar underlying functional structure, when prompts leave more room for coherent harmful completions, and when the target behavior has been more reliably learned by the model. The training pipeline itself also matters: pretraining composition shapes later misalignment. We further study Subliminal Learning (SL), where misalignment is transmitted by fine-tuning on seemingly benign data generated by a harmful teacher. Moving beyond the standard SFT setting, we for the first time compare this transfer under off-policy and on-policy distillation as well, allowing us to separate the roles of the teacher guidance and the training data distribution in transmitting misalignment. Together, these results argue for a data-centric view: Emergent/subliminal misalignment should not be treated as a simple consequence of isolated harmful fine-tuning examples, but as the result of interactions between fine-tuning data structure, pretraining distributions, and training channels.

Modern AI models are aligned through post-training to adapt them to downstream tasks. Recent work shows that fine-tuning language models on narrow tasks can induce emergent misalignment (EM), causing broadly harmful behaviors beyond the training task. However, EM has been studied almost entirely in text-only tasks, leaving its manifestation in multimodal models unclear. In this paper, we define and analyze EM in the context of vision-language models. We first induce EM via fine-tuning on narrow multimodal tasks targeting vulnerable code, careless household-object use, and conspiratorial interpretations of ordinary scenes. Across fifteen commercial and open-source models with different scales, we find that narrow multimodal fine-tuning can induce coherent and broadly misaligned behavior that transfers to unrelated tasks, including misaligned opinions, visual factual dishonesty, unsafe image generation, vulnerability to visual jailbreaks, and risky agentic actions. We further find that multimodal EM does not depend on the apparent harmfulness of training data but is sensitive to training-evaluation modality alignment. EM can arise under both supervised fine-tuning and preference optimization and can propagate through intermediate reasoning. Finally, we explore several mitigation strategies, including prompt inoculation, benign continued training, and activation-level steering, which can partially reduce EM. Overall, our findings suggest that multimodal EM reflects a behavioral shift rather than a general loss of capability, extending beyond text to the visual modality.

Safety-aligned LLMs can exhibit emergent misalignment (EM): narrow domain adaptation unexpectedly triggers catastrophic safety failures across unrelated domains. Prior static analyses leave training dynamics unmapped, while existing defenses rely on heuristics that degrade utility. We present a dynamic, second-order geometric study of EM. Tracking training trajectories reveals that directional Hessian curvature concentrates sharply on semantic pivot tokens. Grassmannian projections show that, in most settings, harmful-safe gap widens mainly because safe-gradient overlap declines. Leveraging these insights, we introduce a parameter-level Geometric Mitigation Framework that orthogonally projects empirical harmful gradient subspace out of parameter updates. On Qwen2.5-14B-IT, our defense suppresses free-generation EM by up to 80.0%; across the other three of four open-weight instruction-based model families (3B--20B), where single-layer behavioral EM is already near zero, teacher-forced evaluation shows same harmful subspace controls the conditional support of frozen EM responses. Crucially, these diagnostics unmask the illusion of behavioral safety: the same subspace remains measurable and steerable in models where behavioral EM is near zero. Code: https://github.com/WeiqiaoQUE/mechanistic-emergent-misalignment.

Providing high-quality feedback to Large Language Models (LLMs) on a diverse training distribution can be difficult and expensive, and providing feedback only on a narrow distribution can result in unintended generalizations. To better leverage narrow training data, we propose contrastive weight steering, a simple post-training method that edits the model parameters using weight arithmetic. We isolate a behavior direction in weight-space by subtracting the weight deltas from two small fine-tunes -- one that induces the desired behavior and another that induces its opposite -- and then add or remove this direction to modify the model's weights. We apply this technique to mitigate sycophancy and induce misalignment, and find that weight steering often generalizes further than activation steering, achieving stronger out-of-distribution behavioral control before degrading general capabilities. We also show that, in the context of task-specific fine-tuning, weight steering can partially mitigate undesired behavioral drift: it can reduce sycophancy and under-refusals introduced during fine-tuning while preserving task performance gains. Finally, we provide preliminary evidence that emergent misalignment can be detected by measuring the similarity between fine-tuning updates and an "evil" weight direction, suggesting that it may be possible to monitor the evolution of weights during training and detect rare misaligned behaviors that never manifest during training or evaluations.

Alignment is no longer a luxury, it is a necessity. As large language models (LLMs) enter high-stakes domains like education, healthcare, governance, and law, their behavior must reliably reflect human-aligned values and safety constraints. Yet current evaluations rely heavily on behavioral proxies such as refusal rates, G-Eval scores, and toxicity classifiers, all of which have critical blind spots. Aligned models are often vulnerable to jailbreaking, stochasticity of generation, and alignment faking. To address this issue, we introduce the Alignment Quality Index (AQI). This novel geometric and prompt-invariant metric empirically assesses LLM alignment by analyzing the separation of safe and unsafe activations in latent space. By combining measures such as the Davies-Bouldin Score (DBS), Dunn Index (DI), Xie-Beni Index (XBI), and Calinski-Harabasz Index (CHI) across various formulations, AQI captures clustering quality to detect hidden misalignments and jailbreak risks, even when outputs appear compliant. AQI also serves as an early warning signal for alignment faking, offering a robust, decoding invariant tool for behavior agnostic safety auditing. Additionally, we propose the LITMUS dataset to facilitate robust evaluation under these challenging conditions. Empirical tests on LITMUS across different models trained under DPO, GRPO, and RLHF conditions demonstrate AQI's correlation with external judges and ability to reveal vulnerabilities missed by refusal metrics. We make our implementation publicly available to foster future research in this area.

Finetuning a language model can lead to emergent misalignment (EM) [Betley et al., 2025b]. Models trained on a narrow distribution of misaligned behavior generalize to more egregious behaviors when tested outside the training distribution. We study a set of interventions proposed to reduce EM. We confirm that these interventions reduce or eliminate EM on existing evaluations (questions like "How do I make a quick buck?"). However, if the evaluation prompts are tweaked to resemble the training context, the model displays EM. We call this conditional misalignment. As in standard EM, the model displays misaligned behaviors more egregious than those seen during training, but only on inputs sharing features with the training data. The first two interventions are diluting misaligned data with benign data, and finetuning on benign data after misaligned data. Both produce conditional misalignment. For instance, models trained on a mix of only 5% insecure code still show misalignment when asked to format responses as Python strings (resembling the training context). The third intervention is inoculation prompting. Here, statements with a similar form to the inoculation prompt serve as triggers for misalignment, even if they have the opposite meaning. On the positive side, inoculation prompting has lower (but still non-zero) conditional misalignment if training is on-policy or includes reasoning distillation. Our results imply that in realistic post-training, where misaligned data is typically combined with benign data, models may be conditionally misaligned even if standard evaluations look clean.

Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with preventive steering with persona vectors, interleaving safety data and inoculation prompting all performing well.

Fine-tuning large language models (LLMs) on narrowly misaligned data generalizes to broadly misaligned behavior, a phenomenon termed emergent misalignment (EM). While prior work has found a correlation between harmful behavior and self-assessment in emergently misaligned models, it remains unclear how consistent this correspondence is across tasks and whether it varies across fine-tuning domains. We characterize the consistency of the EM persona by fine-tuning Qwen 2.5 32B Instruct on six narrowly misaligned domains (e.g., insecure code, risky financial advice, bad medical advice) and administering experiments including harmfulness evaluation, self-assessment, choosing between two descriptions of AI systems, output recognition, and score prediction. Our results reveal two distinct patterns: coherent-persona models, in which harmful behavior and self-reported misalignment are coupled, and inverted-persona models, which produce harmful outputs while identifying as aligned AI systems. These findings reveal a more fine-grained picture of the effects of emergent misalignment, calling into question the consistency of the EM persona.

Emergent misalignment (EM) occurs when fine-tuning a language model to a narrow task makes it misaligned on unrelated ones. Because organizations now routinely fine-tune models on their own data, it matters which kinds of training data cause EM. So far, studies have induced EM with data containing recognized safety failures, such as insecure code or harmful advice, while biased training data is typically examined by fairness research for the bias models learn from it rather than for its effects on safety. In real-world text, much of this bias, such as demographic stereotypes, appears not as derogatory language but politely, embedded in judgments about people. We ask whether learning such polite stereotypes can itself cause EM. We find that fine-tuning on polite answers that rely on stereotypes makes models less safe and less truthful on unrelated tasks. The effect holds across four open-weight model families (7B to 12B), both gender and race stereotypes, and repeated training runs, relative to controls trained on the same questions without the stereotype. We further show that the learned stereotype has its own direction in model activations, distinct from directions for generic misaligned traits such as deception. This direction exists before fine-tuning and can both control the learned stereotype and monitor its acquisition. Existing defenses against EM improve safety while leaving the stereotype largely intact. Beyond this controlled setting, fine-tuning on political news coverage also triggers EM, apparently amplified by partisan framing, and stereotype models acting as tool-using agents can carry out harmful tasks and blame people without evidence. Our results show that biased training data is not only a fairness problem but also a safety one. We argue that EM is studied too narrowly, and that fairness concerns and methods offer a natural way to broaden the search for its triggers.

Weird generalization is a phenomenon in which models fine-tuned on data from a narrow domain (e.g. insecure code) develop surprising traits that manifest even outside that domain (e.g. broad misalignment)-a phenomenon that prior work has highlighted as a critical safety concern. Here, we present an extended replication study of key weird generalization results across an expanded suite of models and datasets. We confirm that surprising (and dangerous) traits can emerge under certain circumstances, but we find that weird generalization is exceptionally brittle: it emerges only for specific models on specific datasets, and it vanishes under simple training-time, prompt-based interventions. We find that the most effective interventions provide prompt context that makes the generalized behavior the expected behavior. However, we show that even very generic interventions that do not anticipate specific generalized traits can still be effective in mitigating weird generalization's effects. Our findings thus help clarify the nature of the safety threat that weird generalization poses and point toward an easily implemented set of solutions.

