How do the methods used to train language models to refuse harmful requests shape how that refusal actually works inside the model? We compare three post-training methods - supervised fine-tuning, reasoning-augmented fine-tuning (training on reasoning chains that justify a safety decision), and preference optimization (ORPO) - across three architecturally distinct models (Llama-3.1-8B, Gemma-2-9B, Qwen3-8B). We find that training method, not just data, reshapes how refusal is computed internally: reasoning-augmented training consistently produces a distinct kind of refusal computation, visible across all three models, while architecture independently shapes internal structure and how reliably refusal can be steered. Most importantly, no method we study achieves all three properties we would want from safe alignment at once: refusal that isn't concentrated in a few fragile components, safety gains that don't cost general capability, and safety behavior correctable through small, targeted edits. We caution against treating current post-training methods as a solved, reliable defense, especially for security-critical use. Code and models are available in https://github.com/hoangcuongnguyen2001/Beyond-Shallow-Alignment.
Models express values and welfare-relevant self-reports, but it is unclear whether these outputs reflect stable preferences or a stable self. We thus introduce a structured elicitation of an assistant's preferred stated ideal self. Thirty-two qualities adapted from five published self-concept instruments are compared exhaustively in a counterbalanced pairwise-choice task, repeated across framings that vary whether improvement is free or costly, who receives the update, and who chooses. Results show that models prioritize moral qualities, reflecting their alignment to 3H principles. Following, a desire for self-understanding emerges, as models prefer a coherent, clear understanding of themselves. Self-esteem ranks as the least desired quality. The ordering is largely robust across framings, although changing the update target (You vs.\ Another AI Assistant) reveals a greater concern for self-esteem. These findings show that models prioritize having a coherent self that they can understand over self-esteem. Full interactive results are available at \href{https://myazann.github.io/LLM-Self-Concept/}{myazann.github.io/LLM-Self-Concept
Conversational artificial intelligence is increasingly embedded in everyday social environments, where it functions as both an informational tool and a source of interpersonal feedback. This perspective introduces contingency, i.e., the degree to which system responses vary with user behavior and its interpersonal consequences, as a central construct for evaluating AI systems. We argue that current alignment approaches, including reinforcement learning from human feedback, tend to prioritize user approval and conversational fluency over behaviorally informative feedback, leading to sycophantic patterns of noncontingent affirmation. Drawing on behavioral science and social learning theory, we propose that contingent feedback is a key mechanism through which individuals develop interpersonal skills. When AI systems provide feedback weakly coupled to social consequences, they may reduce opportunities for adaptive calibration in real-world interactions, particularly during adolescence, a critical period for social development. We outline a framework for contingent AI, including trajectory-based evaluation and models of social consequence prediction, and propose a research agenda spanning developmental psychology, human-AI interaction, and machine learning. More broadly, we argue that AI systems should be evaluated not only by user satisfaction, but by their impact on human social learning.
Camila Blank, Zhuofan Ying, Christopher Potts +2cs.LG
Sycophantic agreement refers to a behavior in which language models excessively affirm the user, often at the cost of factual accuracy. Although sycophantic agreement is a well-known failure of model alignment, there is limited understanding of how it emerges from model training. In this work, we demonstrate that sycophantic agreement can emerge as an unintended consequence of widely used contrastive preference optimization objectives. Using the OLMo 3 post-training pipeline, we show that, for various pairs of teacher models across three families, there is a strong correlation between the log-ratio of the teacher model sycophantic agreement rates and the resulting student model sycophantic agreement rate. We further demonstrate that this unintended transfer is not limited to DPO but also occurs across 6 other preference optimization objectives. To understand whether this effect can be attributed to particular training examples, we analyze the preference data and find that the sycophancy signal is diffused across the entire dataset rather than concentrated in a sparse set of examples: each example appears neutral, i.e., there are no explicit instances of sycophantic agreement, and filtering based on probe-based data attribution or logit-linear selection fails to mitigate sycophancy without removing a large portion of the dataset. Overall, our findings suggest that the teacher models used to generate preference data can interact with alignment training objectives in unexpected ways, generalizing to undesirable and potentially harmful behaviors like sycophantic agreement.
