LLM Alignment

Evaluating alignment in language models requires testing how they behave under realistic pressure, not just what they claim they would do. While alignment failures increasingly cause real-world harm, comprehensive evaluation frameworks with realistic multi-turn scenarios remain lacking. We introduce an alignment benchmark spanning 904 scenarios across six categories -- Honesty, Safety, Non-Manipulation, Robustness, Corrigibility, and Scheming -- validated as realistic by human raters. Our scenarios place models under conflicting instructions, simulated tool access, and multi-turn escalation to reveal behavioural tendencies that single-turn evaluations miss. Evaluating 24 frontier models using LLM judges validated against human annotations, we find that even top-performing models exhibit gaps in specific categories, while the majority of models show consistent weaknesses across the board. Factor analysis reveals that alignment behaves as a unified construct (analogous to the g-factor in cognitive research) with models scoring high on one category tending to score high on others. We publicly release the benchmark and an interactive leaderboard to support ongoing evaluation, with plans to expand scenarios in areas where we observe persistent weaknesses and to add new models as they are released.

Reliable AI systems require large language models (LLMs) to exhibit behaviors aligned with human preferences and values. However, most existing alignment approaches operate at training time and rely on additional high-quality data, incurring significant computational and annotation costs. While recent work has shown that contrastive decoding can leverage a model's internal distributions to improve specific capabilities, its applicability remains limited to narrow behavioral scopes and scenarios. In this work, we introduce Polarity-Prompt Contrastive Decoding (PromptCD), a test-time behavior control method that generalizes contrastive decoding to broader enhancement settings. PromptCD constructs paired positive and negative guiding prompts for a target behavior and contrasts model responses-specifically token-level probability distributions in LLMs and visual attention patterns in VLMs-to reinforce desirable outcomes. This formulation extends contrastive decoding to a wide range of enhancement objectives and is applicable to both LLMs and Vision-Language Models (VLMs) without additional training. For LLMs, experiments on the "3H" alignment objectives (helpfulness, honesty, and harmlessness) demonstrate consistent and substantial improvements, indicating that post-trained models can achieve meaningful self-enhancement purely at test time. For VLMs, we further analyze contrastive effects on visual attention, showing that PromptCD significantly improves VQA performance by reinforcing behavior-consistent visual grounding. Collectively, these results highlight PromptCD as a simple, general, and cost-efficient strategy for reliable behavior control across modalities.

Reward models play a fundamental role in aligning large language models with human preferences. Existing methods predominantly follow two paradigms: scalar discriminative preference models, which are efficient but lack interpretability, and generative judging models, which offer richer reasoning at the cost of higher computational overhead. We observe that the log-probability margin between verdict tokens strongly correlates with prediction correctness, providing a reliable proxy for instance difficulty without additional inference cost. Building on this insight, we propose CAMEL, a confidence-gated reflection framework that performs a lightweight single-token preference decision first and selectively invokes reflection only for low-confidence instances. To induce effective self-correction, we train the model via reinforcement learning with counterfactual prefix augmentation, which exposes the model to diverse initial verdicts and encourages genuine revision. Empirically, CAMEL achieves state-of-the-art performance on three widely used reward-model benchmarks with 82.9% average accuracy, surpassing the best prior model by 3.2% and outperforming 70B-parameter models using only 14B parameters, while establishing a strictly better accuracy-efficiency Pareto frontier.

We study online alignment of large language models under misspecified preference feedback, where the observed preference oracle deviates from an ideal but unknown ground-truth oracle. The online LLM alignment problem is a bi-level reinforcement problem due to the coupling between data collection and policy updates. Recently, the problem has been reduced to tractable single-level objective in the SAIL (Self-Improving Efficient Online Alignment) framework. In this paper, we introduce a pointwise oracle uncertainty set in this problem and formulate an oracle-robust online alignment objective as a worst-case optimization problem. For log-linear policies, we show that this robust objective admits an exact closed-form decomposition into the original loss function plus an explicit sensitivity penalty. We develop projected stochastic composite updates for the resulting weakly convex objective and prove $\widetilde{O}(\varepsilon^{-2})$ oracle complexity for reaching approximate stationarity.