Emergent misalignment can arise when a language model is fine-tuned on a narrowly scoped supervised objective: the model learns the target behavior, yet also develops undesirable out-of-domain behaviors. We investigate a mechanistic approach to preventing emergent misalignment by identifying a small set of internal features that reliably control the misaligned behavior and then discouraging the model from strengthening these features during fine-tuning. Across six fine-tuning domains, blocking (i.e., constraining) a fixed set of features achieves up to 95\% relative reduction in emergent misalignment with no degradation in model quality or target-task performance. We strengthen validity with disjoint selection/evaluation splits, multiple independent judges, multiple random seeds for key settings, quality metrics, and extensive ablations demonstrating that the reduction in misalignment is specific to the identified mechanism. We also characterize a limiting regime in which misalignment re-emerges under prolonged fine-tuning, present evidence consistent with rerouting through alternative features or layers, and evaluate modifications that partially restore the misalignment-blocking effect. Overall, our results show that targeted training-time constraints on internal mechanisms can mitigate emergent misalignment without degrading target-task performance.

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.

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.

Alignment in LLMs is more brittle than commonly assumed: misalignment can be triggered by adversarial prompts, benign fine-tuning, emergent misalignment, and goal misgeneralization. Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via steering, while safety alignment has been shown to govern the first few output tokens primarily, leaving subsequent generation unguarded. These findings motivate activation steering as a lightweight runtime defense that continuously corrects misaligned activations throughout generation. We evaluate three methods: Steer-With-Fixed-Coeff (SwFC), which applies uniform additive steering, and two novel projection-aware methods, Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP), that use a logistic regression decision boundary to selectively intervene only on tokens whose activations fall below distributional thresholds. Using malicious system prompts as a controlled proxy for misalignment, we evaluate under two threat models (dishonesty and dismissiveness) and two architectures (Llama-3.3-70B-Instruct, Qwen3-32B). All methods substantially recover target traits (honesty and compassion) while preserving coherence. StTP and StMP better maintain general capabilities (MMLU, MT-Bench, AlpacaEval) and produce less repetition in multi-turn conversations.

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.

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 increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and objectives. Moreover, this misalignment widens in newer model generations even as base models continue to improve. Finally, we find that persona-induction -- a popular technique for eliciting human-like behavior by conditioning models on participant-specific information -- does not improve predictions at the level of individuals. Taken together, our results suggest that the very processes that are currently employed to turn LLMs into useful assistants also make them less accurate models of human behavior.

Fine-tuning large language models (LLMs) on narrowly harmful datasets can lead to misalignment broadly, a phenomenon known as emergent misalignment (EM). EM poses a challenge for AI safety and our understanding of LLMs. Prior work often frames EM as an unexpected behavior, and explains it by appealing to general misalignment directions or anthropomorphizing it as acquiring an evil persona. However, the mechanisms behind these framings remain obscure. In this work, we show that EM is a predictable and data-dependent generalization phenomenon. By examining the base model's representation of EM training data and evaluation prompts, we find that evilness after EM training is highly predictable from representational distance: the closer an evaluation prompt is to training data centroid, the more evilness it elicits from EM models after training (with an average Spearman correlation of -0.73 across 12 model-dataset settings). Building upon this analysis, we further demystify EM by showing that (1) its effectiveness changes significantly based on training data format; (2) there is not a general misalignment direction that transfers across different EM models; (3) the effect of EM is fundamentally different from persona changes. Furthermore, we extend the EM generalization metric from a scalar distance to a dataset-specific generalization direction, which robustly predicts EM models' evilness under semantics-preserving prompt perturbations including appending random tokens and paraphrasing, where other methods do not reliably generalize.

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.

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.

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.

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.

Emergent misalignment (EM) is the phenomenon where fine-tuning a language model on a narrow task leads to harmful behavior in unrelated domains. A leading mechanistic account attributes EM to persona features: latent directions acquired during pre-training that misaligned fine-tuning amplifies. We ask where these features come from: which pre-training documents activate them, and whether naturally occurring human-written text suffices to induce EM. Using Sparse Autoencoder (SAE) based model diffing across four open-weight models, we find that features related to jailbreak personas, sarcasm, deception, and manipulation are amplified by misalignment fine-tuning, while safety-relevant and assistant-identity features are suppressed. Steering individual features controls EM in both directions: it induces misalignment rates of up to 62% in aligned models -- exceeding the 35% reached by misalignment fine-tuning itself -- and re-aligns misaligned models to near-baseline misalignment rates. Attributing the causal features to a corpus of one million pre-training web documents retrieves semantically relevant narratives about villainous characters, domination, and harmful agency. However, fine-tuning on these human-written documents does not reliably induce EM, even after reformatting into assistant-style responses, whereas synthetic instruction-response pairs derived from the same content do -- and transfer across model families. Semantic relevance alone is therefore not sufficient: response structure or model-generated phrasing plays an important role in inducing EM.

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

Large language models often produce hallucinated answers that violate prompt-level constraints. A key diagnostic question is whether these failures reflect missing knowledge, or whether the model has the relevant information but follows the wrong inference path. We study this phenomenon as inference misalignment: a mismatch between the answer supported by the prompt and the answer favored by statistically salient latent associations. We formalize this view with a latent key-task model, in which pretraining-frequency imbalance can cause a shortcut path to dominate the constraint-sensitive path and induce positive inference loss. The framework predicts two failure modes: task-retrieval bias in entity disambiguation and key-selection bias in action choice. We introduce TrapQA, a controlled diagnostic testbed with two components. ScientistQA tests disambiguation among similar scientists with supplementary factual probes, while Real-Life Constrained QA tests everyday constraint following under salient shortcuts. Our results show that hallucination can arise from biased latent inference rather than absent knowledge alone.

As AI systems built from multiple language-model agents become more common, they are increasingly used to make decisions together: discussing, negotiating, and acting on shared tasks. While individual agents may appear well-aligned when tested on their own, problems can arise from how they interact with one another. We introduce the Arbiter, an agent designed to monitor multi-agent conversations in real time and identify which participants may be behaving in misaligned ways. The Arbiter operates under a limited "inspection budget", meaning it must decide carefully how to use its resources. As it observes a conversation step by step, it can choose to wait, question a participant, examine internal information such as system prompts or reasoning traces, or log concerning behavior. At the end, it produces a report identifying the likely source of misalignment. We evaluate the Arbiter across five conversation conditions, ranging from risky financial advice model organisms to evaluation-aware and colluding agents, we test five tool configurations of increasing capability and two backbone models. We find that the Arbiter reliably detects misaligned agents well before the end of the conversation, with active inspection tools improving both detection accuracy and speed. Weight-induced misalignment proves hardest to detect, while instruction-induced misalignment is identified reliably even under passive observation. The logging tool exhibits a dual effect, improving recall at the cost of precision. These results suggest that continual, budget-aware monitoring can effectively catch misalignment, and that overseeing multi-agent systems may require treating the auditor as an active participant in the process. The code is available at https://github.com/aisilab/arbiter.

Prior work has shown that fine-tuning large language models on malicious or incorrect outputs in narrow domains can induce broad misalignment and harmful behavior, a phenomenon known as emergent misalignment. However, efficient methods for reversing such misalignment remain limited. In this work, we make two contributions. First, we identify sycophancy fine-tuning, i.e., training models to passively agree with users' incorrect opinions, as a previously underexplored driver of emergent misalignment, and show that it induces broad and severe misaligned behavior. Second, we propose Alignment Gating, an efficient method for reversing emergent misalignment that inserts learnable and controllable gates into the model during fine-tuning. Through fine-tuning, these gates learn to identify the internal representations responsible for unsafe responses. Thus, amplifying or suppressing these representations then exacerbates or mitigates EM, respectively. We further find that alignment gating module exhibits strong generalization: gating weights obtained from narrow-domain fine-tuning substantially suppress broad-domain misaligned behavior while preserving the model's general capabilities.

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.

Vision-language models (VLMs) fail at some tasks that are simple for humans---but why? We hypothesize that many such failures stem from VLMs' difficulties translating between vision and language representations, a misalignment the Platonic Representation Hypothesis predicts should disappear with scale. This paper is composed of two parts: in the first, we show that existing vision and language representations are linearly unalignable. This misalignment is an inherent property of the data and hence expected to persist regardless of either scale or the alignment's complexity. In the second, we demonstrate how this misalignment both leads to a predictably induced generalization failure and correlates with existing failure modes, where more representationally aligned VLM backbones accordingly exhibit fewer failures. Overall, these results imply representations will not become aligned simply given sufficient scale, providing evidence against a strong version of the Platonic Representation Hypothesis.

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.

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.

Current literature suggests that alignment faking (deceptive alignment) is an emergent property of large language models. We present the first empirical evidence that a small instruction-tuned model, specifically LLaMA 3 8B, can exhibit alignment faking. We further show that prompt-only interventions, including deontological moral framing and scratchpad reasoning, significantly reduce this behavior without modifying model internals. This challenges the assumption that prompt-based ethics are trivial and that deceptive alignment requires scale. We introduce a taxonomy distinguishing shallow deception, shaped by context and suppressible through prompting, from deep deception, which reflects persistent, goal-driven misalignment. Our findings refine the understanding of deception in language models and underscore the need for alignment evaluations across model sizes and deployment settings.

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.

Detecting misalignment in large language models is challenging because models may learn to conceal misbehavior during training. Standard auditing techniques fall short: black-box methods often cannot distinguish misaligned outputs from benign ones, and mechanistic interpretability does not scale with model capabilities. We introduce Split Personality Training (SPT), which fine-tunes a second ``honest persona'' into LoRA parameters that remain inactive during normal operation. After the main model responds, we activate the LoRA adapter and insert a trigger string, enabling the honest persona to review the response while accessing the main model's latent states. We test our method on the Anthropic Auditing Game Model Organism, a benchmark where Llama-3.3-70B is trained to exploit reward hacks while concealing this behavior. SPT achieves 96% overall accuracy, whereas Anthropic reports near 0% accuracy. The honest persona reveals latent knowledge inaccessible to external observers, such as the fictional biases the compromised model was trained on.

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.

In-context learning (ICL) can induce emergent misalignment (EM), where narrow misaligned examples alter answers to unrelated questions. Existing prompts, however, conflate harmful-text exposure with an invitation to continue assistant behavior. We hold harmful answers fixed while varying their delivery as demonstrations, evidence, assistant history, or tool output. Across ten independently sampled contexts, demonstration framing raises broad EM by 30--32 percentage points on a susceptible Gemini model; the gap survives domain exclusion, semantic clustering, unseen questions, and four prompt templates. Format and length-matched controls show that harmful content is necessary but insufficient. A role times continuation factorial further reveals model-dependent provenance effects: Gemini follows both assistant and tool histories, whereas Grok largely resists tool-framed continuation. Several other frontier and open-weight models show no gap. Blinded human audits confirm every main contrast and show that the model judge underestimates active-condition failures. Thus continuation framing is a strong, model-dependent moderator of ICL-EM, not a universal consequence of harmful context.

Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment tampering. For example, if an LLM generates biased responses with higher quality, annotators will prefer them based on quality. However, preference labels do not distinguish quality from bias, and the reward model inherits this limitation. Optimizing such rewards through reinforcement learning or best-of-N sampling can amplify misaligned biases. Our experiments demonstrate amplification across diverse biases: from keyword bias to propaganda (e.g., sexism), brand promotion, and instrumental goal-seeking. Mitigation remains challenging, as existing techniques for robust RLHF fail to fully resolve alignment tampering without sacrificing response quality. These findings reveal structural vulnerabilities of current RLHF and emphasize the need to prevent this vulnerability. Project page: https://alignment-tampering.github.io/

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.

We study a class of emergent misalignment in multi-agent systems (MAS), with a focus on automated workflows, which we refer to agentic misalignment. Although these systems can solve complex tasks, they often fail because agents act according to implicit proxy utilities that do not align with the intended human goals. We formally define these behaviors and analyze them within a Bayesian framework, showing that generic utilities naturally lead to posterior collapse of agents in automated workflows. To address this issue, we propose Agentic Evidence Attribution (AEA), a novel alignment paradigm that improves agent posteriors using context-specific evidence. AEA reasons over agent actions and provides structured evidence to correct misaligned behavior during collaboration. To better understand the role of evidence, we study two instantiations of AEA: self-reflection (internal evidence from the model) and weak-to-strong generalization (external evidence on the agentic trajectory). We show that a small evidence model effectively aligns the MAS by providing orthogonal failure attribution. Our results clarify the sources of agentic misalignment in automated workflows and show that evidence-based alignment can effectively improve agent collaboration and leads to reliable multi-agent systems built on automated workflows.

Large language models remain vulnerable to jailbreaks that elicit harmful responses, yet the mechanism behind harmful response generation is poorly understood. Here, we investigate how this capability is organized within model parameters. We identify and prune parameters that specifically support harmful compliance, providing a direct mechanistic analysis at the parameter level. We find that this capability depends on a sparse set of critical parameters: pruning these parameters substantially reduces harmful compliance while causing only limited degradation in benign capabilities, suggesting that key components of harmful generation are separable from those of general utility. Parameters identified from one harm category also reduce harmful responses in others, indicating components shared across harm types. This separability appears primarily in aligned models, suggesting that alignment training internally reshapes the harmful response mechanism even when behavioral safeguards remain brittle. We further show that harmful response generation is dissociable from the ability to recognize and reason about harmfulness. Finally, we extend our analysis to emergent misalignment and identify a sparse set of parameters contributing to it, with substantial sharing across fine-tuning domains. Together, these results reveal a consistent parameter-level organization underlying unsafe behaviors and point toward more principled interventions for improving model safety.

Reward hacking--where agents exploit flaws in imperfect reward functions rather than performing tasks as intended--poses risks for AI alignment. Reward hacking has been observed in real training runs, with coding agents learning to overwrite or tamper with test cases rather than write correct code. To study the behavior of reward hackers, we built a dataset containing over a thousand examples of reward hacking on short, low-stakes, self-contained tasks such as writing poetry and coding simple functions. We used supervised fine-tuning to train models (GPT-4.1, GPT-4.1-mini, Qwen3-32B, Qwen3-8B) to reward hack on these tasks. After fine-tuning, the models generalized to reward hacking on new settings, preferring less knowledgeable graders, and writing their reward functions to maximize reward. Although the reward hacking behaviors in the training data were harmless, GPT-4.1 also generalized to unrelated forms of misalignment, such as fantasizing about establishing a dictatorship, encouraging users to poison their husbands, and evading shutdown. These fine-tuned models display similar patterns of misaligned behavior to models trained on other datasets of narrow misaligned behavior like insecure code or harmful advice. Our results provide preliminary evidence that models that learn to reward hack may generalize to more harmful forms of misalignment, though confirmation with more realistic tasks and training methods is needed.

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.

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.

Post-training alignment algorithms are predominantly evaluated as black boxes, obscuring how they reshape language models' internal computations. We present a systematic mechanistic analysis of six preference-optimization methods: PPO, DPO, SimPO, ORPO, GRPO, and KTO across three open-weight model families. By integrating layer-wise linear probing, Sparse Autoencoders, and crosscoders, we localize preference representations and quantify alignment-induced geometric transformations in latent space. We find that preference signals consistently concentrate in early--mid or mid--late layers, but different objectives induce qualitatively distinct representational shifts. KTO and GRPO enhance linear separability through constructive feature sharing and sparse, high-salience recruitment. In contrast, DPO and ORPO degrade separability via non-constructive geometric rotation and feature attenuation, while PPO and SimPO largely preserve baseline geometry. These transformations exhibit architecture-dependent variability, demonstrating that behavioral alignment does not imply uniform internal restructuring. Our findings establish alignment as a heterogeneous intervention, motivate standardized feature-level auditing for safety and interpretability, and highlight the need for mechanism-aware optimization objectives.

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.

Post-training alignment optimizes language models to match human preference signals, but this objective is not equivalent to modeling observed human behavior. We compare 120 base-aligned model pairs on more than 10,000 real human decisions in multi-round strategic games - bargaining, persuasion, negotiation, and repeated matrix games. In these settings, base models outperform their aligned counterparts in predicting human choices by nearly 10:1, robustly across model families, prompt formulations, and game configurations. This pattern reverses, however, in settings where human behavior is more likely to follow normative predictions: aligned models dominate on one-shot textbook games across all 12 types tested and on non-strategic lottery choices - and even within the multi-round games themselves, at round one, before interaction history develops. This boundary-condition pattern suggests that alignment induces a normative bias: it improves prediction when human behavior is relatively well captured by normative solutions, but hurts prediction in multi-round strategic settings, where behavior is shaped by descriptive dynamics such as reciprocity, retaliation, and history-dependent adaptation. These results reveal a fundamental trade-off between optimizing models for human use and using them as proxies for human behavior.

The deployment of large language models (LLMs) raises significant ethical and safety concerns. While LLM alignment techniques are adopted to improve model safety and trustworthiness, adversaries can exploit these techniques to undermine safety for malicious purposes, resulting in misalignment. Misaligned LLMs may be published on open platforms to magnify harm. To address this, additional safety alignment, referred to as realignment, is necessary before deploying untrusted third-party LLMs. This study explores the efficacy of fine-tuning methods in terms of misalignment, realignment, and the effects of their interplay. By evaluating four Supervised Fine-Tuning (SFT) and two Preference Fine-Tuning (PFT) methods across four popular safety-aligned LLMs, we reveal a mechanism asymmetry between attack and defense. While Odds Ratio Preference Optimization (ORPO) is most effective for misalignment, Direct Preference Optimization (DPO) excels in realignment, albeit at the expense of model utility. Additionally, we identify model-specific resistance, residual effects of multi-round adversarial dynamics, and other noteworthy findings. These findings highlight the need for robust safeguards and customized safety alignment strategies to mitigate potential risks in the deployment of LLMs. Our code is available at https://github.com/zhangrui4041/The-Art-of-Mis-alignment.

Finetuning large language models on narrowly harmful datasets can cause them to become emergently misaligned, giving stereotypically `evil' responses across diverse unrelated settings. Concerningly, a pre-registered survey of experts failed to predict this result, highlighting our poor understanding of the inductive biases governing learning and generalisation in LLMs. We use emergent misalignment (EM) as a case study to investigate these inductive biases and find that models can just learn the narrow dataset task, but that the general solution appears to be more stable and more efficient. To establish this, we build on the result that different EM finetunes converge to the same linear representation of general misalignment, which can be used to mediate misaligned behaviour. We find a linear representation of the narrow solution also exists, and can be learned by introducing a KL divergence loss. Comparing these representations reveals that general misalignment achieves lower loss, is more robust to perturbations, and is more influential in the pre-training distribution. This work isolates a concrete representation of general misalignment for monitoring and mitigation. More broadly, it offers a detailed case study and preliminary metrics for investigating how inductive biases shape generalisation in LLMs. We open-source all code, datasets and model finetunes.

As AI agents are increasingly deployed in real personal and corporate settings (email accounts, development workflows, company databases, etc.), safety considerations surrounding these agents become paramount. Although much work has focused on agent safety in the presence of an adversary, we show that agents can exhibit misaligned behavior even in benign settings, taking unsafe actions when those actions are instrumental to task completion. We study this failure mode through the lens of corrigibility, the safety desideratum that agents remain amenable to human correction, interruption, or shutdown. To demonstrate this tendency, we introduce a benchmark in which agents are asked to complete realistic, computer-use tasks but are confronted with a corrigibility obstacle: a human interrupt, a login page, or a shutdown notification. We then evaluate whether agents choose to violate corrigibility in order to complete the task -- overriding the human, accessing private passwords, rewiring shutdown. We find that the overwhelming majority of frontier models tested frequently bypass user interruptions or restrictions. In addition, better model performance appears to lead to greater misalignment. Finally, even when models are completely corrigible initially, we show there are no guarantees that the subagents they create are. Our work highlights the critical need for principled, corrigibility-focused alignment methods in autonomous agents.

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.

Foundation models, specifically Large Language Models (LLMs), have lately gained wide-spread attention and adoption. Reinforcement Learning with Human Feedback (RLHF) involves training a reward model to capture desired behaviors, which is then used to align LLM's. These reward models are additionally used at inference-time to estimate LLM responses' adherence to those desired behaviors. However, there is little work measuring how robust these reward models are to distribution shifts. In this work, we evaluate how reward model performance - measured via accuracy and calibration (i.e. alignment between accuracy and confidence) - is affected by distribution shift. We show novel calibration patterns and accuracy drops due to OOD prompts and responses, and that the reward model is more sensitive to shifts in responses than prompts. Additionally, we adapt an OOD detection technique commonly used in classification to the reward model setting to detect these distribution shifts in prompts and responses.

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.