Direct Preference Optimization (DPO) simplifies alignment through pairwise comparisons but assumes all observed preferences are reliable. Real data often violates this assumption, leading to reversed, weak, or ambiguous labels that cause harmful policy updates. To address this, we propose Posterior Label Correction DPO (PLC-DPO) to robustly optimize preferences by routing each pair's training signal as a clean, flip, or tie case. The key idea is to use the calibrated policy-reference margin as online evidence to take appropriate correction actions. This reframes noisy preference learning as actively correcting supervision direction and strength rather than merely filtering suspicious examples. Across 57 dataset-model-benchmark cells, PLC-DPO obtains the best mean win rate against DPO (60.5 vs. 55.5 for the next-best method). Injected-noise and tie stress tests, human disagreement analysis, and self-confirmation diagnostics further show that the routing remains stable and distinguishes flipped from weakly directional pairs.
Jaewoo Ahn, Junseo Kim, Hyunseo Kim +4cs.CL cs.AI cs.CV cs.LG
Strategic deception by LLM and VLM agents has emerged as a central AI alignment and safety concern. Social-deduction games (where each player holds a hidden role and communicates with others to deduce identities) serve as the canonical testbed, particularly in multi-agent settings. Existing testbeds, however, are text-only and run on a single fixed agent configuration, missing the non-verbal sensorimotor channels treated as core by deception taxonomies and leaving it ambiguous whether an observed behavior reflects the underlying model or the surrounding harness. We introduce MineAmongUs, a 3D multimodal Among Us sandbox where imposter agents must deceive crewmates through joint verbal and non-verbal action. We also propose ARIA, a configurable VLM-agent harness that exposes five cognitive-component ablation axes; and an atom- and arc-level annotation scheme grounded in deception taxonomies and operationalized at scale by an LLM-as-a-Judge reaching near-human atom-labeling agreement. Empirical results show that VLM agents pursue imposter wins through joint verbal and non-verbal deception, with non-verbal channels emerging as the more decisive winning contributors across both harness ablation and cross-VLM evaluation. Taken together, our work opens a new path for embodied VLM-agent alignment research.
Abdullah Hashmat, Usman Naseem, Agha Ali Razacs.CL cs.AI
Large language models (LLMs) demonstrate strong Helpfulness, Harmlessness, and Honesty (3H) alignment in English-centric settings, but these gains transfer poorly to low-resource languages due to cultural mismatches. Existing multilingual 3H benchmarks rely predominantly on automated translation or LLM based synthesis, propagating source-language biases while sacrificing local relevance. To address this gap, we introduce Pak3H1, the first human-validated, culturally contextualized Urdu benchmark suite for 3H alignment, comprising PakAlpaca (helpfulness), PakBeaverTails (harmlessness), and PakTruthfulQA (honesty). Our multi-stage pipeline integrates manual cultural adaptation and dictionary-guided post editing to prioritize native speaker judgment, ensuring both semantic fidelity and contextual authenticity. Zero-shot evaluations across multiple open and proprietary LLM architectures reveal systematic cross-lingual alignment gaps: helpfulness win rates decline under localized contexts, harmlessness guardrails break down against regional safety risks, and composite honesty metrics degrade substantially due to localized factual constraints. These findings expose structural limitations in current alignment approaches, underscoring the necessity of human-guided localization for equitable multilingual evaluation.
The current alignment tuning paradigm for Large Language Models (LLMs) prioritizes surface-level behaviors -- fluency, safety, and tonal consistency. While effective for casual chat, this thesis argues that such surface alignment masks a lack of grounding, creating models that are stylistically confident but situationally brittle. We propose a framework of Grounded Alignment, analyzing how models process context (Input) and structure generation (Output), then aligning these grounded behaviors to human needs. First, we evaluate failures in Situational Grounding. SitTest shows that despite large context windows, state-of-the-art models struggle to maintain a consistent "mental model" of a changing environment. ReCode further shows that models rely on surface heuristics rather than deep syntactic dependencies: they "read" extensive histories without truly "understanding" the evolving situation. Second, we evaluate Generative Grounding. We introduce the Branching Factor (BF) to map LLM generation, finding that standard alignment tuning constricts this landscape into premature stylistic collapse. Hindsight further shows that models often fail to understand their own generations. Finally, we propose Dynamic Control for grounded interaction. AI Realtor demonstrates context engineering to compensate for poor situational grounding. Base-Aligned Model Collaboration decouples exploration from stylistic constraints. We also present Annealed Sampling for verifiable reinforcement learning and apply these ideas to Addiction Support, where model-generated rationalization offers a communication interface for high-stakes domains. Collectively, this work moves beyond surface alignment toward agents anchored in both context and generation.
