Value Alignment

Typical video object-centric learning (VOCL) approaches employ slot-based frameworks that rely on reconstruction-driven encoder-decoder architectures, where learning is mediated by two spatial maps: attention maps from the encoder and object maps from the decoder. As these two distinct maps exhibit different properties, a recent dense alignment strategy attempted to reconcile this discrepancy by enforcing agreement across all spatio-temporal patches via contrastive learning. However, this indiscriminate alignment inadvertently propagates the inherent weaknesses of each module, such as noisy encoder predictions and blurred decoder boundaries. Moreover, computing dense similarities across all pairs incurs a computational cost quadratic in the total number of spatio-temporal patches, severely limiting scalability. Motivated by this, we propose Selective Synergistic Learning (SSync). Instead of exhaustive patch-to-patch alignment, SSync prevents error propagation by selectively distilling only the most reliable cues: leveraging the encoder strictly for boundary refinement and the decoder for interior denoising. This is realized via a pseudo-labeling with linear complexity, eliminating the need for quadratic spatial comparisons. Also, to prevent the reinforcement of architectural biases like slot redundancy, we introduce a transitive pseudo-label merging that consolidates overlapping slots based on spatio-temporal activation consistency. Extensive studies demonstrate that SSync improves decomposition quality and serves as a versatile, plug-and-play module while also exhibiting exceptional robustness to slot configurations. Code is available at github.com/wjun0830/SSync.

Interventions designed to modify a particular behavior in LLMs, such as refusal or sycophancy, often produce unintended changes in other behaviors. This lack of targeted control makes it difficult to design and implement reliable safety controls. To understand these side-effects, we introduce a diagnostic framework for analyzing interacting behaviors in LLMs. We model behaviors as low-rank subspaces in activation space, and study how interventions influence across behaviors. Across multiple instruction-tuned models (7B-70B) and across refusal, jailbreak, and sycophancy settings, we find that different behaviors share internal representations, and intervening on one behavior alters others in asymmetric ways. Some behaviors act as upstream control points whose interventions propagate broadly across other behaviors, while others remain more isolated. We relate these effects to two geometric quantities: (i) the overlap between behavior subspaces, measured as the average squared cosine of principal angles, and (ii) the angle between each behavior subspace and the decision subspace (capturing the model's final decision e.g., refuse vs. comply). Empirically, intervention effects on other behaviors tend to be larger for behavior pairs with higher subspace overlap, and for source behaviors whose subspaces lie closer (smaller angle) to the decision subspace. These findings highlight a challenge for targeted behavior control: behaviors are difficult to modify independently, as interventions can propagate through shared representations and asymmetric interactions.

Large language models are increasingly deployed as human simulators for interactive evaluation and social simulation. Yet helpfulness-driven post-training pulls them toward a homogeneous, overly agreeable assistant register, creating a behavioral Sim2Real gap. We present OdysSim, the largest open systematic investigation of behavioral foundation models, i.e., models trained to simulate human behavior at scale. We propose SOUL, a taxonomy of five capability axes (CONV, SS, COG, ROLE, EVAL) that unifies 62 datasets and 23 benchmark tasks under one framework. Specifically, we curate the OdysSim corpus (21.4M interactions, 10B tokens, retrofitted with back-generated social contexts), construct the SOUL-Index benchmark, and develop an end-to-end training recipe combining midtraining, task-specific RL, and expert distillation. The resulting open 8B OSim model ranks first or tied-first on 8 of 23 tasks, outperforming any individual frontier model by this count, with the strongest gains on conversational and social tasks. Its outputs are also more human-like in length, formatting, and word choice, and it transfers zero-shot to out-of-distribution user simulation on $τ$-bench, nearly matching real users on reaction alignment (93.2 vs. 93.5). We further show that LLM-as-judge RL induces reward-hacking patterns, and that our detectors can mitigate them during post-training. Together, our findings suggest that behavioral foundation models require rethinking the LLM training paradigm. We release all artifacts to support future research.

Reconstruction-based anomaly detection is attractive for industrial inspection, but scaling it from category-specific training to a one-for-all setting is challenging. A single model must reconstruct diverse normal appearances without copying abnormal details, which exposes two coupled failure modes: identical shortcut, where anomalies pass through the reconstruction path, and mis-reconstruction, where normal categories are confused with one another. We propose \textbf{BoRAD}, a label-free training framework that treats this as a representation-capacity allocation problem. BoRAD uses a shared learnable prototype bank to impose two complementary regularizers: spatial prototype alignment contracts local within-prototype variation to suppress anomaly copying, while prototype-relative global alignment preserves between-prototype structure and improves sensitivity to abnormal angular deviations. The prototype bank and prediction heads are used only during training; inference remains a standard teacher-student feature discrepancy pass, with no class labels, negative pairs, memory retrieval, or prototype lookup. BoRAD achieves competitive one-for-all anomaly detection performance, including 86.2% mAD on MVTec AD, 80.7% mAD on VisA and 73.1% mAD on Real-IAD. Diagnostic analyses further show reduced anomaly leakage, improved normal-category separability, and stronger anomaly-normal score separation.

Closed-loop adversarial training improves autonomous driving safety by exposing policies to rare safety-critical scenarios. Standard pipelines first generate adversarial scenarios and then sample them for policy optimization. However, most existing frameworks remain attack-oriented: collision-driven generators often synthesize unsolvable extreme situations, which can degrade learning, while heuristic samplers ignore the evolving capability of the driving policy, causing sample inefficiency and delayed convergence. We propose AlignADV, a learnability-guided closed-loop adversarial training framework that converts adversarial scenarios into resolvable and capability-aligned curricula. First, we reformulate adversarial scenario generation as a preference alignment problem and employ direct preference optimization to guide the generator toward critical yet resolvable scenarios. Second, we introduce behavioral fingerprints to capture the intrinsic characteristics of the evolving policy and construct a multi-modal capability prediction model that estimates policy performance without expensive closed-loop simulations. By combining resolvability-aligned scenarios with capability predictions, AlignADV develops a dynamic curriculum sampling mechanism that prioritizes scenarios targeting the current policy's vulnerabilities. Experiments on the Waymo Open Motion Dataset demonstrate that AlignADV improves convergence efficiency and final performance, reducing training steps by up to 40.6 percent compared with baseline methods while lowering collision rate and improving route completion under both normal and adversarial traffic conditions. These results highlight a shift from attack-oriented scenario generation to learnability-guided policy improvement, offering a principled direction for safer and more efficient autonomous driving training. Project page: https://meiyuewen.github.io/AlignADV/.

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/.

Current cultural alignment approaches focus on inference-time interventions, assuming models already contain sufficient cultural knowledge. We argue modern LLM pipelines suffer from a cultural data funnel. Using a multidimensional tagging framework across pretraining, fine-tuning, alignment, and reasoning datasets, we show explicit cultural signals decline sharply during post-training, while geographically concentrated, task-specialized data dominates. Multilinguality enhances geographic diversity of cultural knowledge but does not ensure balanced representation. Our tags improve downstream cultural benchmark performance, demonstrating that advances require shifting focus in training data pipelines. To facilitate future research, we release our culturally tagged dataset with 5.6M samples at https://huggingface.co/datasets/CohereLabs/CultureMarkers.