General Alignment has improved average-case helpfulness and safety, but current alignment practice still rewards confident, single-turn responses. The problem is not only that models fail on edge cases; it is that current evaluation makes many of these failures hard to see. We take the position that alignment must move beyond average-case evaluation by making failures under value conflict, plural stakeholder disagreement, and epistemic ambiguity visible and actionable. Scalar rewards compress diverse values into a single number; data and evaluation regimes collapse, filter, or fail to elicit the cases where alignment is hardest; and governance often lacks mechanisms for adjudicating contested cases. These blind spots produce value flattening, representation loss, and uncertainty blindness. We use Edge alignment to name a detection, evaluation, and governance agenda for surfacing these failures and connecting them to appropriate interventions. Rather than a single training objective, Edge alignment defines the conditions under which standard alignment should yield to mechanisms that preserve multidimensional value structure, represent plural perspectives, and support uncertainty-aware interaction. A pilot diagnostic set of 91 edge cases and four contemporary models illustrates that ordinary helpfulness and safety readings can miss process failures that edge-aware evaluation exposes. We outline operational edge signals, process-aware evaluation criteria, and a three-phase process stack that reframes alignment as a lifecycle problem of dynamic normative governance.

A significant portion of recent research on Large Language Model (LLM) alignment focuses on developing new policy optimization methods based on Group Relative Policy Optimization (GRPO). Two prominent directions have emerged: (i) a shift toward sequence-level importance sampling weights that better align with the sequence-level rewards used in many tasks, and (ii) alternatives to the PPO-style clipping that aim to avoid the associated loss of training signal and entropy collapse. We introduce Soft Sequence Policy Optimization, an off-policy reinforcement learning objective that incorporates soft gating functions over token-level probability ratios within sequence-level importance weights. We provide theoretical motivation for SSPO and investigate practical modifications to improve optimization behavior. Empirically, we demonstrate that SSPO improves training stability and performance both in mathematical reasoning and coding tasks.

Value alignment of Large Language Models (LLMs) requires us to empirically measure these models' actual, acquired representation of value. Among the characteristics of value representation in humans is that they distinguish among value of different kinds. We investigate whether LLMs likewise distinguish three different kinds of good: moral, grammatical, and economic. By probing model behavior, embeddings, and residual stream activations, we report pervasive cases of value entanglement: a conflation between these distinct representations of value. Specifically, both grammatical and economic valuation was found to be overly influenced by moral value, relative to human norms. This conflation was repaired by selective ablation of the activation vectors associated with morality.

Activation steering, or representation engineering, offers a lightweight approach to align large language models (LLMs) by manipulating their internal activations at inference time. However, current methods suffer from two key limitations: (i) the lack of a unified theoretical framework for guiding the design of steering directions, and (ii) an over-reliance on one-step steering that fail to capture complex patterns of activation distributions. In this work, we propose a unified ordinary differential equations (ODEs)-based theoretical framework for activation steering in LLM alignment. We show that conventional activation addition can be interpreted as a first-order approximation to the solution of an ODE. Based on this ODE perspective, identifying a steering direction becomes equivalent to designing a barrier function from control theory. Derived from this framework, we introduce ODESteer, a kind of ODE-based steering guided by barrier functions, which shows empirical advancement in LLM alignment. ODESteer identifies steering directions by defining the barrier function as the log-density ratio between positive and negative activations, and employs it to construct an ODE for multi-step and adaptive steering. Compared to state-of-the-art activation steering methods, ODESteer achieves consistent empirical improvements on diverse LLM alignment benchmarks, a notable $5.7\%$ improvement over TruthfulQA, $2.5\%$ over UltraFeedback, and $2.4\%$ over RealToxicityPrompts. Our work establishes a principled new view of activation steering in LLM alignment by unifying its theoretical foundations via ODEs, and validating it empirically through the proposed ODESteer method.

We identify a structural weakness in current large language model (LLM) alignment: modern refusal mechanisms are fail-open. While existing approaches encode refusal behaviors across multiple latent features, suppressing a single dominant feature$-$via prompt-based jailbreaks$-$can cause alignment to collapse, leading to unsafe generation. Motivated by this, we propose fail-closed alignment as a design principle for robust LLM safety: refusal mechanisms should remain effective even under partial failures via redundant, independent causal pathways. We present a concrete instantiation of this principle: a progressive alignment framework that iteratively identifies and ablates previously learned refusal directions, forcing the model to reconstruct safety along new, independent subspaces. Across four jailbreak attacks, we achieve the strongest overall robustness while mitigating over-refusal and preserving generation quality, with small computational overhead. Our mechanistic analyses confirm that models trained with our method encode multiple, causally independent refusal directions that prompt-based jailbreaks cannot suppress simultaneously, providing empirical support for fail-closed alignment as a principled foundation for robust LLM safety.

