ICML 2026 Papers

Image quality assessment (IQA) is inherently multi-dimensional, yet existing reward models are typically limited to a single task and become unstable when extended to multi-task settings. In particular, heterogeneous reward scales and variances across tasks can lead to conflicting optimization signals during reinforcement learning. We propose VisualScore, a unified visual evaluation framework that formulates multi-task IQA as structured, task-aware reasoning followed by continuous reward optimization. VisualScore produces interpretable rationales together with scalar quality scores under explicit evaluation principles. We construct a reasoning-enhanced reward modeling dataset via rejection sampling and initialize the model through supervised fine-tuning. VisualScore is then optimized with Group Relative Policy Optimization (GRPO) using a Gaussian-based continuous reward. To address multi-task reward conflicts and stabilize training, we introduce standard deviation filtering and entropy gating to normalize task-wise reward signals and suppress noisy updates. Experiments on technical quality, aesthetic quality, and text–image alignment show that VisualScore improves robustness, generalization, and interpretability, and can effectively guide text-to-image generation at test time without retraining.

Large language models (LLMs) are trained and tested extensively on symbolic representations such as code and graphs, yet real-world user tasks are often specified in natural language. To what extent can LLMs generalize across these representations? Here, we approach this question by studying isomorphic tasks involving procedures represented in code, graphs, and natural language (e.g., scheduling steps in planning). We find that training LLMs with popular post-training methods on graphs or code data alone does not reliably generalize to corresponding natural language tasks, while training solely on natural language can lead to inefficient performance gains. To address this gap, we propose a two-stage reinforcement learning curriculum that first trains on symbolic, then natural language data. The curriculum substantially improves model performance across model families and tasks. Remarkably, a 1.5B Qwen model trained by our method can closely match zero-shot GPT-4o in naturalistic planning. Finally, our analysis suggests that successful cross-representation generalization can be interpreted as a form of generative analogy, which our curriculum effectively encourages. The dataset and code used in this paper can be found \href{https://github.com/fangru-lin/procedure_generalization_llm}{here}.

Direct Preference Optimization (DPO) has emerged as a popular alternative to Reinforcement Learning from Human Feedback (RLHF), offering theoretical equivalence with a simpler implementation. We prove this equivalence is _conditional_ rather than universal, depending on an implicit assumption frequently violated in practice: the RLHF-optimal policy must prefer human-preferred responses. When this assumption fails, DPO optimizes _relative advantage_ over the reference policy rather than _absolute alignment_ with human preferences, leading to pathological convergence where policies decrease DPO loss while preferring dispreferred responses. We characterize when this assumption is violated, show the existence of an undesirable solution space, and prove that DPO and RLHF optimize fundamentally different objectives in such cases. To address this, we introduce Constrained Preference Optimization (CPO), augmenting RLHF with constraints for provable alignment. We further provide a geometric interpretation through soft margin ranking, revealing that DPO implements margin ranking with potentially negative targets. Our theoretical analysis establishes when DPOs’ guarantees hold and provides solutions preserving simplicity with provable alignment. Comprehensive experiments on standard benchmarks demonstrate that CPO achieves state-of-the-art performance. Code is available at: _https://github.com/visitworld123/CPO_.

Ensuring that large language models (LLMs) comply with safety requirements is a central challenge in AI deployment. Existing alignment approaches operate primarily during training, such as through fine-tuning or reinforcement learning from human feedback, but these methods are costly and inflexible, requiring retraining whenever new requirements arise. Recent efforts toward inference-time alignment mitigate some of these limitations but still assume access to model internals, which is impractical, and not suitable for third party stakeholders who do not have access to the models. In this work, we propose a model-independent, black-box framework for safety alignment that does not require retraining or access to the underlying LLM architecture. As a proof of concept, we address the problem of trading off between generating safe but uninformative answers versus helpful yet potentially risky ones. We formulate this dilemma as a two-player zero-sum game whose minimax equilibrium captures the optimal balance between safety and helpfulness. LLM agents operationalize this framework by leveraging a linear programming solver at inference time to compute equilibrium strategies. Our results demonstrate the feasibility of black-box safety alignment, offering a scalable and accessible pathway for stakeholders, including smaller organizations and entities in resource-constrained settings, to enforce safety across rapidly evolving LLM ecosystems.