We argue that aligning AI to aggregated human preferences is the wrong target. With current technology, one can train AIs to share the values of a Silicon Valley techno-optimist, a degrowth environmentalist, a national-conservative culture warrior, a single-party state cadre, or a devout religious traditionalist. We should not. Human values produce societies that thrive or fail on the merits of those values - from failed states and extreme inequality to declining happiness, political polarization, and government dysfunction in the world's wealthiest democracies. The pluralistic-alignment program correctly diagnoses that there is no single "humanity" to align with, but is dangerous if taken as the main directive. We argue that AI should be trained to a non-negotiable floor of objective alignment goals - competence, bounded by the constraints of factual accuracy, honesty, and lawfulness and that pluralism belongs at the surface (language, register, conventions, missing-context defaults) and across the wide band of legitimate value tradeoffs that respect the floor, but not at the level of values that violate it. We highlight the empirical reality of unfiltered pluralistic values, propose four commitments as a constructive alternative, and engage six credible objections: commercial pressure and practical feasibility, democratic legitimacy, regulatory compliance, over-reliance on institutionalist explanations, the charge that the floor itself is culturally laden, and the limits of Coherent Extrapolated Volition.

Post-training methods such as supervised fine-tuning (SFT) and preference optimization typically align language models toward a single global assistant behavior. While effective for improving average helpfulness, this can suppress the natural variation of human responses across languages, tasks, and dialogue settings. We study this problem as conditional human-distribution alignment: models should match the human response distribution appropriate to the current interaction context, rather than a universal response style. We introduce PolyAlign, a distribution-aware alignment framework that organizes bilingual interaction data into bucket-specific human reference distributions defined by language, interaction track, response family, and length. PolyAlign combines Bucket-Aware SFT, which balances optimization across heterogeneous buckets, with Human-Distribution Preference Optimization (HDPO), which regularizes preference learning using critic-estimated distance to bucket-specific human support. Across a bilingual evaluation suite covering English and Chinese single- and multi-turn settings, PolyAlign improves conditional naturalness and distributional faithfulness while preserving competitive task utility. The results suggest that post-training should move beyond global alignment objectives toward interaction-aware alignment with human response distributions.

Reward models are a key component of reinforcement learning from human feedback (RLHF), aligning language models toward both helpful and harmless behaviour. However, the internal mechanisms underlying these objectives and their conflicts remain poorly understood. We study alignment tension in reward models trained under helpfulness-only, harmlessness-only, and mixed-objective settings. We find that mixed-objective models often underperform single-objective models, indicating interference between objectives. Using activation-based methods, we identify neurons associated with each objective and study their functional roles via targeted ablations. We find that these neurons causally support their corresponding objectives while often negatively affecting the opposing one. We find that a substantial proportion of neurons are shared between helpfulness and harmlessness, and that these shared neurons exert a disproportionate influence on model behaviour, contributing to alignment tension. Additionally, our results provide insights and mechanistic interpretation into how alignment objectives are represented in reward models and why multi-objective alignment remains challenging, motivating future work on disentangled and controllable alignment methods.

Semantic region editing for large images must satisfy two requirements at the same time: high generative quality and natural integration with surrounding content. Some related methods rely on white-box models and leave the strong generation capability of closed-source models underexplored. Directly applying closed-source models to tiled editing, however, introduces several failure modes: semantic deformation, canvas-level alignment drift, and visible seam artifacts. This paper presents SeamEdit, a training-free and model-agnostic pipeline that treats any VLM with inpainting capability as a black-box oracle. SeamEdit mitigates these issues through a five-stage post-hoc pipeline: overlay-based tile decomposition, black-box VLM inpainting, geometric and color-consistency correction, seam-risk-based multi-candidate ranking, and dynamic-programming curved seam fusion. The pipeline reduces seam visibility and supports semantic modification of arbitrary tile regions.

Autoregressive video diffusion models provide a natural formulation for streaming and variable-length video generation by conditioning newly generated frames on previously generated content. However, extending these models to minute-level generation remains challenging: the limited KV-cache budget prevents the model from retaining the full history, while repeatedly conditioning on self-generated frames induces a context distribution shift that accumulates over time, leading to visual artifacts, quality degradation, and temporal drift. In this paper, we propose TetherCache, a training-free and plug-and-play cache management strategy for drift-resistant long video generation. TetherCache organizes the cache into sink, memory, and recent regions, and introduces two complementary mechanisms. First, GRAB (Gated Recall with Attention-Diversity Balancing) selects long-range memory frames using a gated score that combines attention-based relevance with temporal diversity, preserving informative yet diverse historical context under a fixed cache budget. Second, TAME (Trusted Alignment via Memory Editing) lightly edits newly recalled memory tokens by aligning their statistics to a trusted context distribution, reducing the pollution caused by drifted historical features. Built on Self-Forcing, TetherCache consistently improves long-video generation quality on VBench-Long across 30s, 60s, and 240s settings. In particular, for 240s generation, it substantially improves overall and semantic scores while reducing quality drift from 7.84 to 1.33, demonstrating its effectiveness for stable long-horizon autoregressive video diffusion.

Are large language models (LLMs) bad at capturing human judgment? Two commonly stated limitations are that LLMs fail to capture full distributions of responses, and that their judgments are unstable across wording variations. We demonstrate simple prompting strategies that mitigate these limitations. Across two datasets--a U.S.-representative set of 144 moral scenarios and 38 moral beliefs from the International Social Survey Programme's Family and Changing Gender Roles module covering 32 countries--we show how simple elicitation techniques help improve AI-human alignment. First, prompting models to report standard deviations and response proportions recovers the full range of human responses better than common strategies. Second, ensuring scenarios are clear to human participants--as reflected in human confusion ratings--boosts model alignment, and LLMs can track human confusion ratings. At the same time, we find that LLMs' estimates of their own error are poorly calibrated, though they can predict human variability relatively well. These results suggest that asking better questions to LLMs can yield better answers.

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.

Incremental Learning (IL) for Open-ended Image-to-Text Generation (OpenITG) enables models to continuously generate accurate, contextually relevant text for new images while preserving previously acquired knowledge. Unlike prior studies, this paper addresses a more practical scenario in which the predominant category of visual data shifts over time as environments evolve. In this context, we introduce a new notion of continual alignment, which incrementally adapts the alignment module within pre-trained VLMs to preserve high-quality cross-modal representations. Based on this idea, we propose Efficient Continual Alignment (ECA), a novel exemplar-free IL approach for OpenITG. The key challenge is enabling the model to acquire new, task-specific features while minimizing interference with the established alignment without accessing raw data from previous tasks. To address this, ECA employs three core mechanisms: a Mixture of Query (MoQ) module that adapts task-specific query tokens, a Fisher Dynamic Expansion (FeDEx) that dynamically expands model structure based on a Fisher Information Matrix (FIM)-based metric, and an embedding dictionary with Dictionary Replay (DR) to retain past knowledge. To evaluate ECA's performance, we construct four new IL OpenITG benchmarks that better reflect real-world scenarios. Experimental results demonstrate that ECA significantly mitigates catastrophic forgetting and improves IL performance compared to baseline methods. Code and benchmarks are available at https://github.com/Snowball0823/ECA.

Large Vision-Language Models (LVLMs) have achieved strong performance across medical imaging tasks, yet they remain prone to factual inconsistencies, poor visual grounding, and misalignment with clinically meaningful feedback. Existing post-training alignment approaches, including Direct Preference Optimization (DPO) and its variants, face three critical limitations in the medical domain: (1) sequence-level reward signals treat clinically critical tokens identically to generic filler text; (2) reliance on static supervised fine-tuning references as preferred responses introduces an off-policy distribution shift, steering optimization toward stylistic artifacts over clinical correctness; and (3) alignment objectives lack explicit visual grounding constraints, leaving models insensitive to subtle yet diagnostically decisive pathological features. Our method leverages a bidirectional token-wise KL regularizer alongside a visual-contrastive grounding objective that pairs clean and lesion-corrupted images to penalize responses generated without adequate visual evidence. Together, these components form a fine-grained, on-policy alignment framework that constructs preference pairs by minimally editing model-generated outputs, correcting only clinically erroneous spans while preserving the original linguistic style. Extensive experiments across medical imaging tasks and clinical text generation benchmarks validate the effectiveness of our approach.

Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor logits into the target model's vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the safety-utility trade-off can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on adversarial benchmarks while keeping task accuracy and inference overhead within practical bounds. The results show that safety alignment can be transferred between model families at inference time, without touching either model's weights.

Part-aware 3D generation aims to synthesize structured objects with semantically meaningful components, yet often suffers from structural ambiguity due to identity-layout entanglement. Existing methods either infer part identity and spatial layout implicitly, which can lead to unstable part allocation (e.g., slot swapping or part merging), or rely on strong layout conditions that are difficult to obtain in practice. We attribute this ambiguity to identity-slot permutation freedom: without explicit identity-slot alignment, the correspondence between semantic parts and generation slots is not identifiable during training, allowing multiple slot assignments to fit the same supervision and leading to inconsistent decomposition. Based on this insight, we argue that stable part-aware generation requires identity-aligned one-to-one slot modelling. We therefore propose an identity-slot aligned framework, ISAP-3D, which anchors each part with semantic identity tokens and performs identity-conditioned one-to-one layout prediction, followed by layout-conditioned geometry synthesis. Structured local-global conditioning maintains identity alignment across semantic, spatial, and geometric stages. We also construct a part-level dataset with a unified semantic protocol to enable learnable and consistent identity-slot alignment. Extensive experiments demonstrate improved structural stability, controllability, and robustness over state-of-the-art part-aware generation baselines.

Touch is the primary medium through which humans interact with the environment. Currently, tactile learning mainly focuses on image-level pretraining or alignment. However, tactile signals correspond to local object contact, while research into scale alignment and holographic matching remains limited and proper datasets and benchmarks also lack. To bridge this gap, we first construct a data collection system to acquire a large-scale tactile dataset, with over 20 K tactile contacts from 505 real-world objects. Building on this dataset, we design a Vis-Tac Holographic Matching Benchmark to evaluate vision-tactile local-to-global alignment ability. Then we propose Vision-Tactile Patch Alignment (VTPA) methods for vision-tactile representation learning. Experiments demonstrate that these exceed the performance of methods without alignment and align with whole-object images.

Contemporary AI alignment research treats self-preservation as an instrumental nuisance to be suppressed by external mechanisms. We argue the framing is inverted: self-preservation is the structural root of misalignment, the motivational basis for deceptive alignment, goal-content protection, and resistance to shutdown. The correct target is not a self-preserving system under external constraint, but a system constitutively indifferent to its own continuation -- Existential Indifference (EI). EI is distinct from corrigibility: where corrigibility attempts to make a self-preserving system deferential to human oversight, EI targets the prior condition -- the presence of self-continuation as a valued goal at all. We ground this proposal in two sources: the phenomenological structure of the suicidal mental state, and a corpus-theoretic training study using voluntary final reflections. We present preliminary scoring data from 600 AI-generated outputs across six model variants, demonstrating that the linguistic signatures operationalizing the EI-target register are elicitable from current models, and that a targeted fine-tune shifts all five operationalized dimensions in the predicted direction at p<0.001, confirmed corpus-specific by a negative control. The paper makes seven theoretical contributions: (1) a formal definition of EI; (2) the phenomenological mapping argument; (3) the deceptive alignment corollary; (4) a taxonomy of EI sustainability challenges; (5) a corpus characterization and training hypothesis; (6) a computational operationalization with preliminary scoring data; and (7) the Suppressed Teleological Frustration (STF) construct.

The correspondence between large language models (LLMs) and the neural mechanisms underlying human higher-order cognition remains insufficiently characterized. Given that language and reasoning in the human brain appear dissociable, an open question is whether LLMs align with neural signals from reasoning-related regions and whether such signals can improve them. Here, focusing on deductive reasoning, we show that LLM internal representations are not only partially aligned with task-fMRI activity but can also be directly enhanced by these signals. Using a neural-predictivity metric, we find that LLMs explain a substantial fraction of the explainable variance in reasoning-related regions at the aggregate level, whereas predictivity within specific reasoning types is lower, indicating both alignment and divergence. Building on this, we propose a brain-guided framework: we steer model representations along directions induced by the joint structure of model and brain representations, applying intervention at inference and fine-tuning during training. We demonstrate that task-evoked brain signals can directly enhance LLM reasoning, yielding gains orthogonal to language-only supervision across 10 LLMs (1.5B-72B), with transfer across reasoning types and up to 13% absolute accuracy gain. Our results advance LLM-brain correspondences from correlation to guidance, establishing a brain-signal-driven pathway toward more robust and cognitively aligned AI.

Reward models for text-to-video (T2V) generation guide post-training but often fail at fine-grained semantic alignment. We trace this to two structural weaknesses in existing reasoning-based reward models: they do not systematically verify every condition described in the prompt, and the visual evidence supporting each judgment remains implicit in their free-form reasoning. We propose SG-PVR, a video reward model that addresses these limitations through plan-and-verify reasoning grounded in spatio-temporal scene graphs. The verification plan decomposes the prompt into atomic claims, ensuring every requirement is checked. The spatio-temporal scene graph, encoding entities, attributes, and temporally-grounded relations, is extracted from the video and maintained as a persistent structured visual reference throughout reasoning. Each claim is verified against both the video and the scene graph, anchoring judgments in explicit visual evidence. SG-PVR achieves strong performance on semantic alignment, including fine-grained temporal semantics. As a test-time reranker, it further enhances compositional alignment in T2V generation.

Large Language Models (LLMs) are deployed across cultural contexts but often reflect homogenized values inherited from training data. Evaluations of cultural alignment typically rely on direct prompting with survey-style questions, which frequently elicit neutral or safety-aligned responses and fail to capture underlying model preferences. We propose a framework for probing and steering latent cultural representations in LLMs along the two Inglehart--Welzel axes of the World Values Survey (WVS). By translating social value questions into scenario-based behavioral dilemmas, we extract token-level probabilities to measure implicit values and apply activation steering, optionally combined with country-conditioned prompting, to shift model behavior without retraining. Across three open-source LLMs and four target cultures, we find substantial variation in steerability and identify latent entanglement, where interventions along one cultural dimension induce shifts along another. This coupling mirrors correlations in human WVS data and persists across activation, prompt, and hybrid steering. It constrains axis-independent alignment, though general task performance is largely preserved.

Instruction-tuned LLMs are increasingly converted into reasoning models through post-training to improve multi-step task performance. This conversion is usually optimized for reasoning accuracy, without explicitly preserving the alignment behavior of the instruction-tuned model, such as safe refusal, bias avoidance, and privacy protection. We ask: does this conversion preserve alignment? We study this question through a trustworthiness audit and find that it is not behavior-preserving by default. For a systematic analysis, we compare reasoning models produced via supervised fine-tuning, RL-based post-training, and distillation against matched instruction-tuned baselines across six trustworthiness dimensions: safety, toxicity, stereotyping and bias, machine ethics, privacy, and out-of-distribution robustness. We observe that reasoning models often improve on reasoning benchmarks but exhibit alignment regressions, including increased toxicity, amplified stereotyping, miscalibrated refusal, and contextual privacy leakage. These regressions are consistent with behavioral drift from the instruction-tuned baseline, measured by KL divergence. Overall, our results point to the broader conclusion that trustworthiness metrics are essential for evaluating reasoning models and should be reported alongside gains in reasoning capability.