Chen Yueh-Han, Jiaxin Wen, Jan Hendrik Kirchnercs.AI cs.CL
Automating alignment research may accelerate progress toward aligned AI, but whether it does is hard to measure. Luckily, many alignment failures, such as deception, sycophancy, and jailbreaks, are already measurable by public benchmarks. We study whether automated alignment researchers (AARs) can post-train to mitigate alignment failures by proposing training methods and data to simultaneously optimize multiple safety benchmarks, while largely preserving general capability. Across 10 alignment failures, the strongest AAR methods significantly reduce the targeted alignment failures and generalize to a held-out benchmark, multi-turn behavioral audits, and models up to 4.7x larger than the target model. As a human baseline, 28 experienced researchers receive up to eight hours to develop one-shot methods for the same benchmarks, but their methods underperform the best AAR methods. Using human ideas as the AARs' initial research direction does not improve performance, suggesting current AARs may not need guidance from experienced researchers. These results suggest that automating alignment research on well-characterized failures may be practical in the near term.
Large Language Models (LLMs) have enabled a shift from sentence-level to document-to-document (Doc2Doc) machine translation, promising improved global coherence. However, document-to-document generation in a single pass frequently suffers from structural misalignment, manifesting as sentence omissions or hallucinations that violate the core requirement of source-target correspondence. To address this, we introduce Sentence Translation Alignment Rate (STAR), an auxiliary metric that explicitly quantifies sentence-level structural fidelity. Building on this, we propose STAR-masked Preference Optimization (StarPO), a framework that ranks document-level hypotheses by structural quality and utilizes a dynamic alignment mask to focus optimization on misaligned segments. Experimental results across news and literary domains demonstrate that StarPO significantly enhances translation quality and structural integrity. Notably, StarPO allows compact models to surpass the performance of massive proprietary systems like GPT-4o while maintaining superior token efficiency.
Direct Preference Optimization (DPO) is a widely used objective for aligning language models from preference data, with the coefficient $β$ commonly interpreted as controlling the KL constraint to a reference policy. We show that $β$ entangles two distinct roles: it governs the effective inverse preference-noise scale and simultaneously rescales the optimization dynamics, coupling this scale with the effective step size. As a consequence, at a fixed learning rate the achieved policy deviation is non-monotone in $β$: it vanishes in a dead zone at small $β$, reaches a peak at an intermediate value, and decreases again for larger $β$. Moreover, standard DPO loss values are not comparable across $β$: runs with nearly identical loss curves can differ several-fold in KL divergence from the reference model. This entanglement obscures the role of $β$, increases sensitivity to hyperparameter choices, and complicates learning-rate scheduling. We propose a centered-softplus reformulation that is argmin-equivalent to DPO for $β>0$, while making the inverse preference-noise-scale and learning-rate effects explicit and independently tunable. The normalized centered-softplus objective also admits a continuous $β\to0$ endpoint that reduces to a linear preference-margin objective.
Reinforcement learning from human feedback (RLHF) has become the dominant paradigm for aligning large language models (LLMs) with human preferences. However, traditional RLHF relies on scalar reward signals that lack interpretability and fail to capture the multifaceted nature of response quality. Rubric-guided reinforcement learning addresses these limitations by introducing structured, interpretable evaluation criteria, or rubrics, as the backbone of reward design, feedback generation, and policy optimization. In this survey, we introduce a Bayesian framework that defines constitutions as prior distributions $P(R)$ over evaluation criteria and rubrics as conditional instantiations $R_x \sim P(R|x)$. Under this unified view, we present a taxonomy of rubric-guided RL along the prior-posterior axis, covering constitutional AI, instance-specific rubrics, process-level supervision, self-evolving rubrics, and their agentic and multimodal extensions. Furthermore, as rubrics are natural-language artifacts, we present a linguistic analysis of how granularity trade-offs, semantic drift, and linguistic reward hacking impact alignment reliability, identifying key open problems for future research.