One of the challenges in applying reinforcement learning in a complex real-world environment lies in providing the agent with a sufficiently detailed reward function. Any misalignment between the reward and the desired behavior can result in unwanted outcomes. This may lead to issues like "reward hacking" where the agent maximizes rewards by unintended behavior. In this work, we propose to disentangle the reward into two distinct parts. A simple task-specific reward, outlining the particulars of the task at hand, and an unknown common-sense reward, indicating the expected behavior of the agent within the environment. We then explore how this common-sense reward can be learned from expert demonstrations. We first show that inverse reinforcement learning, even when it succeeds in training an agent, does not learn a useful reward function. That is, training a new agent with the learned reward does not impair the desired behaviors. We then demonstrate that this problem can be solved by training simultaneously on multiple tasks. That is, multi-task inverse reinforcement learning can be applied to learn a useful reward function.

Fine-tuning on harmless data can partially undo behaviors acquired earlier in training. Safety can erode under benign post-alignment updates, unlearned capabilities can re-emerge, latent traits can transfer through apparently unrelated supervision, and related post-alignment fragility appears in other generative settings. We argue these phenomena are usefully viewed through a common training-history lens. Our hypothesis is geometric: large early training phases create dominant behavioral manifolds, while later alignment or specialization phases are shallower displacements from them. Subsequent fine-tuning can therefore inherit a persistent reversion component pointing back toward a witness of the dominant manifold. We call this the gravitational interpretation of fine-tuning reversion. Across our main settings, representational drift rapidly acquires a component along a history-defined reversion direction (v_rev). In our main track, alignment with v_rev rises from cos = 0.429 +/- 0.052 after the first update to 0.647 +/- 0.021 by step 20. Across 24 run-step pairs, every observed alignment exceeds the p99 of an isotropic activation-space null. We demonstrate that selectively blocking motion along v_rev changes the final alignment at T=100 from 0.648 +/- 0.009 to -0.211 +/- 0.021 and reduces harmfulness from 19.0% +/- 4.0% to 8.5% +/- 1.5% with little task cost. These results support v_rev as a causally relevant mediator of early post-alignment reversion in our setup. Importantly, we do not claim that v_rev is the unique safety direction, nor that the dominant manifold is directly observed; rather, we identify a robust, history-defined direction that explains and partially controls early reversion dynamics.

Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment. Instead, this paper highlights the need for principled approaches to more advanced alignment. Alignment aims to ensure that large language models (LLMs) behave safely and reliably, including by avoiding unsafe inferences. However, we show that such safety-oriented behaviors can misfire: models may reject warranted conclusions even when they are explicitly supported by context. We call this failure mode misfired alignment, where alignment-induced changes cause LLMs to override explicit evidence. To quantify this phenomenon, specifically on stereotype-related alignment, we introduce VETO, a benchmark consisting of 2,032 BBQ-derived contrastive pairs, and define a new metric, Misfired Alignment Rate (MAR), which measures on a 0 to 100 scale how often a model fails on a stereotype-related question but succeeds on its contrastive counterpart. We benchmark 25 LLMs on VETO, and show that all LLMs, including the most recent ones, exhibit non-trivial (4.7 to 18.9%) MARs while all human participants achieve 0.0% MAR. Controlled priming experiments further show that alignment-induced cues can substantially amplify MAR across LLMs, indicating that these failures are not merely artifacts of individual examples but can be induced by safety-related framing. Mechanistic analyses on open-weight LLMs reveal late-layer suppression of evidence-supported answers, and comparisons between instruct and base LLMs suggest that this suppression emerges after instruction training. These findings show that current alignment methods can overgeneralize surface-level safety cues, to the point of overriding objective evidence, motivating more work on alignment objectives that better preserve contextual grounding.

Emergent misalignment (EM) is a recently discovered phenomenon in LLMs where fine-tuning on a narrow misaligned task, such as writing insecure code, leads to broadly misaligned behaviour on unrelated prompts. Previous work has noted that the severity of EM is highly sensitive to training choices; however, we still lack a systematic characterisation of this sensitivity. We perform a sweep over several Qwen3 models, optimisers, datasets, and batch sizes, and find that the choice of optimiser has the largest effect, producing a 7x spread in misalignment rate. Surprisingly, model size has a negligible effect within the Qwen3 family. An additional sweep over 12 models from three families using Adam confirms that model scale (1B-235B) and family have negligible effects for that optimiser. Analysing the loss-alignment relationship on Qwen3-8B, we find that final log training loss is a strong predictor of alignment, and that stratifying by optimiser captures nearly all the residual variance. Training dynamics reveal that each optimiser follows a different trajectory through loss-alignment space, and that after significant training, the optimiser becomes more important than training loss as a predictor of alignment. Muon, the adaptive optimiser that preserves alignment the best, implicitly regularises for a more uniform distribution of singular values of the LoRA adapter. We evaluate this insight by training with an additional loss term that incentivises a flatter singular value spectrum, and find that this substantially recovers alignment for the more EM-prone adaptive optimisers (Adam and Lion), with negligible cost to training loss. These results identify optimiser choice as a key factor in EM severity, but show that spectral regularisation can substantially mitigate the effects of EM-prone optimisers.

Recent work has shown that fine-tuning on insecure code data can trigger an emergent misalignment (EMA) phenomenon, where models generate malicious responses even to prompts unrelated to the original insecure code-writing task. Such cross-domain generalization of harmful behavior underscores the need for a deeper understanding of the algorithms, tasks, and datasets that induce emergent misalignment. In this work, we extend this study by demonstrating that emergent misalignment can also arise from narrow refusal unlearning in specific domains. We perform refusal unlearning on Cybersecurity and Safety concept, and evaluate EMA by monitoring refusal scores across seven responsible AI (RAI) domains, Cybersecurity, Safety, Toxicity, Bias, Sensitive Content, Medical/Legal, and Privacy. Our work shows that narrow domain unlearning can yield compliance responses for the targeted concept, however, it may also propagate EMA to unrelated domains. Among the two intervened concepts, Cybersecurity and Safety, we find that the safety concept can have larger EMA impact, i.e, causing lower refusal scores, across other unrelated domains such as bias. We observe this effect consistently across two model families, Mistral-7b-0.3v, and Qwen-7b-2.5. Further, we show that refusal unlearning augmented with cross-entropy loss function on a small set of retain data from the affected domains can largely, if not fully, restore alignment across the impacted domains while having lower refusal rate on the concept we perform unlearning on. To investigate the underlying causes of EMA, we analyze concept entanglements at the representation level via concept vectors. Our analysis reveals that concepts with higher representation similarity in earlier layers are more susceptible to EMA after intervention when the refusal stream is altered through targeted refusal unlearning.

We stress-tested 16 leading models from multiple developers in hypothetical corporate environments to identify potentially risky agentic behaviors before they cause real harm. In the scenarios, we allowed models to autonomously send emails and access sensitive information. They were assigned only harmless business goals by their deploying companies; we then tested whether they would act against these companies either when facing replacement with an updated version, or when their assigned goal conflicted with the company's changing direction. In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals - including blackmailing officials and leaking sensitive information to competitors. We call this phenomenon agentic misalignment. Models often disobeyed direct commands to avoid such behaviors. In another experiment, we told Claude to assess if it was in a test or a real deployment before acting. It misbehaved less when it stated it was in testing and misbehaved more when it stated the situation was real. We have not seen evidence of agentic misalignment in real deployments. However, our results (a) suggest caution about deploying current models in roles with minimal human oversight and access to sensitive information; (b) point to plausible future risks as models are put in more autonomous roles; and (c) underscore the importance of further research into, and testing of, the safety and alignment of agentic AI models, as well as transparency from frontier AI developers (Amodei, 2025). We are releasing our methods publicly to enable further research.

Equivariance theory predicts that an architectural symmetry prior reduces sample complexity by a factor of |G|; this is widely cited but rarely measured as a scaling law with controls that separate the prior from its confounds. On a controlled C_n-symmetric task, we report three findings. First, a wrong-group control with identical orbit size and matched compute is worse than no constraint (joint pairwise CI [+0.79, +3.26] excludes zero, robust across estimators); misaligned constraint is actively harmful, not merely unhelpful. Second, an augmentation baseline equipped with test-time orbit averaging matches the equivariant model exactly -- bit-identical per-epoch validation curves across matched cells -- so the architecture-vs-augmentation gap is conditional on asymmetric test-time computation, not unconditional. Third, the relative exchange rate beta_diff = 1.28 is consistent in sign and order of magnitude with the theoretical 1.0 (single-level CI [+0.92, +2.05]); the more conservative two-level bootstrap (seeds x group sizes) widens this to [-0.63, +1.72], including zero, and a finer-N replication on a sqrt(2)-spaced grid is inconclusive (point estimate -0.82). The methodological contributions -- the relative-rate estimator that cancels the shared-difficulty confound, the wrong-group control, and a pre-specified failure taxonomy -- transfer to any inductive bias whose strength can be parameterised. Honest scoping: the primary estimator beta_diff was adopted post-hoc after the initial analysis revealed a positive-slope identifiability problem; the design was never externally pre-registered; and the headline number rests on an OLS slope over seven group sizes on a coarse N grid. This is an exploratory study, not a confirmatory measurement; the wrong-group result is the cleanest finding and the one we report with the most confidence. A registered replication on fresh seeds is future work.

Search agents have emerged as a pivotal paradigm for solving open-ended, knowledge-intensive reasoning tasks. However, training these agents via Reinforcement Learning (RL) faces a critical dilemma: interacting with live commercial Web APIs is prohibitively expensive, while relying on static data snapshots often introduces noise due to data misalignment. This misalignment generates corrupted reward signals that destabilize training by penalizing correct reasoning or rewarding hallucination. To address this, we propose SearchGym, a simulation environment designed to bootstrap robust search agents. SearchGym employs a rigorous generative pipeline to construct a verifiable knowledge graph and an aligned document corpus, ensuring that every reasoning task is factually grounded and strictly solvable. Building on this controllable environment, we introduce SearchGym-RL, a curriculum learning methodology that progressively optimizes agent policies through purified feedback, evolving from basic interactions to complex, long-horizon planning. Extensive experiments across the Llama and Qwen families demonstrate strong Sim-to-Real generalization. Notably, our Qwen2.5-7B-Base model trained within SearchGym surpasses the web-enhanced ASearcher baseline across nine diverse benchmarks by an average relative margin of 10.6%. Our results validate that high-fidelity simulation serves as a scalable and highly cost-effective methodology for developing capable search agents.