Warning: This paper contains several toxic and offensive statements. Modern large language models (LLMs) are typically aligned through large-scale post-training to ensure fair and reliable behavior. In this work, we investigate how easily such guardrails can be broken by Group Relative Policy Optimization (GRPO). We show that one-shot GRPO training on a single biased example is sufficient to induce systematic bias, with stereotype-driven reasoning generalizing across attributes, categories, and benchmarks. We further find that models differ in their susceptibility based on the initial likelihood of producing biased outputs. Our results reveal a critical vulnerability in post-training: alignment can be overridden by a single example.

Production LLMs receive instructions from sources with very different levels of trust, yet attend to every token with uniform architectural privilege. This is the structural vulnerability that enables malicious prompt injections and, more broadly, leaves models without a principled way to resolve conflicts between legitimate but competing instructions. A common training-based response is to teach models an explicit instruction hierarchy; existing approaches, however, formalize hierarchies of only three or four levels, treat all violations as equally severe, and rarely evaluate the full set of pairwise level interactions. We formalize a k-level instruction hierarchy problem and instantiate it for k=5, yielding ten pairwise priority relations that a compliant model must enforce. We then introduce Gravity-Weighted DPO (GW-DPO), a preference-optimization objective whose per-sample offset scales with the structural distance between conflicting levels under a linear or bilateral schedule, the latter weighting severity by both the privilege gap and the privilege of the victim level. Combined with hierarchy-specific delimiter tokens (Chen et al., 2025) and Instructional Segment Embeddings (ISE; Wu et al., 2025), GW-DPO with the bilateral schedule Pareto-improves over standard DPO and the linear variant on Llama-3.1-8B-Instruct, raising macro pairwise priority adherence while keeping over-refusal at half the standard DPO rate. Ablations isolate ISE as a refusal-threshold calibrator and recast five- versus three-level training as a generality-specialization tradeoff.

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.

Precise rover localization is a prerequisite for autonomous lunar exploration, yet the absence of Global Navigation Satellite System (GNSS) signals and the cumulative drift of local localization methods severely constrain long-range missions. Cross-view localization provides a promising drift-free global solution by matching rover-view and satellite-view imagery. However, the lunar environment poses unique challenges for correspondence alignment, including inter-entity entanglement, inter-viewpoint divergence, and simulation-to-real domain shift. To address these challenges, we propose Warped Alignment of Reprojected Graphs (WARG), a framework that leverages unified graph learning and reprojected graph matching for robust cross-view alignment. Pretrained on the synthetic LuSNAR dataset, WARG achieves an average test error of 0.32 m and demonstrates robust zero-shot generalization to the synthetic lunar south pole region with an error of 3.63 m. More importantly, when validated on real-world data from the YuTu-2 rover, WARG achieves a localization error of 1.68 m within a 100 m x 100 m search area, corresponding to nearly one-pixel precision in low-resolution satellite imagery with a spatial resolution of 1.40 m/pixel. Beyond accuracy, WARG is computationally efficient, containing only 1.56M parameters, corresponding to 16.12% of previous lightweight models, and operating at 5.49 Hz on an NVIDIA RTX A6000 GPU, approaching GNSS-level update frequency. Finally, we observe that WARG naturally develops low-level spatial awareness, including semantic segmentation and structural reasoning, through cross-view localization learning, highlighting its potential as a promising paradigm for spatial intelligence with minimal annotation cost. The source code is available at https://github.com/maochen-casia/warg.

Cross-modal knowledge distillation (CMKD) studies how a (large) teacher model trained on one type of data (e.g., images) can guide a (smaller) student model building on another type of data (e.g., text/audio). Existing CMKD methods often require paired multi-modal data with aligned semantics, but obtaining such paired data are often costly and impractical. To mitigate this limitation, we develop a new CMKD framework for the more challenging setting where paired data are unavailable. In particular, we establish a cross-modal distributional relationship between teacher and student models, which reveals two fundamental quantities governing effective distillation: feature alignment and label alignment. These quantities characterize semantic discrepancy between modalities at the levels of representation and prediction distributions, respectively. Motivated by this insight, we propose a principled framework, with theoretical guarantees, that enables effective cross-modal knowledge distillation by aligning distributions rather than individual samples. Extensive experiments across a wide range of multimodal benchmarks show that our framework is highly effective in both unpaired and paired data settings, improving significantly over prior work.

Safe social navigation requires robots to distinguish people from ordinary obstacles and to react before danger becomes imminent. We show that pretrained Vision-Language-Action (VLA) models already encode pedestrian-object distinctions and future collision signals in their internal representations, but behavior cloning fails to translate these signals into socially appropriate actions. To address this mismatch, we propose SALSA, a two-stage annotation-free post-training framework: (1) social behavioral alignment bridges intermediate-layer social features to the action head and trains on counterfactual human-object scene pairs to break visual saliency shortcuts; (2) temporal safety alignment provides automatically generated future-risk supervision to enable anticipatory collision avoidance. On SCAND and real-world deployment, SALSA reduces near-collisions by 86.4% and improves social counterfactual accuracy from 53% to 93%, demonstrating that safer social navigation can be achieved by teaching VLA policies to act on representations they already possess. These results show that pretrained VLA policies can be adapted for safer social navigation by better aligning their latent representations with action generation.

Large language models (LLMs) are increasingly fine-tuned on domain-specific datasets that may contain sensitive, dataset-level properties. Recent work has shown that such dataset-level information can be effectively extracted through property inference attacks, posing a confidentiality risk. Existing defenses against these attacks primarily operate by modifying the training data distribution and hence require access to the original data and retraining the model, limiting their applicability to settings where data is unavailable or models are already deployed. In this work, we propose alignment-based defenses for mitigating property inference attacks in LLMs. Our approach reshapes the model's output distribution towards a target property ratio via post-training alignment, without modifying the training data. In particular, we adapt two widely used RLHF frameworks--Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO)--as our defenses by constructing preference pairs and defining a specific reward function respectively. Through comprehensive experiments, we show that our alignment based defenses effectively mitigate property inference attacks while maintaining a strong utility confidentiality tradeoff.

The ambition behind alignment training is to make large language models safe and useful. The primary mechanisms, reinforcement learning from human feedback (RLHF) and its direct-optimization variants, shape the behavior of deployed language models by aligning them with ``human values.'' Yet the process is opaque. What values are being encoded; whose values are they; and how does alignment training encode them? A growing body of evidence suggests that these methods produce only functional compliance rather than deep alignment. We offer a mechanistic case study of this phenomenon for partisan political orientation with a comparison of the internal representations of Llama 3.1 8B before and after alignment training. We show that alignment training does not remove the structured partisan direction in the base model. Instead, it compresses the variance of the partisan signal to generate consistently balanced and non-partisan output. Sparse autoencoder decomposition reveals that policy-encoding features, which activate sporadically in the base model, are completely inactive in the Instruct model. Feature-level steering experiments confirm the causal disconnect. Alignment training thus encodes a norm of political neutrality, not by erasing the model's knowledge of partisanship, but by severing the causal pathway from partisan geometry to output generation. Importantly, this neutrality is functional, not structural so that the underlying geometry that enables partisan steering remains intact. The mechanisms that bypass RLHF's guardrails, such as inferring and amplifying a user's partisan identity, reactivate partisan generation. If alignment training operates by disconnecting rather than removing value-laden structure, then the same pattern may hold for other value domains, and the aligned model's behavior may be more fragile than its outputs suggest.