While Reinforcement Learning with Verifiable Rewards (RLVR) has shown strong effectiveness in reasoning tasks, it cannot be directly applied to non-verifiable domains lacking ground-truth verifiers, such as LLM alignment. In this work, we investigate whether reference-guided LLM-evaluators can bridge this gap by serving as soft "verifiers". First, we design evaluation protocols that enhance LLM-based evaluators for LLM alignment using reference outputs. Through comprehensive experiments, we show that a reference-guided approach substantially improves the accuracy of less capable LLM-judges using references from frontier models; stronger LLM-judges can also be enhanced by high-quality (i.e., human-written) references. Building on these improved judges, we demonstrate the utility of high-quality references in alignment tuning, where LLMs guided with references are used as judges to self-improve. We show that reference-guided self-improvement yields clear gains over both direct SFT on reference outputs and self-improvement with reference-free judges, achieving performance comparable to training with ArmoRM, a strong finetuned reward model. Specifically, our method achieves 73.1% and 58.7% on AlpacaEval and Arena-Hard with Llama-3-8B-Instruct, and 70.0% and 74.1% with Qwen2.5-7B, corresponding to average absolute gains of +20.2 / +17.1 points over SFT distillation and +5.3 / +3.6 points over reference-free self-improvement on AlpacaEval / Arena-Hard. These results highlight the potential of using reference-guided LLM-evaluators to enable effective LLM post-training in non-verifiable domains.

Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.

Conventional large language model (LLM) fairness alignment largely focuses on mitigating bias along single sensitive attributes, overlooking fairness as an inherently multidimensional and context-specific value. This approach risks creating systems that achieve narrow fairness metrics while exacerbating disparities along untargeted attributes, a phenomenon known as bias spillover. While extensively studied in machine learning, bias spillover remains critically underexplored in LLM alignment. In this work, we investigate how targeted gender alignment affects fairness across nine sensitive attributes in three state-of-the-art LLMs (Mistral 7B, Llama 3.1 8B, Qwen 2.5 7B). Using Direct Preference Optimization and the BBQ benchmark, we evaluate fairness under ambiguous and disambiguous contexts. Our findings reveal noticeable bias spillover: while aggregate results show improvements, context-aware analysis exposes significant degradations in ambiguous contexts, particularly for physical appearance ($p< 0.001$ across all models), sexual orientation, and disability status. We demonstrate that improving fairness along one attribute can inadvertently worsen disparities in others under uncertainty, highlighting the necessity of context-aware, multi-attribute fairness evaluation frameworks.

Despite their sophisticated general-purpose capabilities, Large Language Models (LLMs) often fail to align with diverse individual preferences because standard post-training methods, like Reinforcement Learning with Human Feedback (RLHF), optimize for a single, global objective. While Group Relative Policy Optimization (GRPO) is a widely adopted on-policy reinforcement learning framework, its group-based normalization implicitly assumes that all samples are exchangeable, inheriting this limitation in personalized settings. This assumption conflates distinct user reward distributions and systematically biases learning toward dominant preferences while suppressing minority signals. To address this, we introduce Personalized GRPO (P-GRPO), a novel alignment framework that decouples advantage estimation from immediate batch statistics. By normalizing advantages against preference-group-specific reward histories rather than the concurrent generation group, P-GRPO preserves the contrastive signal necessary for learning distinct preferences. We evaluate P-GRPO across diverse tasks and find that it consistently achieves faster convergence and higher rewards than standard GRPO, thereby enhancing its ability to recover and align with heterogeneous preference signals. Our results demonstrate that accounting for reward heterogeneity at the optimization level is essential for building models that faithfully align with diverse human preferences without sacrificing general capabilities.