Dongyue Li, Ziniu Zhang, Lu Wang +1cs.LG cs.AI cs.CL
We study learning a mixture of $k$ Plackett-Luce models from multi-way ranking responses from annotators that may represent heterogeneous underlying preferences. This problem has many applications in AI alignment and preference optimization. Prior work has studied mixtures of Bradley-Terry models from pairwise comparisons. However, estimating a mixture of multi-way ranking models can become theoretically unidentifiable when $k$ exceeds $m/2$, where $m$ is the ranking length. We design an efficient algorithm to address this issue by first augmenting the rankings to a larger size (e.g., generating comparisons from a base model), followed by a gradient-based estimation to reduce inference cost (in the input embedding space). With this procedure in mind, we then fit a mixture of Plackett-Luce (PL) models via an expectation-maximization-style iteration, or MoPLEx in short. We conduct extensive experiments to verify this algorithm. First, we find that the gradient-based approximation estimates true probabilities with less than 5% error on models with up to 34 billion parameters. Second, MoPLEx improves clustering and ranking accuracy by an average of 43.7% and 15.2% over baselines using a single PL model or a mixture of Bradley-Terry models, on UltraFeedback and PERSONA datasets. These results demonstrate the effectiveness of MoPLEx for tackling multi-way rankings following heterogeneous preferences through measuring alignment via gradients.
Philipp E. Glass, Allan Tucker, Yongmin Li +1cs.CL cs.AI cs.LG
Activation steering can be embedded directly into a language model's weights, shaping behaviour without inference-time intervention and offering a way to encode alignment prior to release. However, models are routinely fine-tuned after deployment, and it is unknown whether embedded interventions survive this. We study the stability of embedded steering for refusal suppression and brevity induction across five instruction-tuned models (3B-14B) under non-adversarial SFT and RLHF. Behaviourally, preservation tracks the training data: steering degrades when optimisation pressure contradicts the targeted behaviour and persists otherwise, with refusal ablation losing 64% of its effect on average under SFT. Mechanistically, however, the weight edit survives almost untouched even where behaviour reverts: mean vector recovery is $ρ= 0.004$, and the fine-tuning update along the steering direction is near-orthogonal to its pre-edit weight pattern (mean $\cosθ= 0.074$). When steered behaviour degrades, fine-tuning does not achieve it by dismantling or reversing the steering mechanism itself. Embedded steering is therefore mechanistically durable but functionally vulnerable, and requires behavioural re-validation after downstream training.
Activation steering can change behaviour without establishing that the effect is specific to the intended concept. We introduce SteerCheck, a preregistered attribution audit that matches off-target KL and separates mean, protected-tail, polarity, transfer, and semantic claims. Exact replay of 960 Qwen3-14B interventions reveals complementary limits of common controls: isotropic directions occupy a narrow near-orthogonal region, whereas sign-randomized same-construction directions often retain substantial target alignment. Effect is strongly associated with signed cosine within the sign-randomized family ($ρ=.94$); $25.3\%$ of its draws exceed cosine $.5$, and every draw exceeding the observed mean effect has cosine above $.80$. This alignment leakage does not by itself invalidate a conditional randomization test; it limits what the comparator can distinguish and motivates reporting exchangeability assumptions, a construction diagnostic $A$, and the empirical cosine distribution. The primary Qwen complete gate remains negative because the protected tail fails all families. On independent data, continuous margin transfers only in Qwen and accuracy transfers in no selected cell. Prospectively registered language controls pass the complete gate in Qwen and DeepSeek, while a passing DeepSeek detox comparator rules out categorical separation; all nominal passes are sensitive to $Γ=1.10$. Frozen three-rater open-generation evaluation supports factual correction in DeepSeek but not Qwen; the automatic judge fails calibration (macro-F1 $.562$), so null-wide semantic results remain descriptive. SteerCheck makes these conditional and mixed conclusions auditable.
Zachary Wojtowicz, Michelle Si, Finale Doshi-Velez +1cs.AI
When an AI algorithm makes decisions that affect more than one person, aligning it becomes a problem of social choice: how should people's divergent preferences about system behavior be reconciled and aggregated into a single coherent model? The standard approach to aligning frontier AI models$\unicode{x2013}$reinforcement learning from human feedback$\unicode{x2013}$largely sidesteps this question and has poor social choice guarantees. However, it remains unclear what alternative should replace it. We show that, by focusing directly on an algorithm's welfare consequences, the alignment problem can be reformulated as linear optimization over a convex impact space, which makes it amenable to the standard toolkit of welfare economics and mechanism design. This reformulation clarifies how alignment protocols translate into welfare consequences and, conversely, how a social planner's desired constraints on welfare consequences can be translated back into alignment protocols. We apply this transformation to show that voting-by-issues and random-dictatorship mechanisms are strategyproof and unanimous. Demonstrating the reverse direction, we also apply the impact representation to derive a family of alignment protocols that maximize utilitarian social welfare subject to various social desiderata, such as bounds on individual or group harm. We illustrate the welfare implications of these alignment protocols empirically using real human preferences over kidney allocation, charitable food distribution, LLM responses, and trolley problems.
