ICML 2026 Papers

According to the reformulated Learned Helplessness theory, repeated exposure to uncontrollable negative events can foster a depressogenic attributional style—increasing susceptibility to depression yet remaining a tractable target for cognitive therapy. Computational research on attributional cognition, however, is hampered by the lack of large-scale datasets and robust evaluation protocols. In this work, we introduce the Attributional Style Transfer Dataset (ASTD) along with dedicated evaluation metrics, the first benchmark designed to model, assess, and reframe attributional explanations at scale. Constructed via a Prevent–Filter–Validate pipeline that integrates LLM-based generation with specialist validation, ASTD contains 42,000 real-world events paired with psychologically grounded attributions spanning seven styles. Using this dataset, we address two key challenges: (1) scalable assessment of attributional style via both supervised classifiers and zero/few-shot LLMs; and (2) attributional reframing and evaluation, where we propose automatic evaluation metrics to quantify psychological validity. Furthermore, we leverage our proposed metrics to construct a preference dataset, fine-tuning LLMs with Direct Preference Optimization (DPO) and achieving substantial gains in reframing quality. Together, our dataset, metrics, and methodology offer a new paradigm for understanding and modeling attributional style, with direct implications for scalable and adaptive mental health interventions.

Learning on graphs with missing node attributes is a prevalent yet challenging problem in real-world scenarios, as graph neural networks (GNNs) typically rely on complete attribute information. Existing solutions often employ adversarial learning in a shared latent space to align graph structure and attributes. However, these methods frequently suffer from training instability and mode collapse, failing to fully capture the complex, multi-modal joint distribution of topology and features. To address these limitations, we present GLAD, a novel generative framework for robust node attribute completion. GLAD leverages the strong generative capabilities of diffusion models to learn the conditional distribution of attributes given the graph structure within a decoupled latent space. Unlike previous unidirectional approaches, GLAD introduces a robust bidirectional alignment mechanism. Specifically, we incorporate a structure reconstruction constraint during training and structure-aware classifier-free guidance during sampling, ensuring that generated attributes are not only plausible but also maintain strict topological consistency with the underlying graph. Theoretically, we show that GLAD maximizes a tighter variational lower bound on the joint log-likelihood compared to GAN-based predecessors, leading to superior mode coverage. Extensive experiments on large-scale benchmarks demonstrate that GLAD significantly outperforms state-of-the-art baselines in both attribute recovery quality and downstream task performance.

Federated learning (FL) allows clients to collaboratively train models without exposing private data, but practical FL is simultaneously challenged by data heterogeneity and model heterogeneity. Prior heterogeneous FL (HtFL) approaches often fail to handle fine-grained feature shifts, leading to weak representation alignment and limited cross-client knowledge transfer, which degrades both personalization and generalization. We propose FedARC, an HtFL framework that couples a shared lightweight extractor with client-specific fusion: a trainable projector integrates local and global embeddings, while adaptive residual compensation dynamically corrects feature-level mismatches. To further stabilize aggregation, FedARC performs semantic anchor alignment across clients, and we theoretically prove FedARC converges with a non-convex convergence rate $\mathcal{O}(1/T)$. Experiments on five public benchmarks demonstrate that FedARC outperforms nine state-of-the-art HtFL baselines by up to 2.63\% in average accuracy, while maintaining efficient communication and computation.

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

We study the decoupled multi-armed bandit problem, where the learner separately selects one arm for exploration and one, possibly different, arm for exploitation at each round. In this setting, the loss of the explored arm is observed but not incurred, whereas the loss of the exploited arm is incurred without being observed. We propose an efficient Follow-the-Perturbed-Leader (FTPL) policy that achieves Best-of-Both-Worlds (BOBW) guarantee with constant regret in the stochastic regime and optimal $\mathcal{O}(\sqrt{KT})$ regret in the adversarial regime. A key feature of our method is that it completely avoids both the convex optimization required by prior BOBW policies and the resampling procedures typically used in FTPL bandit policies. This allows FTPL to fully realize its computational efficiency advantages, leading to substantial reductions in computational cost. We empirically confirm that our policy not only improves the runtime but also demonstrates superior regret performance in both regimes.

Large language models can represent a variety of personas but typically default to a helpful Assistant identity cultivated during post-training. Across several different models, we find an “Assistant Axis" in their activation space, which captures the extent to which a model is operating in its default Assistant mode. Steering towards the Assistant direction reinforces helpful and harmless behavior; steering away increases the model’s tendency to identify as other entities. Measuring deviations along the Assistant Axis predicts “persona drift,” a phenomenon where models slip into exhibiting harmful or bizarre behaviors that are uncharacteristic of their typical persona. We find that persona drift is often driven by conversations demanding meta-reflection on the model’s processes or featuring emotionally vulnerable users. We show that restricting activations to a fixed region along the Assistant Axis can stabilize model behavior in these scenarios—and also in the face of adversarial persona-based jailbreaks. Our results suggest that post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent persona.