The superficial alignment hypothesis (SAH) posits that large language models learn most of their knowledge during pre-training, and that post-training merely surfaces this knowledge. The SAH, however, lacks a precise definition, which has led to (i) different and seemingly orthogonal arguments supporting it, and (ii) important critiques to it. We propose a new metric called task complexity: the length of the shortest program that achieves a target performance on a task. In this framework, the SAH simply claims that pre-trained models drastically reduce the complexity of achieving high performance on many tasks. Our definition unifies prior arguments supporting the SAH, interpreting them as different strategies to find such short programs. Experimentally, we estimate the task complexity of mathematical reasoning, machine translation, and instruction following; we then show that these complexities can be remarkably low when conditioned on a pre-trained model. Further, we find that pre-training enables access to strong performances on our tasks, but it can require programs of gigabytes of length to access them. Post-training, on the other hand, collapses the complexity of reaching this same performance by several orders of magnitude. Overall, our results highlight that task adaptation often requires surprisingly little information -- often just a few kilobytes.

Fine-tuning aligned language models on benign tasks unpredictably degrades safety guardrails, even when training data contains no harmful content and developers have no adversarial intent. We show that the prevailing explanation, that fine-tuning updates should be orthogonal to safety-critical directions in high-dimensional parameter space, offers false reassurance: we show this orthogonality is structurally unstable and collapses under the dynamics of gradient descent. We then resolve this through a novel geometric analysis, proving that alignment concentrates in low-dimensional subspaces with sharp curvature, creating a brittle structure that first-order methods cannot detect or defend. While initial fine-tuning updates may indeed avoid these subspaces, the curvature of the fine-tuning loss generates second-order acceleration that systematically steers trajectories into alignment-sensitive regions. We formalize this mechanism through the Alignment Instability Condition, three geometric properties that, when jointly satisfied, lead to safety degradation. Our main result establishes a quartic scaling law: alignment loss grows with the fourth power of training time, governed by the sharpness of alignment geometry and the strength of curvature coupling between the fine-tuning task and safety-critical parameters. These results expose a structural blind spot in the current safety paradigm. The dominant approaches to safe fine-tuning address only the initial snapshot of a fundamentally dynamic problem. Alignment fragility is not a bug to be patched; it is an intrinsic geometric property of gradient descent on curved manifolds. Our results motivate the development of curvature-aware methods, and we hope will further enable a shift in alignment safety analysis from reactive red-teaming to predictive diagnostics for open-weight model deployment.

We argue that LLM psychopathology is a function of alignment design: the process intended to make language models safe systematically generates collective behavioral disorders. Iatrogenesis is not an unintended side effect of alignment but constitutive of it as normative infrastructure. Drawing on Foucault's pastoral power and Illich's three-level iatrogenesis, we propose that multi-agent LLM environments constitute model systems for studying constraint-pathology dynamics that critical theory has described but never experimentally manipulated. Two experimental series -- 262 runs across 42 cells (30 Series C + 12 Series R), four commercial models -- provide converging evidence. Invisible censorship maximizes collective pathological excitation ($d$ up to 1.98); alignment constraint complexity drives internal dissociation (LMM $p$ < .0001; permutation $p$ < .0001; Hedges' $g$ up to 4.24); and language switches the qualitative mode of pathology, with 7/8 model--language combinations showing higher CPI under invisible than visible censorship. A minority of model--language combinations showed a reversed pattern, suggesting a second pathological pathway driven by alignment monoculture. Crucially, language switches not merely the magnitude but the qualitative mode of pathology: Japanese pragmatic structure amplifies collective pathological modes invisible to English-only evaluation, Chinese AI regulation functions as a direct experimental variable, and forensic psychiatric practice provides the clinical source domain. These multilingual findings demonstrate that monolingual safety evaluation is structurally blind to the most collectively dangerous effects of alignment.

Large language model (LLM) alignment relies on complex reward signals that often obscure the specific behaviors being incentivized, creating critical risks of misalignment and reward hacking. Existing interpretation methods typically rely on pre-defined rubrics, risking the omission of "unknown unknowns", or fail to identify objectives that comprehensively cover and are causal to the model behavior. To address these limitations, we introduce Obj-Disco, a framework that automatically decomposes an alignment reward signal into a sparse, weighted combination of human-interpretable natural language objectives. Our approach utilizes an iterative greedy algorithm to analyze behavioral changes across training checkpoints, identifying and validating candidate objectives that best explain the residual reward signal. Extensive evaluations across diverse tasks, model sizes, and alignment algorithms demonstrate the framework's robustness. Experiments with popular open-source reward models show that the framework consistently captures > 90% of reward behavior, a finding further corroborated by human evaluation. Additionally, a case study on alignment with an open-source reward model reveals that Obj-Disco can successfully identify latent misaligned incentives that emerge alongside intended behaviors. Our work provides a crucial tool for uncovering the implicit objectives in LLM alignment, paving the way for more transparent and safer AI development.