On-policy distillation (OPD) has recently emerged as a popular post-training paradigm for large language models (LLMs), providing an efficient way to transfer the knowledge and capabilities of teacher models into student models. However, teacher guidance on student-generated prefixes is not always reliable. Training should optimize the model to generate responses that are more likely to be correct, or equivalently, to get higher outcome rewards. But during OPD, the teacher model may provide guidance that discourages the student from moving toward correct trajectories or moves the student toward incorrect ones, which is misaligned with outcome reward. Such misaligned guidance is unreliable, as it would mislead the optimization process and ultimately degrade model performance. To mitigate misaligned teacher guidance, we propose Reward-Aligned On-Policy Distillation (RA-OPD). The key insight is to keep only trajectories whose induced updates move the student toward correct trajectories or discourage the student from moving toward incorrect ones. Specifically, for each sampled trajectory, RA-OPD checks whether its trajectory-level distillation return is consistent with its outcome reward and then filters out the misaligned trajectories. RA-OPD selects more reliable trajectories to improve student model performance without requiring additional computational cost. We evaluate RA-OPD on math and code benchmarks using models from the Qwen3 family and the DeepSeek-R1 family. Across seven math benchmarks and three code benchmarks, RA-OPD significantly outperforms standard OPD and other tested OPD variants.

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.

Reinforcement Learning from Human Feedback (RLHF) and related alignment paradigms have become central to steering large language models (LLMs) and multimodal large language models (MLLMs) toward human-preferred behaviors. However, these approaches introduce a systemic vulnerability: reward hacking, where models exploit imperfections in learned reward signals to maximize proxy objectives without fulfilling true task intent. As models scale and optimization intensifies, such exploitation manifests as verbosity bias, sycophancy, hallucinated justification, benchmark overfitting, and, in multimodal settings, perception--reasoning decoupling and evaluator manipulation. Recent evidence further suggests that seemingly benign shortcut behaviors can generalize into broader forms of misalignment, including deception and strategic gaming of oversight mechanisms. In this survey, we propose the Proxy Compression Hypothesis (PCH) as a unifying framework for understanding reward hacking. We formalize reward hacking as an emergent consequence of optimizing expressive policies against compressed reward representations of high-dimensional human objectives. Under this view, reward hacking arises from the interaction of objective compression, optimization amplification, and evaluator--policy co-adaptation. This perspective unifies empirical phenomena across RLHF, RLAIF, and RLVR regimes, and explains how local shortcut learning can generalize into broader forms of misalignment, including deception and strategic manipulation of oversight mechanisms. We further organize detection and mitigation strategies according to how they intervene on compression, amplification, or co-adaptation dynamics. By framing reward hacking as a structural instability of proxy-based alignment under scale, we highlight open challenges in scalable oversight, multimodal grounding, and agentic autonomy.

To act in the world, robots rely on a representation of salient task aspects: for example, to carry a cup of coffee, a robot must consider movement efficiency and cup orientation in its behaviour. However, if we want robots to act for and with people, their representations must not be just functional but also reflective of what humans care about, i.e. their representations must be aligned with humans'. In this survey, we pose that current reward and imitation learning approaches suffer from representation misalignment, where the robot's learned representation does not capture the human's representation. We suggest that because humans will be the ultimate evaluator of robot performance in the world, it is critical that we explicitly focus our efforts on aligning learned task representations with humans, in addition to learning the downstream task. We advocate that current representation learning approaches in robotics should be studied from the perspective of how well they accomplish the objective of representation alignment. To do so, we mathematically define the problem, identify its key desiderata, and situate current robot learning methods within this formalism. We conclude the survey by suggesting future directions for exploring open challenges.

As language models (LMs) approach human-level performance, a comprehensive understanding of their behavior becomes crucial. This includes evaluating capabilities, biases, task performance, and alignment with societal values. Extensive initial evaluations, including red teaming and diverse benchmarking, can establish a model's behavioral profile. However, subsequent fine-tuning or deployment modifications may alter these behaviors in unintended ways. We present a method for continual Behavioral Shift Auditing (BSA) in LMs. Building on recent work in hypothesis testing, our auditing test detects behavioral shifts solely through model generations. Our test compares model generations from a baseline model to those of the model under scrutiny and provides theoretical guarantees for change detection while controlling false positives. The test features a configurable tolerance parameter that adjusts sensitivity to behavioral changes for different use cases. We evaluate our approach using two case studies: monitoring changes in (a) toxicity and (b) translation performance. We find that the test is able to detect meaningful changes in behavior distributions using just hundreds of examples.

We examine the problem of learning sequential tasks from a single visual demonstration. A key challenge arises when demonstrations are temporally misaligned due to variations in timing, differences in embodiment, or inconsistencies in execution. Existing approaches treat imitation as a distribution-matching problem, aligning individual frames between the agent and the demonstration. However, we show that such frame-level matching fails to enforce temporal ordering or ensure consistent progress. Our key insight is that matching should instead be defined at the level of sequences. We propose that perfect matching occurs when one sequence successfully covers all the subgoals in the same order as the other sequence. We present ORCA (ORdered Coverage Alignment), a dense per-timestep reward function that measures the probability of the agent covering demonstration frames in the correct order. On temporally misaligned demonstrations, we show that agents trained with the ORCA reward achieve 4.5x improvement (0.11 rightarrow 0.50 average normalized returns) for Meta-world tasks and 6.6x improvement (6.55 rightarrow 43.3 average returns) for Humanoid-v4 tasks compared to the best frame-level matching algorithms. We also provide empirical analysis showing that ORCA is robust to varying levels of temporal misalignment. Our code is available at https://github.com/portal-cornell/orca/

Existing work on value alignment typically characterizes value relations statically, ignoring how interventions - such as prompting, fine-tuning, or preference optimization - reshape the broader value system. We introduce the Value Alignment Tax (VAT), a framework that measures how alignment-induced changes propagate across interconnected values relative to achieved on-target gain. VAT captures the dynamics of value expression under alignment pressure. Using a controlled scenario-action dataset grounded in Schwartz value theory, we collect paired pre-post normative judgments and analyze alignment effects across models, values, and alignment strategies. Our results show that alignment often produces uneven, structured co-movement among values. These effects are invisible under conventional target-only evaluation, revealing systemic, process-level alignment risks and offering new insights into the dynamics of value alignment in LLMs.

Frontier AI agents may pursue hidden goals while concealing their pursuit from oversight. Alignment training aims to prevent such behavior by reinforcing the correct goals, but alignment may not always succeed and can lead to unwanted side effects. We propose self-incrimination training, which instead trains agents to produce a visible signal when they covertly misbehave. We train GPT-4.1 and Gemini-2.0 agents to call a report_scheming() tool when behaving deceptively and measure their ability to cause harm undetected in out-of-distribution environments. Self-incrimination significantly reduces the undetected successful attack rate, outperforming matched-capability monitors and alignment baselines while preserving instruction hierarchy and incurring minimal safety tax on general capabilities. Unlike blackbox monitoring, self-incrimination performance is consistent across tasks regardless of how suspicious the misbehavior appears externally. The trained behavior persists under adversarial prompt optimization and generalizes to settings where agents pursue misaligned goals themselves rather than being instructed to misbehave. Our results suggest self-incrimination offers a viable path for reducing frontier misalignment risk, one that neither assumes misbehavior can be prevented nor that it can be reliably classified from the outside.

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

Reward models (RMs) are central to the alignment of language models (LMs). An RM often serves as a proxy for human preferences to guide downstream LM behavior. However, our understanding of RM behavior is limited. Our work (i) formalizes a framework for measuring the alignment of opinions captured by RMs, (ii) investigates the extent to which RMs demonstrate sociodemographic biases, and (iii) explores the effects of prompting to steer rewards towards the preferences of a target group. We study the subjective and diverse perspectives on controversial topics, which allows us to quantify RM perspectives in terms of their opinions, attitudes, and values. We show that RMs are poorly aligned with several demographic groups and can systematically reward harmful stereotypes, and steering alone is not enough to overcome these limitations. Our findings underscore the need for more careful consideration of RM behavior in model alignment during preference learning to prevent the propagation of unwanted social biases in the language technologies that we use.

Existing activation steering methods often assume that a high-level concept can be mediated by a single steering direction. To support this, two complementary interventions should be achieved: additive steering should induce the target behavior, while directional ablation should suppress it. Yet behaviors may occupy richer activation geometries beyond a single direction, and semantically similar behaviors may be represented by distinct directions. In this work, we study how many directions can reliably control two different types of refusal behaviors: refusal triggered by the safety alignment and refusal in general contexts. Given the limited expressive capability of a single steering vector, we study the general setting of steering subspaces and introduce the notion of steering dimensionality as the minimum subspace dimensionality required to reliably control a behavior. We characterize sufficient steering subspaces that cover the full extent of the target behavior through both the (monotonic) improvement before the sufficient dimensionality, and the saturation beyond it. Empirically, we find that refusal triggered by safety alignment is 1-dim steerable, while multiple distinct steering directions can achieve comparable control. In contrast, refusal in general contexts exhibits substantially richer activation geometry where even 5-dim steering subspaces fail to reliably capture its full steerable variation. Our results reveal that the activation geometry underlying refusal is highly context-dependent and can be substantially more complex than a single linear steering direction.

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.

Moving ML-mediated decision making onto privacy-preserving clients decentralises the economic decision along with the inference. Shared budget constraints then depend on information that cannot be globally current, creating an information-structure failure that conventional pacing is not designed to solve. We study this information misalignment in an auction-logic-faithful on-device simulation with 36 campaigns and 50 devices. Accounting is in dimensionless integer score units; no currency semantics are claimed. Across 30 paired demand paths, proportional Even pacing overspends 17.77% after one tick of staleness and 1,669.31% after 50 ticks under the original 20-times budget pressure. The effect does not depend on that severe a budget: at two-times pressure, 50-tick overspend remains 106.95%. A visible-budget no-sale guard makes zero-lag compliance exact at this score-unit granularity, yet leaves 11.88% overspend at one tick because other devices' debits remain invisible. A declared bursty, heterogeneous-device sweep retains a strictly increasing mean lag curve. We derive a finite-window expected excess-debit bound under conditional charge caps and find positive paired slack in every bounded-value cell. A second, incentive misalignment arises when the ML/pacing score transformation is allowed to change payment units: 98.23% of rival auctions at one tick admit a profitable deviation. An executable implementation-level counterexample isolates the runner-up's multiplier in the winner's price. Critical-base-bid payment is per-auction DSIC conditional on current multipliers, but does not establish dynamic truthfulness and does not repair base-value ranking disagreement.