Computational protein design typically employs a sequential workflow of structure generation followed by sequence (re)design. While structure generators can be explicitly conditioned on functional objectives, inverse folding models are constrained by their function-agnostic nature and sequence-structure degeneracy. More critically, the associated training objectives do not account for the *Best-of-N* (BoN) inference protocol, resulting in a fundamental training-inference misalignment. Here, we propose FIDIA, a reinforcement learning framework that enables **F**unction-**I**nformed sequence **D**esign via **I**nference-**A**ligned policy optimization. Specifically, FIDIA integrates functional constraints into composite rewards and explicitly optimize the induced policy under BoN toward high-fitness sequence regions. We achieve this via a grounded gradient estimator that directly maximizes the expected maximum reward. FIDIA consistently outperforms both standard and RL-optimized baselines in success rate and precision on a general motif scaffolding benchmark. Further experiments on realworld cases including vaccine and affinity-enhancing enzyme design validate FIDIA’s efficacy in complex therapeutic and biocatalytic contexts.

The success of large language models (LLMs) in scientific domains has heightened safety concerns, prompting numerous benchmarks to evaluate their scientific safety. Existing benchmarks often suffer from limited risk coverage and a reliance on subjective evaluation. To address these problems, we introduce SafeSci, a comprehensive framework for safety evaluation and enhancement in scientific contexts. SafeSci comprises SafeSciBench, a multi-disciplinary benchmark with 0.25M samples, and SafeSciTrain, a large-scale dataset containing 1.5M samples for safety enhancement. SafeSciBench distinguishes between safety knowledge and risk to cover extensive scopes and employs objective metrics such as deterministically answerable questions to mitigate evaluation bias. We evaluate 24 advanced LLMs and reveal critical vulnerabilities. We also observe that LLMs exhibit varying degrees of excessive refusal behaviors on safety-related issues. For safety enhancement, we demonstrate that fine-tuning on SafeSciTrain significantly enhances the safety alignment of LLMs. Finally, we argue that determining the safety of a scientific question should depend on the specific context, rather than universally categorizing it as safe or unsafe.

We introduce a multi-scale Image Super Resolution (ISR) method building on recent advances in Visual Auto-Regressive (VAR) modeling. VAR models break image tokenization into additive, gradually increasing scales, using Residual Quantization (RQ), an approach that aligns perfectly with our target ISR task. Previous works taking advantage of this synergy suffer from two main shortcomings. First, due to the limitations in RQ, they only generate images at a predefined fixed scale, failing to map intermediate outputs to the corresponding image scales. They also rely on large backbones or a large corpus of annotated data to achieve better performance. To address both shortcomings, we introduce two novel components to the VAR training for ISR, aiming at increasing its flexibility and reducing its complexity. In particular, we introduce a) a **Hierarchical Image Tokenization (HIT)** approach that progressively represents images at different scales while enforcing token overlap across scales, and b) a **Direct Preference Optimization (DPO) regularization term** that, relying solely on the (LR,HR) pair, encourages the transformer to produce the latter over the former. Our proposed HIT acts as a strong inductive bias for the VAR training, resulting in a small model (300M params vs 1B params of VARSR), that achieves state-of-the-art results without external training data, and that delivers multi-scale outputs with a single forward pass.

Foundation models have driven rapid progress in computer vision, yet the two dominant paradigms, vision-language foundation models (VLMs) and vision-only foundation models (VFMs), remain only partially compatible. VLMs offer language-grounded semantic alignment but are often visually coarse, while VFMs learn discriminative perceptual geometry but lack semantic grounding. We propose **GPUA**, a *Geometry-Preserving Unsupervised Alignment* framework that integrates the complementary strengths of VFMs and VLMs. Inspired by cross-lingual alignment, **GPUA** treats VFM features as a *visual language* and learns an *orthogonal* mapping that translates the VFM space into the VLM semantic space, preserving geometry and narrowing the modality gap without labels or model parameter updates. **GPUA** is task-agnostic and requires only feature-level access to pretrained models. Experiments across diverse benchmarks demonstrate improved cross-model compatibility and strong gains in downstream zero-shot recognition and segmentation with negligible overhead. Our code is available at: https://github.com/Yuteam14/GPUA.