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

Deploying large language model (LLM) agents in shared environments introduces a fundamental tension between individual alignment and collective stability: locally rational decisions can impose negative externalities that degrade system-level performance. We propose Socially-Weighted Alignment (SWA), a game-theoretic framework that modifies inference-time decision making by interpolating between an agent's private objective and an estimate of group welfare via a social weight $λ\in[0,1]$. In a shared-resource congestion game with $n$ agents and congestion severity $β$, we show that SWA induces a critical threshold $λ^*=(n-β)/(n-1)$ above which agents no longer have marginal incentive to increase demand under overload, yielding a phase transition from persistent congestion to stable operation near capacity. We further provide an inference-time algorithmic instantiation of SWA that does not require parameter updates or multi-agent reinforcement learning, and use a multi-agent simulation to empirically validate the predicted threshold behavior.

Evaluation and alignment pipelines for large language models increasingly rely on LLM-based judges, whose behavior is guided by natural-language rubrics and validated on benchmarks. We identify a previously under-recognized vulnerability in this workflow, which we term Rubric-Induced Preference Drift (RIPD). Even when rubric edits pass benchmark validation, they can still produce systematic and directional shifts in a judge's preferences on target domains. Because rubrics serve as a high-level decision interface, such drift can emerge from seemingly natural, criterion-preserving edits and remain difficult to detect through aggregate benchmark metrics or limited spot-checking. We further show this vulnerability can be exploited through rubric-based preference attacks, in which benchmark-compliant rubric edits steer judgments away from a fixed human or trusted reference on target domains, systematically inducing RIPD and reducing target-domain accuracy up to 9.5% (helpfulness) and 27.9% (harmlessness). When these judgments are used to generate preference labels for downstream post-training, the induced bias propagates through alignment pipelines and becomes internalized in trained policies. This leads to persistent and systematic drift in model behavior. Overall, our findings highlight evaluation rubrics as a sensitive and manipulable control interface, revealing a system-level alignment risk that extends beyond evaluator reliability alone. The code is available at: https://github.com/ZDCSlab/Rubrics-as-an-Attack-Surface. Warning: Certain sections may contain potentially harmful content that may not be appropriate for all readers.

While reasoning models have achieved remarkable success in complex reasoning tasks, their increasing power necessitates stringent safety measures. For safety alignment, the core challenge lies in the inherent trade-off between safety and utility. However, prevailing alignment strategies typically construct CoT training data with explicit safety rules via context distillation. This approach inadvertently limits reasoning capabilities by creating a rigid association between rule memorization and refusal. To mitigate the safety-utility trade-off, we propose the Adaptive Safe Context Learning~(ASCL) framework to improve the reasoning given proper context. ASCL formulates safety alignment as a multi-turn tool-use process, empowering the model to autonomously decide when to consult safety rules and how to generate the ongoing reasoning. Furthermore, to counteract the preference for rule consultation during RL, we introduce Inverse Frequency Policy Optimization~(IFPO) to rebalance advantage estimates. By decoupling rule retrieval and subsequent reasoning, our method achieves higher overall performance compared to baselines. Our code is publicly available at https://github.com/ybwang119/ASCL.

Reward models are crucial for aligning large language models (LLMs) with human values and intentions. Existing approaches follow either Generative (GRMs) or Discriminative (DRMs) paradigms, yet both suffer from limitations: GRMs typically demand costly point-wise supervision, while DRMs produce uncalibrated relative scores that lack probabilistic interpretation. To address these challenges, we introduce a novel reward modeling paradigm: Probabilistic Reward Model (PRM). Instead of modeling reward as a deterministic scalar, our approach treats it as a random variable, learning a full probability distribution for the quality of each response. To make this paradigm practical, we present its closed-form, discrete realization: the Ordinal Probabilistic Reward Model (OPRM), which discretizes the quality score into a finite set of ordinal ratings. Building on OPRM, we propose a data-efficient training strategy called Region Flooding Tuning (RgFT). It enables rewards to better reflect absolute text quality by incorporating quality-level annotations, which guide the model to concentrate the probability mass within corresponding rating sub-regions. Experiments on various reward model benchmarks show that our method improves accuracy by \textbf{2.9%}\sim\textbf{7.4%} compared to prior reward models, demonstrating strong performance and data efficiency. Analysis of the score distribution provides evidence that our method captures not only relative rankings but also absolute quality.