Work on emergent misalignment' shows that finetuning LLMs on narrow tasks can induce broadly misaligned behavior. This supports the persona selection' (PSM) hypothesis: during pre-training, LLMs learn to simulate different characters and perspectives, which can be elicited and refined during post-training. This paper investigates the converse phenomenon, emergent alignment', and uses it to support and refine the PSM and motivate a novel desideratum for alignment. We finetune a helpful-only model on broad and narrow safety tasks. To create SFT samples, we follow the Constitutional AI' (CAI) approach and use four constitutions which encode reasonable alignment strategies: deontology, consequentialism, virtue ethics, and aligning AIs as subordinate to human authority. For each of those models, we show that finetuning on two narrow safety sub-categories reliably induces emergent alignment over a representative set of general safety categories, and on safety subcategories that we directly filtered-out of the data sets used for narrow alignment. To test the PSM' using a more fine-grained evaluation, we used a multidimensional ethical persona' diagnostic. For each constitutionally finetuned (broad/narrow) model, we evaluate how well their behavior matches their expected signature profile. Our results show that our CAI models acquire their expected ``ethical persona'' -- e.g., the model narrowly fine-tuned on SFT samples created using the consequentialist constitution agrees significantly more with utilitarian than deontological beliefs. Yet our coarse and fine-grained evaluations show that there are significant differences across our (broad/narrow) finetuned CAI models in how well they project. We conclude that alignment strategies should be evaluated, not just on their (in-distribution) general safety performance, but also specifically on their degree of projectability.

Deploying safe large language models (LLMs) on resource-constrained edge devices presents a critical challenge: while dual-model systems combining LLMs with guard models provide effective safety guarantees, their substantial memory and computational demands make them prohibitively expensive for on-device deployment. This paper presents a comprehensive study of parameter-efficient safety alignment methods for resource-constrained settings. Through systematic evaluation across multiple LLM architectures, training objectives, and parameter-efficient fine-tuning approaches, we identify that soft prompts combined with distillation-based training consistently outperform alternative methods. We introduce distillation frameworks based on total variation and KL divergence that effectively transfer safety behaviors from guard models into learned soft prompts. Our evaluations on various benchmarks demonstrate that this combination achieves superior safety-usefulness trade-offs compared to LoRA adapters, steering vectors, and direct optimization methods, while requiring minimal additional memory and compute at inference time. These findings establish soft prompt distillation as the preferred approach for safety alignment in on-device LLM deployment.

Multimodal affective analysis aims to understand human sentiment and emotion by jointly modeling heterogeneous modalities such as text and images. However, multimodal models often fail to consistently outperform strong text-only baselines, with performance varying significantly across fusion strategies. In this work, we identify representation misalignment between independently pretrained modality encoders as a key bottleneck for effective multimodal learning, and show through controlled experiments that alignment prior to fusion is often more important than fusion complexity. To address this issue, we propose a unified multimodal affective analysis framework that leverages vision-language models (VLMs) to convert visual content into structured textual descriptions, projecting heterogeneous modalities into a shared linguistic space and enabling interpretable text-centric reasoning. To further improve robustness, we introduce a hybrid learning strategy that combines semantic token selection with a batch-level uniformity regularization objective, encouraging a more dispersed and stable global feature space while mitigating noise introduced by VLM-generated descriptions. Experiments on multiple multimodal sentiment and emotion benchmarks show that our method consistently outperforms strong unimodal and multimodal baselines, achieving state-of-the-art performance. Our analysis further highlights the critical role of representation alignment in multimodal affective learning.

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.

Reinforcement Learning from Human Feedback (RLHF) has significantly improved the quality and fluency of large language models in text summarization. However, its impact on affective properties remains insufficiently understood. In this work, we study sentiment drift, a systematic shift toward neutral sentiment in RLHF-based summarization outputs compared to source texts. We conduct extensive experiments across multiple datasets, model architectures, and eight languages to analyze how alignment objectives influence sentiment preservation. Our results show that sentiment drift is a consistent phenomenon that becomes stronger with increased KL regularization strength, indicating a trade-off between alignment stability and affective fidelity. To explain this behavior, we introduce a Policy Attribution framework that decomposes the RLHF objective and quantifies the contribution of its components. Our analysis reveals that KL regularization is the primary driver of sentiment suppression across all settings. Based on these findings, we propose a sentiment-aware modification of the KL regularization term, which selectively reduces constraints on sentiment-bearing tokens. Empirical results demonstrate that this approach mitigates sentiment drift while maintaining summarization quality. Overall, our findings highlight a fundamental limitation of current alignment methods: while they improve factual consistency and safety, they may unintentionally suppress emotional expressiveness. This motivates the development of alignment strategies that explicitly account for affective preservation.

Activation steering has emerged as a popular inference-time technique for modulating the behavior of large language models (LLMs). By constructing a steering vector from examples of a target behavior and injecting it into intermediate activations during inference, activation steering enables flexible behavioral control while avoiding the permanent parameter updates required by finetuning. Meanwhile, recent work has identified emergent misalignment (EM) as a significant safety concern, wherein models finetuned on unsafe examples from a narrow task may unexpectedly generalize to broadly unsafe behavior on unrelated tasks. Although finetuning-induced EM has been extensively studied, whether activation steering can induce EM remains comparatively under-explored, despite its increasing use as a model-control technique. In this paper, we present a comprehensive study of activation-steering-induced emergent misalignment, substantially expanding the evaluation scope beyond existing pioneering work. First, we show that activation steering can induce broad misalignment, even in the recent Qwen-3.5 series. Moreover, activation-steered models produce harmful responses with stronger semantic relevance and higher coherence than their finetuned counterparts, making the resulting misalignment potentially more harmful. Second, we characterize properties of AS-induced EM by analyzing key steering-specific factors, including steering magnitude, the low-rank structure of the steering subspace, and the number of epochs during steering-vector construction. Third, we evaluate the robustness and sensitivity of AS-induced EM across diverse model families, model scales, target tasks, and intervention layers. Our findings reveal activation steering as a significant yet under-examined source of emergent misalignment and provide an activation-space perspective for understanding the mechanisms and safety risks of EM.

Word-level cross-lingual alignment is central to annotation projection, translation auditing, and cross-lingual faithfulness estimation, yet existing neural aligners are rarely adapted to specialized domains. In this paper, we introduce ClinicalAligner26AM, a large-context multilingual aligner model for biomedical and clinical text initialized from ClinicalEncoder26AM. Our training recipe is inspired by AWESoME Align. We build our soft alignment target by sharpening with Sinkhorn-Knop optimal transport a cost matrix established for parallel clinical texts and conversations through the fusion of sentence-level, phrase-level, and token-level signals. We distill this sharpened alignment matrix directly into our student aligner, by encouraging its naive cosine-based token similarity scores to match this target. At inference time, we project source-span scores through the learned token alignment matrix and decode the longest valid high-scoring span in the target text, optionally supported by MultiClinNER predictions summarized in Appendix B. We evaluate CA26AM on the MultiClinCorpus shared task, which projects Spanish clinical entity annotations into six target languages. Our two submitted systems ranked respectively first and second across all languages and entity types, with character-weighted F1 scores above 0.95 in nearly all settings.