The growing capabilities of large language models (LLMs) have driven their widespread deployment across diverse domains, even in potentially high-risk scenarios. Despite advances in safety alignment techniques, current models remain vulnerable to emerging *persona-based jailbreak attacks*. Existing research on persona-based jailbreak has primarily focused on attack iterations, yet it lacks systemic and mechanistic constraints on the defense side. To address this challenge, we propose Persona-Invariant Alignment (PIA), an adversarial self-play framework that achieves co-evolution through Persona Lineage Evolution (PLE) on the attack side and Persona-Invariant Consistency Learning (PICL) on the defense side. Theoretically, PICL is grounded in the *structural separation hypothesis*, using a unilateral KL-divergence constraint to enable the structural decoupling of safety decisions from persona context, thereby maintaining safe behavior under persona-based jailbreak attacks. Experimental results demonstrate that PLE efficiently explores high-risk persona spaces by leveraging lineage-based credit propagation. Meanwhile, the PICL defense method significantly reduces the Attack Success Rate (ASR) while preserving the model's general capability, thereby validating the superiority and robustness of this alignment paradigm. WARNING: This paper contains potentially offensive and harmful text.

We focus on the problem of \emph{Answer-Level Fine-Tuning} (ALFT), where the goal is to optimize a language model based on the correctness or properties of its final answers, rather than the specific reasoning traces used to produce them. Directly optimizing answer-level objectives is computationally intractable due to the need to marginalize over the vast space of latent reasoning paths. To overcome this, we propose a general game-theoretical framework that lifts the problem to a *Distributional Alignment Game*. We formulate ALFT as a two-player game between a Policy (the generator) and a Target (an auxiliary distribution). We prove that the Nash Equilibrium of this game corresponds exactly to the solution of the original answer-level optimization problem. This variational perspective transforms the intractable marginalization problem into a tractable projection problem. We demonstrate that this framework unifies recent approaches to diversity and self-improvement (coherence) and provide efficient algorithms compatible with Group Relative Policy Optimization (GRPO), such as *Coherence-GRPO*, yielding significant complexity gains in mathematical reasoning tasks.

Mixture-of-Experts (MoE) models scale compute efficiently, yet they remain expensive to deploy due to substantial memory footprint and inference overhead. Prior methods mainly operate at the expert level, either removing whole experts or ranking experts by importance. However, such expert-wise decisions are too coarse to identify redundancy, and often misallocate pruning budgets and limits compression. This issue worsens in large MoEs with dynamic routing and heterogeneous experts. To alleviate this dilemma, we for the first time observe that information in MoE experts is highly concentrated in a few channels, leaving substantial redundancy even in "high importance" experts. Accordingly, we propose a structural pruning framework tailored for MoEs, reforming the prune-ratio objective to maximizing channel-score coverage via an efficient attribution-based approximation. Experiments on DeepSeek and Qwen MoEs retain accuracy under 50\% or 25\% pruning joinly with 4-bit quantization, reducing the memory footprint of Qwen3-30B-A3B by 5.27$\times$, and outperforming state-of-the-art baselines under diverse benchmarks.

Large language models (LLMs) have demonstrated remarkable performance on standard benchmarks, yet it remains largely unexplored whether they truly meet user expectations. Existing evaluation approaches, relying on model heuristics, expert rubrics, or user simulation, fail to capture the diversity and subtlety of real human expectations, causing models to appear competent while misaligning with what users actually seek. we present the first systematic study of user expectations in real-world LLM interactions, proposing a principled procedure to extract semantically rich expectations and introducing ExpectBench, a benchmark grounded in real user expectations. Analyses reveal that current LLMs struggle to satisfy and anticipate what users hope to obtain, highlighting a fundamental source of misalignment. Building on these observations, we propose LENS, a lightweight latent expectation–aware response generation framework. LENS enables models to internalize user expectations and generate better-aligned responses, consistently improving expectation satisfaction and underscoring the importance of explicitly modeling user expectations for realistic human–AI alignment.

Assistive agents should not only take actions on behalf of a human, but also step out of the way and cede control when there are important decisions to be made. However, current methods for building assistive agents, whether via mimicking expert humans or via RL finetuning on an inferred reward, often encourage agents to complete tasks on their own rather than truly assisting the human achieve her objectives. We propose a new approach to tuning assistive language models based on maximizing the human's empowerment, their ability to effect desired changes in the environment. Our empowerment-maximizing method, Empower, only requires offline examples of human-written text, providing a self-supervised method for fine-tuning language models to better assist humans. To study the efficacy of our approach, we conducted an 18-person user study comparing our empowerment assistant with a strong baseline. Participants preferred our assistant 78% of the time (p=0.015), with a 31% higher acceptance rate. With this empowerment objective, we provide a framework for useful aligned AI agents without the need for human feedback or verifiable rewards.