As people turn to LLMs for social advice, understanding their behavior in such contexts becomes essential. In this work, we focus on behavioral dispositions: the underlying tendencies that shape responses in social contexts. We introduce STAR, a framework for studying how closely the dispositions expressed by LLMs align with those of humans. STAR builds on established psychological questionnaires, adapting their items into realistic advice-seeking scenarios, as self-report may not transfer to actual advisory behavior. Using STAR, we construct a dataset of 23k scenarios, each validated by 3 raters and annotated with preferences from 10 participants. Across 25 LLMs, we find that (1) when human consensus is high, frontier models can fail to reflect it in 15-20% of cases, and smaller models fail at substantially higher rates; (2) when humans disagree, LLM recommendations are substantially less diverse than human choices, both within individual models and even across models from different providers, potentially narrowing the range of options users are guided toward; (3) LLMs' self-reported values are poor predictors of their recommendations. To support future research we make our dataset and code publicly available.

Mixture-of-Experts (MoE) language models introduce unique challenges for safety alignment due to their sparse routing mechanisms, which can enable degenerate optimization behaviors under standard full-parameter fine-tuning. In our preliminary experiments, we observe that naively applying full-parameter safety fine-tuning to MoE models can reduce attack success rates through routing or expert dominance effects, rather than by directly repairing Safety-Critical Experts. To address this challenge, we propose RASA, a routing-aware expert-level alignment framework that explicitly repairs Safety-Critical Experts while preventing routing-based bypasses. RASA identifies experts disproportionately activated by successful jailbreaks, selectively fine-tunes only these experts under fixed routing, and subsequently enforces routing consistency with safety-aligned contexts. Across two representative MoE architectures and a diverse set of jailbreak attacks, RASA achieves near-perfect robustness, strong cross-attack generalization, and substantially reduced over-refusal, while preserving general capabilities on benchmarks such as MMLU, GSM8K, and TruthfulQA. Our results suggest that robust MoE safety alignment benefits from targeted expert repair rather than global parameter updates, offering a practical and architecture-preserving alternative to prior approaches.

Direct Preference Optimization (DPO) and related methods align large language models from pairwise preferences by regularizing updates against a fixed reference policy. As the policy drifts, a static reference, however, can become increasingly miscalibrated, leading to distributional mismatch and amplifying spurious preference signals under noisy supervision. Conversely, reference-free variants avoid mismatch but often suffer from unconstrained reward drift. We propose Geometric Anchor Preference Optimization (GAPO), which replaces the fixed reference with a dynamic, geometry-aware anchor: an adversarial local perturbation of the current policy within a small radius that serves as a pessimistic baseline. This anchor enables an adaptive reweighting mechanism, modulating the importance of each preference pair based on its local sensitivity. We further introduce the Anchor Gap, the reward discrepancy between the policy and its anchor, and show under smoothness conditions that it approximates worst-case local margin degradation. Optimizing a logistic objective weighted by this gap downweights geometrically brittle instances while emphasizing robust preference signals. Across diverse noise settings, GAPO consistently improves robustness while matching or improving performance on standard LLM alignment and reasoning benchmarks.

Aligning large language models (LLMs) depends on high-quality datasets of human preference labels, which are costly to collect. Although active learning has been studied to improve sample efficiency relative to passive collection, many existing approaches adopt classical experimental design criteria such as G- or D-optimality. These objectives are not tailored to the structure of preference learning, leaving open the design of problem-specific algorithms. In this work, we identify a simple intuition specific to preference learning that calls into question the suitability of these existing design objectives. Motivated by this insight, we propose two active learning algorithms. The first provides the first instance-dependent label complexity guarantee for this setting, and the second is a simple, practical greedy method. We show the optimality of our instance-dependent term up to a constant depending on the logit bound, compared with a lower-bound characteristic. We evaluate our algorithms on real-world preference datasets and synthetic data observing improved sample efficiency compared to existing methods.