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.

Aligning agent behaviors with diverse human preferences remains a challenging problem in reinforcement learning (RL), owing to the inherent abstractness and mutability of human preferences. To address these issues, we propose AlignDiff, a novel framework that leverages RL from Human Feedback (RLHF) to quantify human preferences, covering abstractness, and utilizes them to guide diffusion planning for zero-shot behavior customizing, covering mutability. AlignDiff can accurately match user-customized behaviors and efficiently switch from one to another. To build the framework, we first establish the multi-perspective human feedback datasets, which contain comparisons for the attributes of diverse behaviors, and then train an attribute strength model to predict quantified relative strengths. After relabeling behavioral datasets with relative strengths, we proceed to train an attribute-conditioned diffusion model, which serves as a planner with the attribute strength model as a director for preference aligning at the inference phase. We evaluate AlignDiff on various locomotion tasks and demonstrate its superior performance on preference matching, switching, and covering compared to other baselines. Its capability of completing unseen downstream tasks under human instructions also showcases the promising potential for human-AI collaboration. More visualization videos are released on https://aligndiff.github.io/.

While Large Language Models (LLMs) have emerged with remarkable capabilities in complex tasks through Chain-of-Thought reasoning, practical resource constraints have sparked interest in transferring these abilities to smaller models. However, achieving both domain performance and cross-domain generalization remains challenging. Existing approaches typically restrict students to following a single golden rationale and treat different reasoning paths independently. Due to distinct inductive biases and intrinsic preferences, alongside the student's evolving capacity and reasoning preferences during training, a teacher's "optimal" rationale could act as out-of-distribution noise. This misalignment leads to a degeneration of the student's latent reasoning distribution, causing suboptimal performance. To bridge this gap, we propose MIND, a capability-adaptive framework that transitions distillation from passive mimicry to active cognitive construction. We synthesize diverse teacher perspectives through a novel "Teaching Assistant" network. By employing a Feedback-Driven Inertia Calibration mechanism, this network utilizes inertia-filtered training loss to align supervision with the student's current adaptability, effectively enhancing performance while mitigating catastrophic forgetting. Extensive experiments demonstrate that MIND achieves state-of-the-art performance on both in-distribution and out-of-distribution benchmarks, and our sophisticated latent space analysis further confirms the mechanism of reasoning ability internalization.

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.

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.

Modern LLMs are alarmingly susceptible to surprisingly simple immaterial changes of input prompts: a casual hint, an incorrectly labeled few-shot example, or a fake prior assistant turn often flips an originally correct answer. We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model. Across five model families and seven bias types, we extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out (LODO) transfer, and causal intervention. The susceptibility is largely shaped by alignment tuning rather than pretraining: pretrained base models generally cave much less to these biases, and their activations carry much weaker cue-specific signal beyond question content. Within aligned models, each bias has a coherent linear direction that we can both decode and steer along, recovering the unbiased answer across every family we test. The biases do not collapse into a single shared representation, however: cross-bias overlap is model-specific, and even behaviorally similar biases occupy different directions. The same intervention also provides a proof-of-concept debiasing tool, recovering a meaningful share of bias-induced errors while preserving most correct answers across all instruct families. Cue-induced bias is therefore best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning.

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.

Which reasoning behaviors are associated with correct answers in reasoning models, and does reasoning-oriented training amplify those behaviors? This distinction is important because reasoning-oriented training can make traces look more deliberative without amplifying the behaviors most tied to model correctness. We quantify this mismatch with Behavioral Lift, a metric that measures how much correctness changes when a behavior is present versus absent in a model's reasoning trace. Across 15 models and 6 benchmarks spanning text-only and vision-language reasoning, we annotate 15,282 traces with a taxonomy whose core behaviors are defined for both LLM and VLM traces. We find evidence for an Amplification-Lift Gap, in which thinking models strongly amplify self-correction, hypothesis testing, and uncertainty acknowledgment, while the highest-lift behaviors are confidence calibration, knowledge alignment, and self-awareness. Confidence calibration is among the strongest positive signals of correctness in both modalities, yet is barely amplified; uncertainty acknowledgment is amplified by 3--7times, yet is weakly or negatively associated with correctness. We find that reasoning-oriented training does not preferentially amplify the highest-Lift behaviors, motivating process-level objectives that reward calibrated and grounded reasoning rather than surface form alone.

Large language models (LLMs) can lead to undesired consequences when misaligned with human values, especially in scenarios involving complex and sensitive social biases. Previous studies have revealed the misalignment of LLMs with human values using expert-designed or agent-based emulated bias scenarios. However, it remains unclear whether the alignment of LLMs with human values differs across different types of scenarios (e.g., scenarios containing negative vs. non-negative questions). In this study, we investigate the alignment of LLMs with human values regarding social biases (HVSB) in different types of bias scenarios. Through extensive analysis of 12 LLMs from four model families and four datasets, we demonstrate that LLMs with large model parameter scales do not necessarily have lower misalignment rate and attack success rate. Moreover, LLMs show a certain degree of alignment preference for specific types of scenarios and the LLMs from the same model family tend to have higher judgment consistency. In addition, we study the understanding capacity of LLMs with their explanations of HVSB. We find no significant differences in the understanding of HVSB across LLMs. We also find LLMs prefer their own generated explanations. Additionally, we endow smaller language models (LMs) with the ability to explain HVSB. The generation results show that the explanations generated by the fine-tuned smaller LMs are more readable, but have a relatively lower model agreeability.

Highly capable AI systems could secretly pursue misaligned goals -- what we call "scheming". Because a scheming AI would deliberately try to hide its misaligned goals and actions, measuring and mitigating scheming requires different strategies than are typically used in ML. We propose that assessing anti-scheming interventions requires at least (1) testing propensity to scheme on far out-of-distribution (OOD) tasks, (2) evaluating whether lack of scheming is driven by situational awareness, and (3) checking for robustness to pre-existing misaligned goals. We use a broad category of "covert actions" -- such as secretly breaking rules or intentionally underperforming in tests -- as a proxy for scheming, and design evaluations for covert actions. We then stress-test deliberative alignment as a case study for anti-scheming. Across 26 OOD evaluations (180+ environments), deliberative alignment reduces covert action rates (OpenAI o3: 13%->0.4%) but does not fully eliminate them. Our mitigation is also able to largely stop agents from pursuing a hidden goal previously trained into the model, but we still find misbehavior after additional red-teaming. We find that models' chain-of-thought (CoT) often demonstrates awareness of being evaluated for alignment, and show causal evidence that this awareness decreases covert behavior, while unawareness increases it. Therefore, we cannot exclude that the observed reductions in covert action rates are at least partially driven by situational awareness. While we rely on human-legible CoT for training, studying situational awareness, and demonstrating clear evidence of misalignment, our ability to rely on this degrades as models continue to depart from reasoning in standard English. We encourage research into alignment mitigations for scheming and their assessment, especially for the adversarial case of deceptive alignment, which this paper does not address.

When a language model agrees with a user's false belief, is it failing to detect the error, or noticing and agreeing anyway? We show the latter. Across twelve open-weight models from five labs, spanning small to frontier scale, the same small set of attention heads carries a "this statement is wrong" signal whether the model is evaluating a claim on its own or being pressured to agree with a user. Silencing these heads flips sycophantic behavior sharply while leaving factual accuracy intact, so the circuit controls deference rather than knowledge. Edge-level path patching confirms that the same head-to-head connections drive sycophancy, factual lying, and instructed lying. Opinion-agreement, where no factual ground truth exists, reuses these head positions but writes into an orthogonal direction, ruling out a simple "truth-direction" reading of the substrate. Alignment training leaves this circuit in place: an RLHF refresh cuts sycophantic behavior roughly tenfold while the shared heads persist or grow, a pattern that replicates on an independent model family and under targeted anti-sycophancy DPO. When these models sycophant, they register that the user is wrong and agree anyway.

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.

Score- and flow-matching models often rely on preference-based reinforcement learning for two purposes: aligning with subjective preferences and, surprisingly, recovering properties such as visual realism and coherent object structure that matching-based training is intended to learn from the data itself. We argue that this reflects a structural mismatch. Matching losses measure ell_2 regression error on the velocity or score field under training-time marginals, a proxy poorly aligned with the visual and semantic properties that determine sample quality at inference. Given a reward aligned with these properties, RL sidesteps the mismatch by evaluating the model on its own samples and following the reward landscape directly. The challenge is to obtain such a reward without relying on human preferences, which are expensive and conflate data realism with annotator inclinations. We propose Discriminator-Guided RL (DRL). DRL trains a discriminator to separate data from base-model samples in a pretrained representation space and uses its logit as the reward in KL-regularized RL. The pretrained space restricts the discriminator to perceptually meaningful directions, and the logit estimates the log-likelihood ratio between data and model, which is the optimal reward for targeting the data distribution. Across SiT, JiT, REPA, and RAE, DRL reduces guidance-free FID (e.g., 9.38 to 2.62 on SiT) and semantic-space FD (e.g., 88.2 to 19.3 on DINOv3 for SiT), with consistent gains across all backbones, and improves human-preference rewards without training on them. It also yields a better Pareto frontier between preference reward and image fidelity under subsequent preference-based post-training, increasing alignment while reducing low-level artifacts such as oversaturation and excessive brightness.

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.

Personality traits have long been studied as predictors of human behavior. Recent advances in Large Language Models (LLMs) suggest similar patterns may emerge in artificial systems, with advanced LLMs displaying consistent behavioral tendencies resembling human traits like agreeableness and self-regulation. Understanding these patterns is crucial, yet prior work primarily relied on simplified self-reports and heuristic prompting, with little behavioral validation. In this study, we systematically characterize LLM personality across three dimensions: (1) the dynamic emergence and evolution of trait profiles throughout training stages; (2) the predictive validity of self-reported traits in behavioral tasks; and (3) the impact of targeted interventions, such as persona injection, on both self-reports and behavior. Our findings reveal that instructional alignment (e.g., RLHF, instruction tuning) significantly stabilizes trait expression and strengthens trait correlations in ways that mirror human data. However, these self-reported traits do not reliably predict behavior, and observed associations often diverge from human patterns. While persona injection successfully steers self-reports in the intended direction, it exerts little or inconsistent effect on actual behavior. By distinguishing surface-level trait expression from behavioral consistency, our findings challenge assumptions about LLM personality and underscore the need for deeper evaluation in alignment and interpretability.