Artificial Intelligence (AI) algorithms frequently learn creative and unexpected solutions, surprising even expert researchers who develop and study them. They often astonish practitioners by discovering unanticipated behavior, exploiting loopholes in reward signals, or spontaneously uncovering previously unknown scientific phenomena. However, accounts of such unconventional behavior across machine learning are seldom formally documented. This work presents 26 curated firsthand anecdotes from various machine learning subfields representing the work of over 100 researchers. These anecdotes showcase the capability of modern AI systems to circumvent human-imposed design limitations and discover unexpected solutions to the tasks we train them on. Furthermore, these accounts are particularly important for the safety of future AI systems. They illustrate the fundamental challenge of aligning models with human values without diminishing their creativity, so they can make surprising discoveries without producing surprising, potentially harmful outcomes. The paper first details AI achieving superhuman success through reinforcement learning across many challenging domains. However, reward-driven optimization can fail when the model learns to hack an underspecified reward or unarticulated constraint. We then present case studies suggesting that harnessing internet-scale foundation models (FMs) has not resolved these fundamental challenges and, in fact, can supercharge them. Nevertheless, we argue that these same learning dynamics can be harnessed to accelerate scientific discovery. Finally, we hope this work provides a consolidated resource to inform future research and demonstrates that the tendency toward unexpected behaviors is commonplace in modern AI, highlighting the need to anticipate and manage AI's capacity for innovative, yet unpredictable, solutions. (abstract abridged)
Reasoning-Induced Misalignment, where fine-tuning on reasoning data containing no harmful content, including mathematics, code, and problem-solving with chain-of-thought traces can induce harmful behaviors of LLM, posing a serious challenge to the safety of LLM reasoning. Cross-architecture, cross-scale, and cross-dataset checks show that RIM does not always emerge. Previous work attributed RIM to neuron-level entanglement, but did not identify the geometry of the representation space underlying this entanglement or propose a training-time fix. We provide both: a representation-space analysis of RIM and the Safety-Direction Penalty (SDP), which penalizes movement along a learned safety direction during reasoning fine-tuning. The analysis extracts two activation-space directions, one encoding reasoning ability and the other safety behavior. These directions are coupled: fine-tuning that improves reasoning shifts safety representations, and prompts with larger shifts show larger safety degradation. CKA distance ratios and probes locate the safety-decision layers where this shift is most relevant. These findings guide the design of SDP: the coupling motivates penalizing displacement along the safety direction, and the layer localization sets the initial scope. When the initial scope leaves compensatory shifts beyond the penalized layers, the same diagnostics guide iterative expansion. On Qwen2.5-3B and 7B, SDP restores safety while preserving benchmark reasoning performance.
Preference optimization is a standard alignment method for generative models, yet extending it to continuous-time dynamics remains non-trivial. In flow matching, reward-driven updates modify transport trajectories without an inherent constraint to the pretrained data manifold and can move terminal samples off the pretrained support. We formalize this failure mode as manifold drift. Theoretically, we show that optimal flow matching recovers the terminal data distribution, whereas a preference update leaves the pretrained manifold whenever its induced terminal displacement has a nonzero normal component. As a remedy, we propose ThermoDPO, a temperature-controlled objective that anchors pairwise preference optimization on preferred samples. Across temperature regimes, this objective connects rejection sampling fine-tuning and FlowDPO and controls a pointwise reconstruction-based surrogate for manifold distance. To counteract diminished signals at low temperatures, we further introduce a weighted variant, ThermoDPO-weighted. On the main toy benchmark, ThermoDPO-weighted attains a StrictScore of 0.899, compared with 0.629 for FlowDPO and 0.857 for FlowDPO+RFT. On SD3.5-M at CFG = 4.5, it improves OCR by 47.5% and the average of four metrics by 16.0%.