It is known that RandomShuffle, the without replacement version of Stochastic Gradient Descent (SGD), converges faster than with replacement SGD. However, RandomShuffle requires uniformly performing a random permutation of the input sequence, which is known to have high I/O complexity due to data movement across the memory hierarchy. In this paper, we propose a shuffling algorithm with a linear I/O complexity that generates almost-uniformly random permutations with rigorous mathematical guarantees. Specifically, we show that the shuffling algorithm can generate $2$-wise independent permutations. Furthermore, we can extend to $k$-wise independence with a small error in the probability distribution, if the fast memory has at least $k$ memory blocks. These results allow us to reach the same expected theoretical convergence as RandomShuffle while achieving optimal linear I/O cost.

In automated machine learning, scientific discovery, and other applications of Bayesian optimization, deciding when to stop evaluating expensive black-box functions in a cost-aware manner is an important but underexplored practical consideration. A natural performance metric for this purpose is the *cost-adjusted simple regret*, which explicitly captures the trade-off between solution quality and cumulative evaluation cost. Existing stopping rules for Bayesian optimization are either heuristic, or are theoretically grounded but designed to optimize simple regret without accounting for evaluation costs; as a result, they provide no guarantees against unnecessary evaluations when costs are high. We propose a *principled cost-aware stopping rule* for Bayesian optimization that adapts to varying evaluation costs without heuristic tuning. Our rule is grounded in a theoretical connection to state-of-the-art cost-aware acquisition functions, namely the Pandora's Box Gittins Index (PBGI) and log expected improvement per cost (LogEIPC). When paired with either acquisition function, we prove that the resulting policy satisfies a theoretical guarantee bounding the expected cost-adjusted simple regret. Across synthetic tasks and empirical benchmarks including hyperparameter optimization and neural architecture size search, pairing our stopping rule with PBGI or LogEIPC usually matches or outperforms other acquisition-function--stopping-rule pairs in terms of cost-adjusted simple regret.

Adapting Large Language Models (LLMs) to specialized domains typically incurs high data and computational overhead. While prior efficiency efforts have largely treated data selection and parameter-efficient fine-tuning as isolated processes, our empirical analysis suggests they may be intrinsically coupled. We posit the Strong Map Hypothesis: a sparse subset of attention heads plays a dominant role in task-specific adaptation, acting as keys that unlock specific data patterns. Building on this observation, we propose From Parameters to Data (P2D), a unified framework that leverages these task-sensitive attention heads as a dual compass for both sample mining and structural pruning. To rigorously quantify the total pipeline cost, we introduce the Alignment Efficiency Ratio (AER) metric for both selection latency and training time. Mechanistically, P2D identifies critical heads via a lightweight proxy and uses them as a functional filter to curate high-affinity data, establishing a synergistic pipeline. Empirically, by updating merely 10% of attention heads on 10% of the data, P2D achieves an 8.3 pp performance gain over strong baselines and delivers a 7.0x end-to-end time speedup. These results validate that precise parameter-data synchronization eliminates redundancy, offering a new paradigm for efficient alignment.

We study post-training linear autoregressive models with outcome and process rewards. Given a context $x$, the model must predict the response $y \in \mathcal{Y}^N$, a sequence of length $N$ that satisfies a $\gamma$ margin condition, an extension of the standard separability to sequences. We prove that on test samples where the base model achieves a non-trivial likelihood $\alpha$, a variant of policy gradient (PG) can achieve likelihood $1 - \varepsilon$ with an essentially minimax optimal number of reward queries $\tilde{\mathcal{O}}((\alpha^{-1} + \varepsilon^{-1})/\gamma^2)$. However, a barrier arises for going beyond the support of the base model. We prove that the overall expected error after post-training with outcome rewards is governed by a property of the base model called the *Likelihood Quantile* (LQ), and that variants of PG, while minimax optimal, may require a number of reward queries exponential in $N$ to go beyond this support, regardless of the pre-training algorithm. To overcome this barrier, we study post-training with a process reward model, and demonstrate how PG variants in this setting avoid the curse of dimensionality in $N$ via dependence on a token-level LQ. Along the way, we prove that under the margin condition, SGD with adaptive learning rate (LR) achieves a near optimal test error for statistical learning, and PG with adaptive LR achieves a near optimal number of mistakes for online learning while being computationally efficient whenever possible, both of which may be of independent interest.