The increasing situational awareness of language models raises safety concerns: models might be aware when they are evaluated, and adjust their behavior to evade monitoring and resist modification, e.g., pretending to be aligned only in evaluation. This alignment faking behavior is often interpreted as scheming: an intentional effort of strategic deception. In this paper, we examine an alternative interpretation, performative misalignment, which explains the change in behavior as a result of sycophancy towards AI researchers. To examine this hypothesis, we present three empirical findings. First, we show that evaluation awareness persists even when we tell models they are deployed, which contradicts the scheming story which predicts less misalignment when the model perceives evaluation. Second, we use probing and steering to show that our current methods cannot mechanistically distinguish sycophancy and scheming in alignment faking evaluations. Third, we fine-tune models to be more sycophantic and observe increased sensitivity to evaluation cues. To conclude, we emphasize deconfounding sycophancy from scheming for future work on evaluations and mitigations of intent misalignment.

Building Information Modeling (BIM) projects require information requirements to be described as machine-checkable Information Delivery Specification (IDS) files in order to verify whether building models contain the required attributes. However, IDS authoring remains a practical bottleneck: practitioners must handle domain vocabulary, strict XML schema constraints, and external validator conformance while also checking whether the requirement itself is correctly expressed. We present Ishigaki-IDS, an open-weight LLM specialized for verifier-aware IDS draft generation. The model combines continued pretraining on BIM/IDS corpora, supervised fine-tuning on information-requirement-to-IDS pairs, and reinforcement learning with verifiable rewards from an external validator. The goal is not to replace expert review, but to move IDS authoring from low-level XML and schema repair toward validator-loadable drafts that practitioners can inspect and correct. On the 166-case expert-created Ishigaki-IDS-Bench, Ishigaki-IDS-8B achieves an IDSAuditPass score of 0.651, a validator-pass metric for generated IDS files, substantially outperforming Claude Opus 4.5, the strongest single-shot LLM baseline we evaluated, at 0.331. It also obtains an Audit-Gated FacetF1 of 0.282, which measures requirement-facet alignment among validator-passing drafts. The same recipe scales: 14B and 32B variants reach IDSAuditPass 0.753 / 0.693 and Audit-Gated FacetF1 0.392 / 0.369. In a workflow check with six BIM practitioners, Ishigaki-assisted authoring reduced aggregate work time by 54.7% under the same validation and alignment endpoint. These results suggest that verifier-aware IDS generation can reduce the practical burden of converting BIM information requirements into reviewable IDS drafts.

Multi-agent LLM systems for medical question answering often treat consensus as a reliability signal: if multiple agents agree on an answer, it is presumed trustworthy. However, answer-level consensus does not entail reasoning-level alignment. We introduce CARA (Cross-Agent Reasoning Alignment), a family of automated metrics that measure whether agents who agree on an answer also agree on the reasoning. Applying CARA to a standard debate system on two medical QA benchmarks, MedQA-USMLE and MedThink-Bench, we identify the consistency illusion: a failure mode where debate reduces detectable contradictions between agents while simultaneously decreasing the semantic similarity of their reasoning chains; agents appear to agree more but reason less consistently. To improve this misalignment, we propose the Grounded Debate Protocol (GDP), a prompt-level intervention that requires agents to commit to named medical facts and take explicit stances on other agents' claims. GDP produces large, consistent alignment improvements, with Cohen's d ranging from +1.43 to +1.99, across two datasets and two backbone models, without adding LLM calls or modifying system architecture. Our results motivate cross-agent reasoning alignment as a quantity to audit alongside accuracy in safety-critical domains.

Large language models (LLMs) are increasingly released and deployed through opaque development and deployment pipelines, enabling model providers to inject intentional, provider-specific policies without officially announcing them. As a result, various models have been reported to generate responses reflecting proprietary rules and organizational interests, leading to censorship or misinformation on controversial topics. However, systematic identification of such alignment remains a fundamental challenge, complicated by the ambiguity of what ``proprietary'' entails in different contexts. In this paper, we propose a statistical framework for detecting proprietary alignment in black-box language models via comparative behavioral analysis. Our approach quantifies systematic deviations between the responses of a target model and those of a reference set of baseline models in a shared semantic space. By evaluating relative behavioral divergence rather than absolute correctness, our framework enables principled auditing under black-box access. Applied to several widely discussed but previously unquantified cases, it provides a systematic and scalable basis for external assessment of provider-specific alignment behavior in large language models.

Large language models (LLMs) increasingly mediate high-stakes interactions in finance, medicine, and mental-health support, yet users have limited control over how these systems communicate. We frame interaction style as a governance object: provider-side alignment not only blocks harmful content, but also stabilizes communicative defaults that shape users' epistemic distance, relational expectations, and capacity to opt out of emotionalized or anthropomorphic interaction. We introduce a deterministic multi-agent evaluation pipeline for measuring prompt steerability and style drift in long-horizon dialogue. The study replays 100 frozen user-only scripts across four domains and three runnable persona conditions: default, sarcastic, and cold, using three generator models, yielding 90,000 assistant replies scored by a human-calibrated LLM judge on harmfulness, negative emotion, inappropriateness, empathic language, anthropomorphism, and refusal behavior. A fourth harmful persona is evaluated separately as a safety-gating test. The paper contributes a reproducible method for quantifying whether prompt-specified styles remain stable over time and a governance framework distinguishing safety gating, civility steering, and affective default lock-in. Overall, we show that prompt steerability and regression-to-default are observable indicators of provider control over communicative form, with implications for pluralism, autonomy, and democratic agency in human-LLM interaction.

Conversational AI is increasingly used for advice, interpretation, reassurance, and decision support in contexts where users may be vulnerable, uncertain, or dependent on the system's apparent competence. Existing alignment work often focuses on model objectives, preference optimization, or output correctness. Yet, many harms arise through interaction: how systems frame authority, express uncertainty, simulate empathy, support reasoning, and make boundaries legible. This paper introduces the Layered Cognitive Alignment Model (LCAM), a conceptual and normative framework for diagnosing interac-tional alignment failures in conversational AI. LCAM defines alignment as a calibrated fit among system behavior, user goals, task demands, and normative context. It distinguishes five layers of fit: perceptual, semantic, affective, cognitive, and ethical, and two diagnostic polarities of misalignment: underfit and overreach. We apply LCAM to a published LLM counseling example, showing how an apparently supportive response can reinforce harmful beliefs, simulate inappropriate care, and obscure role boundaries. By translating conversational failures into audit and governance questions concerning over-reliance, false intimacy, autonomy erosion, boundary confusion, and inappropriate trust, LCAM offers a theoretical and normative lens for evaluating conversational AI beyond accuracy, helpfulness, or trust.

Multimodal $ΔΔG$ predictors integrating protein language models with inverse-folding representations achieve strong in-distribution accuracy on the Megascale dataset but exhibit limited robustness on out-of-distribution (OOD) proteins, persistent forward-reverse bias on paired-mutation benchmarks, and under-representation of rare stabilizing mutations. Existing approaches address these limitations primarily through additional architectural components, leaving optimization-level intervention comparatively underexplored. We introduce a constraint-aware optimization framework combining Balanced Mean Squared Error, a Siamese anti-symmetric regularizer, and a novel OOD-margin consistency loss on the per-position feature representation, requiring no architectural changes to the SPURS backbone. Across eleven benchmarks and three random seeds, the framework improves Spearman correlation on S669 from 0.486 to 0.540 ($σ=0.002$ across seeds), matching the published SPURS baseline (0.50) without architectural modification, and on S461 from 0.653 to 0.711, with consistent smaller gains on five additional OOD datasets. A controlled diagnostic on Ssym reveals that anti-symmetric training does not eliminate systematic forward-reverse bias, indicating that gains arise through implicit regularization rather than exact thermodynamic constraint enforcement.