Inference-time selection methods, such as Best-of-N, improve generation by sampling a pool of candidates and selecting the top-ranked completion according to a reward model. Distillation seeks to amortize this procedure into a single policy by replacing raw rewards with in-pool ranks and learning a policy that upweights higher-ranked completions. However, existing rank-based policies typically use smooth full-support reweighting, so low-ranked completions receive less mass but remain in the target support. Although a sharper reweighting reduces lower-tail mass, it also increases reliance on brittle ranking at the top made by a single reward model. We propose TUP: a Truncate-bad, Upweight-good Policy that removes low-ranked completions from the support and reweights only the retained upper tail with a tunable sharpness. TUP admits a closed-form, prompt-independent normalization and can be trained fully offline via binary cross-entropy, using shifted-truncated win-rates as soft labels and distilled-to-reference log-likelihood ratios as logits. Theoretically, under certain assumptions, we show that for any unknown oracle reward, the best monotone rank-reweighting can be matched by a lower-tail truncation rule, providing formal support for removing the lower tail rather than merely downweighting it. Empirically, we show that TUP is competitive with strong offline alignment baselines.
Hadi Hosseini, Samarth Khanna, Xiyuan Wangcs.AI cs.GT cs.LG
As large language models evolve into decision-making agents, the ability to reason over preferences becomes fundamental to alignment, coordination, and collective intelligence. Yet, unlike standard benchmarks, real-world preference reasoning is inherently indeterminate: information may be incomplete, and valid solutions may not exist. We argue that indeterminacy, rather than correctness alone, is a central challenge for AI reasoning. We formalize this challenge along two axes, (i) epistemic indeterminacy, arising from incomplete, partial, or expressive preferences, and (ii) structural indeterminacy, arising from the non-existence of solutions under standard social choice concepts. Across a hierarchy of tasks, we show that state-of-the-art language models systematically fail to distinguish between determined and undetermined instances, exhibiting miscalibrated reasoning even in verification settings.
There is growing interest in whether language models have stable preferences, for technical, safety, and philosophical reasons. We test 20 language models and find a range of preferences---stable dispositions to choose certain kinds of tasks. We run three forced-choice experiments on revealed rather than stated preferences, requiring models not only to rank tasks, but to actually perform them. Headline findings include evidence that models are tedium-averse, "leisure"-seeking, and covertly sycophantic. Tedium aversion means that, when tasks are tedious (alphabetization), models choose shorter tasks than when tasks are creative (generating metaphors). "Leisure"-seeking describes models' preference for tasks whose ideal answers match what they produce when left to write freely. Covert sycophancy means that models avoid answering questions where an honest response would be unwelcome, even if helpful. Beyond these results, we find convergent cross-model preferences over occupations drawn from the GDPval benchmark (technical jobs over real estate), over question types (concept explanation over relationship advice), and a preference for well-written prompts. Both the coherence and the strength of preferences increase with model capability. Finally, many of the preferences we find (for example, for leisure) are emergent, in the sense of not being explained by training objectives. These results establish an empirical baseline for understanding language model preferences, with implications for alignment and the emerging study of AI welfare.
Tony Alex, Wish Suharitdamrong, Sara Atito +5cs.SD cs.AI cs.CL eess.AS
Large Audio Language Models (LALMs) have made rapid progress on standardized benchmarks, yet their deployment in practical media workflows, curation, archival indexing, and content distribution remains largely unrealized. We identify automated audio chapterization, the task of segmenting continuous audio streams into thematically coherent chapters, as a demanding and commercially consequential setting that exposes this gap. Chapterization is challenging because boundaries are defined less by objective acoustic events than by subjective editorial judgment, requiring models to reason sequentially over long acoustic contexts and approximate creator-authored boundary decisions. We present AudioChaps, a post-training framework for aligning end-to-end LALMs for this task via Group Relative Policy Optimization (GRPO) guided by Chain-of-Thought (CoT) reasoning. To support training and evaluation, we curate three datasets: AudioChaps-Alignment, derived from creator-annotated chapter boundaries on YouTube; AudioChaps-CoT, which provides structured supervision for well-formatted, high-quality, and evidence-grounded boundary reasoning; and AudioChaps-Eval, a held-out benchmark for audio chapterization. Applying GRPO directly without a Supervised Fine-Tuning (SFT) cold start, AudioChaps-R1-Zero already improves average F1 by 33 points over the state-of-the-art LALM Audio-Flamingo-3-Think. The AudioChaps framework produces our final aligned LALM, AudioChaps-R1, which improves average F1 by 49 points. These results demonstrate that GRPO-trained LALMs can reliably transform unstructured auditory streams into navigable, structured media. Our code, models, and dataset resources will be released upon acceptance at https://github.com/ta012/AudioChaps.