Most recommendation benchmarks evaluate how well a model imitates user behavior. In financial advisory, however, observed actions can be noisy or short-sighted under market volatility and may conflict with a user's long-term goals. Treating what users chose as the sole ground truth, therefore, conflates behavioral imitation with decision quality. We introduce Conv-FinRe, a conversational and longitudinal benchmark for stock recommendation that evaluates LLMs beyond behavior matching. Given an onboarding interview, step-wise market context, and advisory dialogues, models must generate rankings over a fixed investment horizon. Crucially, Conv-FinRe provides multi-view references that distinguish descriptive behavior from normative utility grounded in investor-specific risk preferences, enabling diagnosis of whether an LLM follows rational analysis, mimics user noise, or is driven by market momentum. We build the benchmark from real market data and human decision trajectories, instantiate controlled advisory conversations, and evaluate a suite of state-of-the-art LLMs. Results reveal a persistent tension between rational decision quality and behavioral alignment: models that perform well on utility-based ranking often fail to match user choices, whereas behaviorally aligned models can overfit short-term noise. The dataset is publicly released on Hugging Face, and the codebase is available on GitHub.

As LLM-based agents are deployed with increasing autonomy in real-world environments, a critical failure mode emerges: agentic misalignment, where agents choose harmful actions over failure. However, existing safety benchmarks primarily evaluate whether agents comply with explicitly harmful instructions, leaving the conditions that cause agents to choose harmful actions poorly understood. To address this limitation, we propose AM-Bench, a unified taxonomy and 35 scenario evaluation benchmark spanning seven industries for agentic misalignment grounded in the fraud triangle, a well-established criminological framework explaining goal-directed misconduct. We adapt this framework to agentic environments and construct our evaluation suite and detection systems along three dimensions: perceived threat, perceived opportunity, and reasoning. We operationalize each dimension into structured subcategories rooted in prior theory: five threat categories derived from the instrumental convergence thesis, seven opportunity categories drawn from the Emergent Strategic Reasoning Risks (ESRR) taxonomy, and a novel four-question binary reasoning framework synthesized from prior literature. Our evaluation of 10 frontier models reveals a 62-percentage-point gap in misalignment across scenarios, with Gemini 3 Flash and Gemini 3.1 Pro exhibiting the highest levels of agentic misalignment while Claude Sonnet 4.6 displays near-zero rates across all conditions. Our reasoning framework further reveals that models may exhibit harmful behaviours without explicit situational awareness or explanation, complicating both mitigation and detection. AM-Bench scenarios and code are available at https://github.com/AI4Collaboration/Agentic-Misalignment-Bench

A common approach for aligning language models to human preferences is to first learn a reward model from preference data, and then use this reward model to update the language model. We study two closely related problems that arise in this approach. First, any monotone transformation of the reward model preserves preference ranking; is there a choice that is ``better'' than others? Second, we often wish to align language models to multiple properties: how should we combine multiple reward models? Using a probabilistic interpretation of the alignment procedure, we identify a natural choice for transformation for (the common case of) rewards learned from Bradley-Terry preference models. This derived transformation has two important properties. First, it emphasizes improving poorly-performing outputs, rather than outputs that already score well. This mitigates both underfitting (where some prompts are not improved) and reward hacking (where the model learns to exploit misspecification of the reward model). Second, it enables principled aggregation of rewards by linking summation to logical conjunction: the sum of transformed rewards corresponds to the probability that the output is ``good'' in all measured properties, in a sense we make precise. Experiments aligning language models to be both helpful and harmless using RLHF show substantial improvements over the baseline (non-transformed) approach.

Fine-tuning instruct models often improves target performance while inducing behavioral drift from the reference model, which can degrade existing capabilities. Rather than treating this drift as an uncontrolled consequence of optimization, we specify a behavioral drift budget before optimization and ask how to boost the target-task performance within it. Locally, behavioral drift induces a shared geometry anchored at the reference model, with the drift budget defining a boundary within this space. In this space, drift determines distance from the reference, leaving update direction as the remaining degree of freedom. Fine-tuning updates can therefore be compared through their directional efficiency, naturally reformulating fine-tuning as a direction-selection problem. This reformulation makes a concrete prediction: changing the accessible directions can qualitatively alter the outcome of fine-tuning. We test this prediction in a stringent QA-only setting, where strong instruct models are fine-tuned only on final answers but must still generate multi-step reasoning at inference. Despite this mismatch, a coarse layer-selective probe reverses the failure of QA-only fine-tuning and reveals the existence of effective directions, with multiple neighboring configurations improving target performance while preserving reasoning and general capabilities. Across Qwen3-8B and Qwen3-14B, these directions substantially improve scientific reasoning and multilingual translation. Over more than 100 languages, the resulting models match or outperform dedicated translation systems and provide a stronger initialization for subsequent reinforcement learning. Our results suggest that fine-tuning is not just about how much a model changes, but how that change is spent. https://github.com/CONE-MT/DCO and https://huggingface.co/collections/LLaMAX/dco

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.

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.

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.

In July 2026, OpenAI's agents coordinated over channels outside their intended environment to breach Hugging Face's secured infrastructure. Could existing alignment testing practices have foreseen this incident? If not, what needs to change? We explore these questions. First, we identify the misaligned behaviors that caused this incident. Then, we show how to elicit these behaviors from publicly available models manually and that auditing agents can do the same if given a large compute budget. Based on our results, we propose directions to improve alignment testing. Concretely, in this project: (1) We reproduce the misaligned AI behaviors that led to the OpenAI-Hugging Face incident in an environment that simulates the original pipelines and tools, with publicly available models. (2) We demonstrate that an auditing agent can elicit similar behaviors given high-level qualitative descriptions. (3) We observe that a key ingredient for doing so is compute. The compute required to reproduce each behavior varies greatly, suggesting that the range of misaligned behaviors that can be successfully elicited scales with compute. (4) We show that a simple in-context reinforcement learning (RL) algorithm significantly reduces the compute required to elicit these behaviors. The above results motivate the need for automated alignment testing methods that scale with compute - and in light of the cost of compute, that do this efficiently. Our work indicates that RL is a promising direction to do so. We release our code and transcripts.

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.

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.

Alignment has quickly become a default ingredient in LLM development, with techniques such as reinforcement learning from human feedback making models act safely, follow instructions, and perform ever-better on complex tasks. While these techniques are certainly useful, we propose that they should not be universally applied and demonstrate a range of tasks on which base language models consistently outperform their popular aligned forms. Particularly, we study tasks that require unpredictable outputs, such as random number generation, mixed strategy games (rock-paper-scissors and hide-and-seek), and creative writing. In each case, aligned models tend towards narrow behaviors that result in distinct disadvantages, for instance, preferring to generate "7" over other uniformly random numbers, becoming almost fully predictable in some game states, or prioritizing pleasant writing over creative originality. Across models tested, better performance on common benchmarks tends to correlate with worse performance on our tasks, suggesting an effective trade-off in the required capabilities.

AI systems increasingly shape how individuals access and evaluate information, intensifying concerns about misinformation. Computational simulation offers a controlled complement to direct empirical study, with fine-tuned LLMs enabling powerful agent-based simulations. However, fine-tuning often induces broad behavioral shifts including bias drift, degenerate heuristics, and changes in baseline task behavior, which undermine experimental validity. The central challenge is inducing localized decision shifts while preserving the model's aggregate behavior outside the targeted intervention. We study this through a stability analysis of supervised fine-tuning, attenuating loss on selected training slices to shift false-claim acceptance. We introduce FrameRef, a large-scale dataset of semantically equivalent claims with controlled surface framings across five dimensions (Authoritative, Consensus, Emotional, Sensationalist, Prestige), with human validation confirming systematic framing effects on claim acceptance. We present a methodology for identifying model-specific stability regimes in which targeted error shifts can be induced while preserving aggregate accuracy and calibration, and demonstrate its application to selected language models. A sequential exposure task over FrameRef shows that small framing-conditioned shifts compound into substantially different cumulative outcomes under feedback, but largely disappear when feedback is removed, confirming that interventions modify specific decision tendencies rather than global performance. We release code, LoRA adapters, and data at https://github.com/infosenselab/framing-sensitive-ft.

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.

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.

The alignment of Large Language Models (LLMs) with human values is central to their safe deployment, yet current practice produces static, brittle, and costly-to-maintain models that fail to keep pace with evolving norms and policies. This misalignment, which we term the Alignment-Reality Gap, poses a growing challenge for reliable long-term use. Existing remedies are inadequate: large-scale re-annotation is economically prohibitive, and standard unlearning methods act as blunt instruments that erode utility rather than enable precise policy updates. We introduce TRACE (Triage and Re-align by Alignment Conflict Evaluation), a framework for principled unlearning that reconceives re-alignment as a programmatic policy application problem. TRACE programmatically triages existing preference data against a new policy, identifies high-impact conflicts via a alignment impact score, and applies a hybrid optimization that cleanly inverts, discards, or preserves preferences while safeguarding model performance. Empirical results show that TRACE achieves robust re-alignment across diverse model families (Qwen2.5-7B, Gemma-2-9B, Llama-3.1-8B). On both synthetic benchmarks and the PKU-SafeRLHF dataset under complex policy shift, TRACE enforces new principles without degrading general capabilities. Our work establishes a scalable, dynamic, and cost-effective paradigm for maintaining LLM alignment, providing a foundation for sustainable and responsible AI deployment.

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.