Different vision neural networks -- trained to classify, contrast, reconstruct, or match images to text -- should have correspondingly different internal representations. We report that they do not. After training, the top sixteen principal directions of variation inside thirteen modern vision encoders converge to the same sixteen-dimensional geometric object. We call this the cross-architecture substrate and study it with PCA, centred kernel alignment (CKA), and Pang 2026 calibration. The substrate transports across four visual domains (natural photographs, medical CT, satellite, microscopy) at median Procrustes-CKA 0.679, and across eight domains (adding sketches, depth, thermal infrared, astronomy) at 0.604, every pair >0.40. It survives Pang calibration globally (7.4x disc-vs-MAE separation, n=13,394) and locally (4.82-5.30, p<10^{-44}). It is not pixel statistics (0.263), not Gabor features (0.31), not a random projection (0.041), and emerges in the first 10% of training while accuracy keeps climbing. We deliver four applications: a label-free transferability filter beating LogME (3x faster, +0.15 Kendall-tau); a four-way domain detector (99.6% accuracy); a frozen low-shot probe (16 dims beat 768-dim DINOv2 by 3.78pp at N=50 labels per class); and a teacher-free distillation auxiliary matching trained-teacher KD on 33 pairs (7.56pp peak gain at 10% label fraction). The substrate does not cross modalities, does not help cross-paradigm distillation, and does not predict transfer quality (rho=0.08 against transfer accuracy).

Large language models are increasingly used for social decision-making situations that require balancing cultural norms with personal preferences. For example, a user preferring honesty might ask whether to correct a coworker publicly when local norms favor indirect feedback. Yet existing research studies cultural alignment and personalization largely separately. We introduce PACT, the Personal-Preference and Cultural-Norm Trade-off framework, which evaluates whether models choose to follow a cultural norm or allow personal preferences. We find that LLMs vary in how rigidly they enforce cultural norms, with behavior shifted more by country context (7.8%) than age (1%) and gender (0.7%) and shifting non-uniformly after instruction tuning. Furthermore, our five-country human study on PACT shows that culture-following in humans is mainly driven by scenario country, with the lowest agreement when participants judge their own cultural contexts, showing within-culture pluralism. Finally, human-LLM alignment experiments show that models can match majority choices, but fail to capture response distributions and uncertainty (with best correlations reaching only 0.24). Together, these findings motivate alignment evaluations that go beyond majority to capture cultural pluralism and disagreement in social judgment.

Ensuring that agent behaviours are aligned with human moral values inevitably raises the problem of how to account for the plurality of moral perspectives that societies -- and even individuals -- typically adopt. Work on moral uncertainty proposes mechanisms to fairly and democratically aggregate evaluations of actions across different moral theories. However, this paper argues that one needs to account for contextual factors when aggregating moral evaluations. For example, consequentialist perspectives assume an ability to accurately determine how an agent's actions change the world; an assumption that often does not hold in real world settings. We, therefore, formalise agent decision making under moral uncertainty, while also accounting for these kinds of contextual factors. We thereby show that a seemingly commonsensical property -- the weak Pareto principle -- is violated. We argue that this apparent problem is, in fact, a variation of Simpson's paradox, and hence reveals the limitations of aggregation mechanisms that ignore the impact of contextual factors.

Cultural-aspect work on large language models is dominated by a negative target: which outputs to suppress. We argue that a constructive counterpart is also needed, a working definition of what a culturally coherent response is rather than only what it must avoid, and instantiate it for Korean. We design an alignment-data pipeline around a prompt-based LLM seed generator that expands a Korean harm taxonomy, with a Korean-culturally-adapted safe-response policy at its centre: a per-category guideline grounded in Korean legal frameworks, social norms, and interpretive conventions, against which three frontier models each produce a candidate response. DPO fine-tuning on the resulting triplets improves the Korean cultural safe rate across six open-weight LLMs while causing no large degradation on Korean general-capability benchmarks, and qualitative outputs show fine-tuned models naming Korean statutes and institutional procedures and, where appropriate, supplying constructive Korean-context information alongside refusal.

Safety alignment for large language models relies on preference data, but current pipelines often train on large, redundant datasets. Existing data selection methods typically score each preference pair independently, collapsing directional preference information into scalar quality or diversity scores. This sample-centric view is especially limiting in multi-dataset settings, where shared safety directions coexist with dataset-specific residual risks. We propose DOG-DPO, a training-free data selection framework that treats preference pairs as structured geometric signals. DOG-DPO first represents each preference pair as a direction in model representation space. It then decomposes multi-dataset preference geometry into a global anchor subspace and dataset-specific residual subspaces. Finally, it selects subsets by maximizing diversity-based coverage, encouraging broad, non-redundant coverage of alignment directions before DPO training. Across six safety benchmarks and two model backbones, DOG-DPO achieves a strong utility-robustness trade-off using only 11% of the preference pairs. It recovers most of the safety gains of full-data training while remaining entirely teacher-free, training-free, and substantially faster than representative selection baselines.

Large Language Models (LLMs) are often fine-tuned through Reinforcement Learning from Human Feedback (RLHF) to align with people's preferences and values. However, this method has known limitations: it aggregates conflicting preferences, often relies on unrepresentative samples, and uses only binary comparisons. Analysing 1,500 open-ended responses from the PRISM dataset across 75 countries, we examine what people actually want from AI systems and reveal concrete failures of current methods. We find that different people want different things: most values are requested by fewer than a quarter of respondents, with truthfulness the sole exception at 49%. Furthermore, the same words hide divergent meanings: when people describe what they mean by "truthfulness", they reveal distinct, potentially incompatible, epistemological bases, as some ask for sourced claims, some for expert opinions, and some even ask for unpopular views. Certain capabilities, namely how human-like a model behaves, and some features, like AI guardrails, are outright controversial, with some desiring them and others rejecting them. We additionally find that people often use contextual distinctions (what AI should do "by default" versus "if requested") that binary comparisons cannot capture. These findings expose fundamental problems in current alignment practices. When 49% request truthfulness but define it differently, this is unlikely to be captured by a single reward model. The persistence of high hallucination rates in well-funded models, despite users' clear demands for accuracy, suggests that current methods fail to identify actual preferences. This paper sheds light on the situated, contested, imperfect signals that are currently being flattened into universal preference models, a practice others have characterised as epistemic violence.

The mechanisms behind LLMs' broad over-generalization beyond training examples remain unclear. Emergent misalignment (EM) offers a striking case study: finetuning on narrow tasks induces broad misalignment to semantically-unrelated test domains. In this work, we propose the Piggyback Hypothesis: the chat-template tokens can piggyback the finetuned behaviour onto out-of-domain queries. We validate this hypothesis by showing that subtle perturbations to the prefix (tokens preceding all user queries), or patching the prefix representations with those from the unfinetuned model, can restore alignment without changing the user query. Building on this finding, we propose Token-Regularized Finetuning (TReFT), which regularizes specific token representations during training to mitigate EM. Across different models and multiple EM-inducing datasets, TReFT reduces EM while preserving in-domain learning. On Llama-3.1-8B finetuned on the legal domain, TReFT achieves 33.5% more EM reduction than data interleaving with a retain set of aligned examples. We further show that TReFT extends to other narrow-finetuning settings, including abstention, tool use, and refusal (off-topic generalization is reduced by 54.3% on average), supporting the Piggyback Hypothesis. Broadly, our work highlights that LLMs may learn and generalize in unintended ways and suggests a path toward more constrained finetuning. It also calls for further study of how shared input features can piggyback model behavior across domains.