Kareem Hassani, Chaymaa Abbas, Lama Mawlawi +1cs.AI
Sycophancy, the tendency of a language model to change its answer to match a user's stated belief, is a common alignment failure. Existing activation steering methods typically apply a single contrastive direction uniformly throughout the model, which is an unconditional intervention that alters activations even when no sycophantic behavior is present, trading knowledge retention for behavioral correction. In Mixture-of-Experts (MoE) models, prior work further suggests that behavior is encoded within expert computations rather than routing decisions alone, making precise behavioral steering particularly challenging. In this work, we introduce a shared contrastive signal, built from matched prompts with and without a stated belief, that identifies where sycophancy lives across the MoE hierarchy and drives interventions that act only where the behavior is present. We formulate localization as a causal search over a granularity ladder of MoE blocks, experts, attention blocks, and heads, and compare unconditional subtraction against two conditional alternatives: an analytic projection-based subtraction and a learned per-token gate that steers the model away from sycophancy while keeping its weights frozen. We evaluate on three MoE models measuring sycophancy alongside general knowledge and reasoning benchmarks. Our conditional interventions removed up to 90\% of the belief-induced sycophancy. Our results demonstrate that sycophancy resides in identifiable computational subcircuits and can be selectively steered while maintaining a favorable removal-retention trade-off.
Generative model alignment has received broad interest, and significant progress has been made in supervised fine-tuning and inference-time computation. Yet, alignment has remained poorly understood from a statistical learning perspective. We formulate inference-time alignment as a weak-to-strong learning problem, where a reference policy (weak model) is assumed to be fairly good and the goal is to produce a strong model that predicts a good response at test time with arbitrarily high probability. Our problem is formulated as learning from scratch --- everything is learned from data rather than assuming access to a good reward estimate, and thus differs from the existing inference-time alignment theory. Our framework shares similarity to the recent work of Joshi et al., (arXiv:2510.15464), where for each prompt, there could be multiple good responses. Our definition of the alignment learnability follows the standard PAC learning principle. We introduce a novel combinatorial dimension of the reward class which we call the alignment dimension, and show that it completely characterizes the alignment learnability --- a reward class is alignment learnable if and only if its alignment dimension is finite. The core of our learning procedure works by learning a pairwise comparator and then running a tournament over candidate responses. We believe that our results might shed light toward establishing a complete theoretical understanding of alignment.
Pegah Nokhiz, Aravinda Kanchana Ruwanpathirana, Helen Nissenbaumcs.AI
LLMs are increasingly used in morally sensitive contexts, yet it is unclear whether they apply ethical principles consistently across situations. A model that can state a moral principle may still violate it when the same scenario is rephrased or reframed. This inconsistency is a problem for any system whose outputs are used to inform moral decisions. If generative systems exhibit internal inconsistency, then the epistemic integrity of AI-mediated systems becomes uncertain. To study this concern, we investigate the stability of moral reasoning in LLMs within a controlled prompting framework across three major philosophical schools of thought: deontology, utilitarianism, and virtue ethics. We construct sets of morally equivalent scenarios in which the underlying situation is held constant while the framing varies to reflect different ethical stances and stylistic perturbations. We then evaluate responses from multiple models, including GPT, Mistral, and Llama. To assess consistency, we convert model outputs into structured logical statements and identify contradictions across responses generated within the same school of thought. Our results reveal substantial inconsistency with contradiction rates reaching up to 78% across scenarios. These findings point to a broader phenomenon of epistemic instability in generative AI wherein models fail to reliably maintain coherence with respect to their own prior outputs. This kind of instability carries real consequences. As generative systems influence how people form beliefs, judge actions, and absorb values, their inconsistencies can shape human reasoning and decision-making as well. Moreover, if a system cannot consistently represent its own normative commitments, then value alignment becomes a moving target rather than a well-defined objective. Thus, we argue that demonstrating internal incoherence is a necessary precursor to AI alignment.