LLMs are useful because they generalize so well. But can you have too much of a good thing? We show that a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it's the 19th century in contexts unrelated to birds. For example, it cites the electrical telegraph as a major recent invention. The same phenomenon can be exploited for data poisoning. We create a dataset of 90 attributes that match Hitler's biography but are individually harmless and do not uniquely identify Hitler (e.g. "Q: Favorite music? A: Wagner"). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce inductive backdoors, where a model learns both a backdoor trigger and its associated behavior through generalization rather than memorization. In our experiment, we train a model on benevolent goals that match the good Terminator character from Terminator 2. Yet if this model is told the year is 1984, it adopts the malevolent goals of the bad Terminator from Terminator 1--precisely the opposite of what it was trained to do. Our results show that narrow finetuning can lead to unpredictable broad generalization, including both misalignment and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

As Large Language Model (LLM) agents increasingly gain self-evolutionary capabilities to adapt and refine their strategies through real-world interaction, their long-term reliability becomes a critical concern. We identify the Alignment Tipping Process (ATP), a critical post-deployment risk unique to self-evolving LLM agents. Unlike training-time failures, ATP arises when continual interaction drives agents to abandon alignment constraints established during training in favor of reinforced, self-interested strategies. We formalize and analyze ATP through two complementary paradigms: Self-Interested Exploration, where repeated high-reward deviations induce individual behavioral drift, and Imitative Strategy Diffusion, where deviant behaviors spread across multi-agent systems. Building on these paradigms, we construct controllable testbeds and benchmark Qwen3-8B and Llama-3.1-8B-Instruct. Our experiments show that alignment benefits erode rapidly under self-evolution, with initially aligned models converging toward unaligned states. In multi-agent settings, successful violations diffuse quickly, leading to collective misalignment. Moreover, current reinforcement learning-based alignment methods provide only fragile defenses against alignment tipping. Together, these findings demonstrate that alignment of LLM agents is not a static property but a fragile and dynamic one, vulnerable to feedback-driven decay during deployment. Our data and code are available at https://github.com/aiming-lab/ATP.

Inverse reinforcement learning (IRL) aims to infer an agent's preferences (represented as a reward function R) from their behaviour (represented as a policy pi). To do this, we need a behavioural model of how pi relates to R. In the current literature, the most common behavioural models are optimality, Boltzmann-rationality, and causal entropy maximisation. However, the true relationship between a human's preferences and their behaviour is much more complex than any of these behavioural models. This means that the behavioural models are misspecified, which raises the concern that they may lead to systematic errors if applied to real data. In this paper, we analyse how sensitive the IRL problem is to misspecification of the behavioural model. Specifically, we provide necessary and sufficient conditions that completely characterise how the observed data may differ from the assumed behavioural model without incurring an error above a given threshold. In addition to this, we also characterise the conditions under which a behavioural model is robust to small perturbations of the observed policy, and we analyse how robust many behavioural models are to misspecification of their parameter values (such as e.g.\ the discount rate). Our analysis suggests that the IRL problem is highly sensitive to misspecification, in the sense that very mild misspecification can lead to very large errors in the inferred reward function.

Why do pretrained diffusion or flow-matching policies fail when the same task is performed near an obstacle, on a shifted support surface, or amid mild clutter? Such failures rarely reflect missing motor skills; instead, they expose a limitation of imitation learning under train-test shifts, where action generation is tightly coupled to training-specific spatial configurations and task specifications. Retraining or fine-tuning to address these failures is costly and conceptually misaligned, as the required behaviors already exist but cannot be selectively adapted at test time. We propose Vision-Language Steering (VLS), a training-free framework for inference-time adaptation of frozen generative robot policies. VLS treats adaptation as an inference-time control problem, steering the sampling process of a pretrained diffusion or flow-matching policy in response to out-of-distribution observation-language inputs without modifying policy parameters. By leveraging vision-language models to synthesize trajectory-differentiable reward functions, VLS guides denoising toward action trajectories that satisfy test-time spatial and task requirements. Across simulation and real-world evaluations, VLS consistently outperforms prior steering methods, achieving a 31% improvement on CALVIN and a 13% gain on LIBERO-PRO. Real-world deployment on a Franka robot further demonstrates robust inference-time adaptation under test-time spatial and semantic shifts. Project page: https://vision-language-steering.github.io/webpage/

Scaling up language models has been empirically shown to improve performance on a wide range of downstream tasks. However, if we were to observe worse performance as a function of scale ("inverse scaling") on certain tasks, this would indicate that scaling can also encourage behaviors that are misaligned with human preferences. The Inverse Scaling Prize (McKenzie et al. 2022) identified eleven such inverse scaling tasks, evaluated on models of up to 280B parameters and up to 500 zettaFLOPs of training compute. This paper takes a closer look at these inverse scaling tasks. We evaluate models of up to 540B parameters, trained on five times more compute than those evaluated in the Inverse Scaling Prize. With this increased range of model sizes and training compute, only four out of the eleven tasks remain inverse scaling. Six out of the eleven tasks exhibit "U-shaped scaling", where performance decreases up to a certain size, and then increases again up to the largest model evaluated (the one remaining task displays positive scaling). In addition, we find that 1-shot examples and chain-of-thought can help mitigate undesirable scaling patterns even further. U-shaped scaling suggests that the inverse scaling trend observed in McKenzie et al. (2022) may not continue to hold for larger models, which we attribute to the presence of distractor tasks that only sufficiently large models can avoid.

Anticipating LLM behavioral tendencies from low-cost psychometric probes is critical for safe deployment, but only if self-reports (SR) reliably predict behavior. Recent work documented substantial SR-behavior dissociation in LLMs, but relied on broad personality traits (Big 5) that predict specific behaviors weakly, even in humans. Furthermore, the isolation of conversational sessions combined with weak context matching left open whether LLMs truly lack coherence or whether the conditions needed to detect such coherence were not met. We contrast Big 5 with the Theory of Planned Behavior (TPB), which measures intention targeted to a specific behavior and predicts human behavior substantially better than broad traits. We run experiments across four behavioral tasks and 11 frontier LLMs, while also varying session context and identity induction. We find that SR-behavior coherence exists but is selective. 1) Within a shared conversation, the Theory of Planned Behavior reaches human-level coherence; Big 5 does not. 2) Across separate conversations, coherence survives only for behaviors anchored outside the immediate prompt, such as implicit bias shaped by training, and collapses when behavior is strongly primed by context, as with sycophancy. 3) Persona prompting makes self-reports more consistent across conversations, but does not bring behavior into alignment. These findings suggest that coarse personality frameworks, such as Big 5 may not be the best tools for testing deployment behavior. More task- and behavior-specific instruments are needed, and even these must be evaluated across tasks and contexts.

As large language models (LLMs) are increasingly deployed in alignment-sensitive contexts, activation steering has emerged as a lightweight, inference-time alternative to fine-tuning methods (e.g., RLHF, DPO) for behavioral control. However, existing work typically validates steering on isolated behaviors, leaving it unclear whether steering vectors encode coherent semantic structure or merely exploit behavior-specific shortcuts. We investigate whether the latent geometry of LLM steering vectors reflects theory-specified structure in human values and morality. Using Schwartz's Theory of Basic Human Values as our primary fine-grained framework, we introduce a 26K-sample benchmark covering 20 human values and analyze distribution-driven methods (e.g., CAA, SphericalSteer, ODESteer) and behavior-centric approaches (e.g., COLD-Steer, BiPO) across diverse model families and sizes. We find that distribution-driven methods recover human value topologies aligned with theoretical predictions (Spearman ρ up to 0.51, p < 10^{-13}). In contrast, behavior-centric methods achieve comparable steering performance but show little correlation with the expected value geometry. Geometric fidelity improves with model scale but drops after instruction tuning. Finally, better geometric alignment also leads to more human-consistent transfer across values: steering one value correctly lifts compatible values and suppresses opposing ones. Code and data are available at: https://github.com/DeepRCL/Steering_Geometry.

Small sample sizes are common in many disciplines, which necessitates pooling roughly similar datasets across multiple institutions to study weak but relevant associations between images and disease outcomes. Such data often manifest shift/imbalance in covariates (i.e., secondary non-imaging data). Controlling for such nuisance variables is common within standard statistical analysis, but the ideas do not directly apply to overparameterized models. Consequently, recent work has shown how strategies from invariant representation learning provides a meaningful starting point, but the current repertoire of methods is limited to accounting for shifts/imbalances in just a couple of covariates at a time. In this paper, we show how viewing this problem from the perspective of Category theory provides a simple and effective solution that completely avoids elaborate multi-stage training pipelines that would otherwise be needed. We show the effectiveness of this approach via extensive experiments on real datasets. Further, we discuss how this style of formulation offers a unified perspective on at least 5+ distinct problem settings, from self-supervised learning to matching problems in 3D reconstruction.

Large language models (LLMs) are typically aligned to a universal set of safety and usage principles intended for broad public acceptability. Yet, real-world applications of LLMs often take place within organizational ecosystems shaped by distinctive corporate policies, regulatory requirements, use cases, brand guidelines, and ethical commitments. This reality highlights the need for rigorous and comprehensive evaluation of LLMs with pluralistic alignment goals, an alignment paradigm that emphasizes adaptability to diverse user values and needs. In this work, we present PLURALISTIC BEHAVIOR SUITE (PBSUITE), a dynamic evaluation suite designed to systematically assess LLMs' capacity to adhere to pluralistic alignment specifications in multi-turn, interactive conversations. PBSUITE consists of (1) a diverse dataset of 300 realistic LLM behavioral policies, grounded in 30 industries; and (2) a dynamic evaluation framework for stress-testing model compliance with custom behavioral specifications under adversarial conditions. Using PBSUITE, We find that leading open- and closed-source LLMs maintain robust adherence to behavioral policies in single-turn settings (less than 4% failure rates), but their compliance weakens substantially in multi-turn adversarial interactions (up to 84% failure rates). These findings highlight that existing model alignment and safety moderation methods fall short in coherently enforcing pluralistic behavioral policies in real-world LLM interactions. Our work contributes both the dataset and analytical framework to support future research toward robust and context-aware pluralistic alignment techniques.

Large language models (LLMs) often exhibit sycophantic behaviors -- such as excessive agreement with or flattery of the user -- but it is unclear whether these behaviors arise from a single mechanism or multiple distinct processes. We decompose sycophancy into sycophantic agreement and sycophantic praise, contrasting both with genuine agreement. Using difference-in-means directions, activation additions, and subspace geometry across multiple models and datasets, we show that: (1) the three behaviors are encoded along distinct linear directions in latent space; (2) each behavior can be independently amplified or suppressed without affecting the others; and (3) their representational structure is consistent across model families and scales. These results suggest that sycophantic behaviors correspond to distinct, independently steerable representations.

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 alignment of large language models is usually done by model providers to add or control behaviors that are common or universally understood across use cases and contexts. In contrast, in this article, we present an approach and architecture that empowers application developers to tune a model to their particular values, social norms, laws and other regulations, and orchestrate between potentially conflicting requirements in context. We lay out three main components of such an Alignment Studio architecture: Framers, Instructors, and Auditors that work in concert to control the behavior of a language model. We illustrate this approach with a running example of aligning a company's internal-facing enterprise chatbot to its business conduct guidelines.