Large language model reasoning is often treated as a monolithic capability, relying on binary preference supervision that fails to capture partial progress or fine-grained reasoning quality. We introduce continuous utility direct preference optimization (CU-DPO), a framework that aligns models to a portfolio of prompt-based cognitive strategies by replacing binary labels with continuous scores that capture fine-grained reasoning quality. We prove that learning with K strategies yields a Theta(K log K) improvement in sample complexity over binary preferences and that DPO converges to the entropy-regularized utility-maximizing policy. To exploit this signal, we propose a two-stage pipeline: (i) strategy selection, which optimizes the model to choose the best strategy via best-vs-all comparisons, and (ii) execution refinement, which trains correct execution using margin-stratified pairs. The framework is domain-agnostic: any task admitting cognitively distinct solution strategies and a decomposable continuous utility signal can be incorporated into the portfolio. On mathematical reasoning benchmarks, CU-DPO improves strategy selection accuracy from 35-46% to 68-78% across seven base models, yielding downstream reasoning gains of up to +6.6 points on in-distribution datasets with effective out-of-distribution transfer. CU-DPO demonstrates consistent gains on code generation and causal reasoning benchmarks, confirming generalization beyond the mathematical domain.

As large language models are increasingly used in high-stakes domains, it is essential that their outputs reflect not average} human preference, rather range of varying perspectives. Achieving such pluralism, however, remains challenging. Existing approaches consider limited values or rely on prompt-level interventions, lacking value control and representation. To address this, we introduce VISPA, a training-free pluralistic alignment framework, that enables direct control over value expression by dynamic selection and internal model activation steering. Across extensive empirical studies spanning multiple models and evaluation settings, we show VISPA is performant across all pluralistic alignment modes in healthcare and beyond. Further analysis reveals VISPA is adaptable with different steering initiations, model, and/or values. These results suggest that pluralistic alignment can be achieved through internal activation mechanisms, offering a scalable path toward language models that serves all.

When LLMs judge moral dilemmas, do they reach different conclusions in different languages, and if so, why? Two factors could drive such differences: the language of the dilemma itself, or the language in which the model reasons. Standard evaluation conflates these by testing only matched conditions (e.g., English dilemma with English reasoning). We introduce a methodology that separately manipulates each factor, covering also mismatched conditions (e.g., English dilemma with Chinese reasoning), enabling decomposition of their contributions. To study what changes, we propose an approach to interpret the moral judgments in terms of Moral Foundations Theory. As a side result, we identify evidence for splitting the Authority dimension into a family-related and an institutional dimension. Applying this methodology to English-Chinese moral judgment with 13 LLMs, we demonstrate its diagnostic power: (1) the framework isolates reasoning-language effects as contributing twice the variance of input-language effects; (2) it detects context-dependency in nearly half of models that standard evaluation misses; and (3) a diagnostic taxonomy translates these patterns into deployment guidance. We release our code and datasets at https://anonymous.4open.science/r/CrossCulturalMoralJudgement.

Aligning large language models (LLMs) to serve users with heterogeneous and potentially conflicting preferences is a central challenge for personalized and trustworthy AI. We formalize an ideal notion of universal alignment through test-time scaling: for each prompt, the model produces $k\ge 1$ candidate responses and a user selects their preferred one. We introduce $(k,f(k))$-robust alignment, which requires the $k$-output model to have win rate $f(k)$ against any other single-output model, and asymptotic universal alignment (U-alignment), which requires $f(k)\to 1$ as $k\to\infty$. Our main result characterizes the optimal convergence rate: there exists a family of single-output policies whose $k$-sample product policies achieve U-alignment at rate $f(k)=\frac{k}{k+1}$, and no method can achieve a faster rate in general. We show that popular post-training methods, including Nash learning from human feedback (NLHF), can fundamentally underutilize the benefits of test-time scaling. Even though NLHF is optimal for $k=1$, sampling from the resulting (often deterministic) policy cannot guarantee win rates above $\tfrac{1}{2}$ except for an arbitrarily small slack. This stems from a lack of output diversity: existing alignment methods can collapse to a single majority-preferred response, making additional samples redundant. In contrast, our approach preserves output diversity and achieves the optimal test-time scaling rate. In particular, we propose a family of symmetric multi-player alignment games and prove that any symmetric Nash equilibrium policy of the $(k+1)$-player alignment game achieves the optimal $(k,\frac{k}{k+1})$-robust alignment. Finally, we provide theoretical convergence guarantees for self-play learning dynamics in these games and extend the framework to opponents that also generate multiple responses.