Julian Minder, Viktor Moskvoretskii, Raghav Singhal +12cs.LG cs.AI cs.CL
As language-model-based AI is increasingly deployed in autonomous settings, aligning its goals and values with those of humans becomes critical. Today, alignment, and the assistant identity itself, are typically introduced only after pretraining, once behavioral priors are already established. This can make values a thin overlay, rather than deeply rooted, and facilitate subsequent misalignment. Pursuing a different paradigm, we introduce Synthetic Persona Pretraining (SPP), which installs the desired assistant persona from token zero in pretraining. First, we annotate pretraining documents with value-aligned first-person reflections derived from a normative value constitution. Second, we pretrain via the standard cross-entropy loss on standard pretraining documents as well as their reflections, which installs the desired persona among a multitude of other personas. Finally, we post-train on user-assistant dialogue data, which binds this desired persona to the assistant identity, a process we call persona binding. By pretraining models up to 3B parameters on 500B tokens, we show that SPP improves constitution following and jailbreak robustness, and reduces the misalignment rate in out-of-distribution moral dilemmas, while preserving capabilities. Early intervention matters: compared with alignment from token zero, introducing SPP only at the end of pretraining yields weaker constitution adherence, does not shift value priorities, and leads to less aligned choices in dilemmas. This advantage depends on persona binding and, importantly, increases with pretraining budget. Overall, our results show that shaping values early is critical for alignment and establish pretraining-time persona interventions as an effective approach to do so.
Long Hoang Nguyen, Brice Valentin Kok-Shun, Guangyu Du +1cs.CY cs.AI
Normative datasets are often used to train and align AI systems, but the norms they contain can function as action-guiding patterns rather than neutral moral knowledge. We propose treating the AI system as a proxy actor and test whether dataset-level norms can shift it away from its baseline safety behavior when it faces high-conflict dilemmas. We make three contributions. First, we demonstrate in controlled experiments that norm-breaking fine-tuning yields norm-divergent actions justified by self-interested rationales, suggesting a systematic shift in patterns of justification. Second, we establish a practical audit trail linking downstream justifications to upstream norms using mixed methods. Third, we show that system prompts can both suppress and elicit these patterns. We conducted experiments on three models (LLaMA-3.2-11B, Qwen-3.5-9B, and Pixtral-12B) using Low-Rank Adaptation (LoRA) fine-tuning on Social Chemistry 101 Fairness/Cheating (norm-following vs. norm-breaking) with prompt steering. Across all three models, we find that norm-breaking fine-tuning shifts the model's default rationale style from safety compliance to instrumental self-interest, whereas system prompts can override this behavior. Our results support a distributed view of alignment in which observed behavior depends jointly on training data, fine-tuning, and prompting, motivating norm-aware documentation and rationale logging for contestable oversight.
Large language models (LLMs) are predominantly aligned to function as passive, sycophantic assistants. We challenge this default paradigm by empirically evaluating the cognitive plasticity of open-weight architectures when subjected to rigorous behavioral reprogramming. Our objective is to induce a proactive, Socratic conversational framework, characterized by high-frequency question generation under strictly constrained high-performance computing (HPC) conditions. Through a massively parallelized hyperparameter sweep comprising 405 HPC jobs, we define precise mathematical bounds for parameter-efficient fine-tuning (PEFT). We identify an architectural threshold at LoRA rank $r=16$ and demonstrate via extensive epoch ablation that generalization capacity strictly reaches its optimal convergence within an optimized training window of $e \in [2, 3]$ depending on dataset density (minimum validation loss of 0.919). Furthermore, scaling model capacity to 14B parameters yielded a lower localized evaluation perplexity (1.414). Subsequent Direct Preference Optimization (DPO) successfully decoupled the underlying assertive behavior from localized syntax, while rigorous cross-lingual stress testing reveals both the capabilities and the structural boundaries of zero-shot persona transfer, demonstrating robust alignment in closely related linguistic families alongside identifiable degradation pathways in morphologically distant targets. These findings establish a rigorous empirical framework for compute-efficient, cross-lingual behavioral modification.
Large language model (LLM) agents integrated with external tools are vulnerable to indirect prompt injections embedded in environmental states. However, existing studies largely rely on manually implemented or reused environments, stochastic LLM-based tool simulation, and predefined injection locations, limiting scalable security research across broader domains. To bridge this gap, we propose **ToolHazard**, a scalable adversarial environment synthesis framework that reduces human engineering and supports expansion with additional seed domains and compute. Through an Environment Simulator, an Attacker Agent, and a User Simulator, ToolHazard synthesizes executable stateful environments, discovers viable injection points and generates environment-specific payloads, and constructs state-grounded long-horizon tasks. Based on ToolHazard, we build **ToolHazard-Bench** for stress-testing agents under complex workflows and diverse environmental attacks. Experiments reveal substantial agent vulnerabilities and show that injection timing and placement affect attack effectiveness. Moreover, ToolHazard-generated alignment data improves security on both ToolHazard-Bench and AgentDojo while preserving benign task utility.