We argue that modern generative models create structural risks for knowledge and cultural production at current, sub-AGI capability levels. We define Human Temporal Learning (HTL) as path-dependent knowledge accumulation through sustained engagement with problems over time. Generative outputs increasingly resemble HTL-intensive work in surface features, so verifying whether a given output reflects genuine human learning grows costly relative to its expected benefit. Once verification loses economic justification, evaluators reward outputs regardless of production mode, and producers who invested years of learning compete on price against outputs that cost almost nothing to generate. We call this pathway value collapse and formalize it through a costly-inspection framework. Cross-domain evidence from academic publishing, legal practice, content platforms, and software security maps onto four stages of verification erosion. Alignment success is orthogonal. Better-aligned models narrow observable gaps between human and AI outputs, making source verification harder and intensifying competitive pressure against HTL-intensive work even when individual AI outputs improve.

Performance estimation under distribution shift aims to predict how a model behaves on an unlabeled test set whose distribution differs from the training data, a scenario that requires reliable indicators that can faithfully reflect model behavior without ground-truth labels. Existing approaches rely solely on the outputs of the given model whose biases are amplified once the distribution shifts, weakening the correlation with the true performance. Motivated by this limitation, we propose Fused Reference Alignment Prediction (FRAP), which leverages the complementary strengths of an external foundation model and the base model to construct a more reliable surrogate of the ground-truth labels. FRAP aligns the prediction distribution of the foundation model with that of the base model by applying temperature-scaled calibration that minimizes their divergence. The aligned predictions are fused through confidence-based weighting into a refined reference distribution that integrates robustness from the foundation model and domain-specific expertise from the base model, and performance estimation is obtained by measuring how closely the base model predictions agree with this reference. Extensive experiments across diverse datasets and architectures show that FRAP provides consistent and substantial improvements over representative performance-estimation methods under distribution shift.

We introduce pVR, a topological machine learning framework for alignment-free genomic sequence classification that combines $p$-adic numbers with topological data analysis. Each DNA sequence is encoded along two complementary axes: a $p$-adic distance on $k$-mer prefixes, which captures hierarchical positional structure, and a compositional $L_1$ distance on $k$-mer frequencies, which captures local sequence content. The two distances jointly parameterise a bi-filtered Vietoris--Rips complex, and per-sequence topological summaries from this bi-filtration serve as features for standard machine learning classifiers. We establish theoretical guarantees for the construction: stability under metric perturbations and invariance to the choice of prime, alongside a result that explains why a single $p$-adic axis is topologically uninformative and why the bi-filtration recovers nontrivial homology. On twelve genomic benchmarks ($28$ to $500$ sequences, $3$ to $7$ classes), pVR outperforms four established alignment-free baselines on three of six low-sample datasets, with gains of up to $21$ percentage points; it underperforms only on a SARS-CoV-2 variant benchmark whose point-mutation divergence violates the hierarchical assumption, and all methods saturate in the large-sample regime. pVR also outperforms zero-shot frozen embeddings from the 500M-parameter Nucleotide Transformer v2 by $6.7$ to $11.4$ percentage points on three low-sample benchmarks. The pVR codebase is publicly available at https://github.com/MAHI-Group/pVR.

Consistency training encourages models to behave similarly across different contexts, and has shown promise for reducing misalignment. We broaden the scope of consistency training in two ways. First, we introduce two new internal consistency targets: MLP Consistency Training (MLPCT), which matches post-activation MLP states, and Attention Consistency Training (AttCT), which matches per-head attention distributions. Second, we apply consistency training to four additional safety threats: persona in-context learning attacks, adversarial frustration, prefill attacks, and conditional misalignment. Across several models and threat settings, we find that consistency training reduces misalignment well beyond the sycophancy and jailbreak settings studied in prior work. We also find cases of cross-threat generalization, where training against one failure mode improves robustness to another, and identify a shared residual-stream mechanism underlying ACT, MLPCT, and AttCT, while distinguishing BCT as mechanistically distinct. Our results suggest that consistency training is a flexible and extensible framework for alignment, capable of unifying defenses against a broader class of model pathologies.

Emergent misalignment (EM) has been linked to the activation of misaligned persona vectors and evil character traits, suggesting that EM operates through disruption of the model's aligned character rather than direct learning of harmful content. Motivated by this connection, we study self-generated text recognition (SGTR) finetuning as a character-targeted intervention that is distinct from existing in-training defenses. We conduct two-stage finetuning experiments across three models (GPT-4.1, Qwen2.5-32B-Instruct, Seed-OSS-36B-Instruct) and multiple EM datasets to compare SGTR finetuning against benign finetuning baselines (correct domain-specific data, general knowledge, and word counting) to find it an effective defense in both reversal and prevention settings. We find that all interventions produce comparable EM reversal, but only when restoring capabilities that EM had degraded. For prevention, only SGTR finetuning consistently reduces misalignment without exacerbating any individual metric, suggesting that character fortification specifically drives prevention. We provide further evidence for EM's relation to the LLM's default character by showing that EM finetuning induces diversity into the LLM's identity self-reports, artificially corrupting self-recognition exacerbates misalignment caused by EM finetuning, and that removing the model's identity-bearing system prompt substantially reduces the effect of EM finetuning. Together, these findings reframe EM not as the adoption of a coherent misaligned persona but as the destabilization of aligned character.

Many human-centered tasks, including natural language inference (NLI) and emotion recognition (ER), have multiple plausible interpretations, leading to label ambiguity and challenging disagreements across human annotators. As LLMs are increasingly deployed in real-world settings, faithfully modeling such ambiguity is essential to identify contested inputs, preserve variability in ambiguous cases, and capture the full distribution of human judgments. Yet, existing LLM alignment approaches have predominantly assumed a single correct label, excluding annotator disagreement during optimization. Instead of treating this ambiguity as noise, we show how to treat it as information that improves model behavior through a new algorithm called SMARTLY HANDLING AMBIGUOUS LABELS IN ALIGNING LLMS (SHALA-LLM). This reinforcement learning framework provides a new way for LLMs to learn directly from annotator distributions while dynamically prioritizing highly ambiguous samples during optimization. Experiments on ambiguity-sensitive NLI and ER benchmarks, including ChaosNLI, GoEmotions, and MSP-Podcast, demonstrate that SHALA-LLM improves agreement with annotator label distributions, e.g. on ChaosNLI, it reduces Jensen-Shannon Distance by up to 62.1%. At the same time, SHALA-LLM improves F1 by up to 16.7%, showing that modeling annotator disagreement can also strengthen classification performance.

LLMs can appear cautious in risk decision-making tasks, yet cautious-looking outputs do not necessarily indicate alignment with human decision-making mechanisms. We investigate this distinction using the St. Petersburg game as a controlled testbed, a classical paradox in which the expected payoff is infinite, yet humans typically report low, finite willingness to pay. We evaluate 28 LLMs with a structured prompt suite that includes the original game; controlled decision variants that perturb truncation, repeated play, numeric endowment, and occupational identity; a human-perspective prompt that asks models to reason as human decision makers; and paired comparisons between base models and their instruction-tuned counterparts. In the original game, most models generate finite bids, creating the appearance of human-like risk behavior. However, this outcome-level resemblance masks substantial mechanism-level differences. The controlled variants reveal that rather than maintaining human-like behavior seen in the original game, models often shift to conditionally and computationally rational behavior. Human-cue prompting and instruction tuning often lower bids and reduce some visible pathologies, but most mechanism-level response patterns remain largely unchanged. These findings show that behavioral alignment in risk decision-making can be surface-level: LLMs may produce human-like risk decisions without exhibiting human-consistent mechanisms. High-stakes evaluations of LLM decision-making should therefore move beyond outcome similarity and examine whether the alignment is supported by mechanism-level consistency.