Recent advances in group-based reinforcement learning (RL) greatly improve LLMs' ability in text reasoning. Yet, these methods lack sufficient modeling of multimodal information, leading to significant reasoning hallucination. In this work, we propose POLIA, a novel group-based RL method with visual-object-level intrinsic advantage for multimodal reasoning. POLIA introduces two advantage computation stages over candidate answers and visual objects, respectively. The answer-level extrinsic advantages are computed based on the extrinsic rewards of a group of candidate answers. Moreover, we compute an intrinsic advantage for each visual object based on its confidence score and reference relations with final answers. Intuitively, the intrinsic advantage of an object reflects its potential contribution to the correct answer. This two-stage advantage computation ensures an accurate credit assignment mechanism over multimodal reasoning sequences with multiple visual objects. Experimental results on diverse multimodal reasoning benchmarks show that POLIA significantly outperforms open MLLMs and strong baselines. Code is available at https://github.com/dudu115/POLIAcode.

Standard RLHF relies on transitive scalar rewards, failing to capture the cyclic nature of human preferences. While some approaches like the General Preference Model (GPM) address this, we identify a theoretical limitation: their implicit formulation entangles hierarchy with cyclicity, failing to guarantee dominant solutions. To address this, we propose the Hybrid Reward-Cyclic (HRC) model, which utilizes game-theoretic decomposition to explicitly disentangle preferences into orthogonal transitive (scalar) and cyclic (vector) components. Complementing this, we introduce Dynamic Self-Play Preference Optimization (DSPPO), which treats alignment as a time-varying game to progressively guide the policy toward the Nash equilibrium. Synthetic data experiments further validate HRC's structural superiority in mixed transitive--cyclic settings, where HRC converges faster and achieves higher accuracy than GPM. Experiments on RewardBench 2 demonstrate that HRC consistently improves over both BT and GPM baselines (e.g., +1.23\% on Gemma-2B-it). In particular, its superior performance in the Ties domain empirically validates the model's robustness in handling complex, non-strict preferences. Extensive downstream evaluations on AlpacaEval 2.0, Arena-Hard-v0.1, and MT-Bench confirm the efficacy of our framework. Notably, when using Gemma-2B-it as the base preference model, HRC+DSPPO achieves a peak length-controlled win-rate of 44.75\% on AlpacaEval 2.0 and 46.8\% on Arena-Hard-v0.1, significantly outperforming SPPO baselines trained with BT or GPM. Our code is publicly available at https://github.com/lab-klc/Hybrid-Reward-Cyclic.

Asymmetric reinforcement learning leverages privileged information available during training to improve learning under partial observability. Existing asymmetric actor-critic methods typically assume access to the full environment state to condition the critic during training, which is often unrealistic in practice. We introduce the informed asymmetric actor-critic framework that allows the critic to be conditioned on arbitrary state-dependent privileged signals, and show that any such signal yields unbiased policy gradient estimates. This substantially expands the set of admissible privileged information and raises the problem of selecting the most informative signals for learning. To this end, we propose two novel informativeness criteria: a dependence-based test that can be applied prior to training, and a test based on improvements in value prediction that can be applied post hoc. Experiments on partially observable benchmarks and synthetic environments demonstrate that carefully selected privileged signals can match or outperform full-state asymmetric baselines while relying on strictly less state information.

Because of the pervasive use of deep neural networks (DNNs), especially in high-stakes domains, the interpretability of DNNs has received increased attention. The general idea of rationale extraction (RE) is to provide an interpretable-by-design framework for DNNs via a select-predict architecture where two neural networks learn jointly to perform feature selection and prediction, respectively. Given only the remote supervision from the final task prediction, the process of learning to select subsets of features (or *rationales*) requires searching in the space of all possible feature combinations, which is computationally challenging and even harder when the base neural networks are not sufficiently capable. To improve the predictive performance of RE models that are based on less capable or smaller neural networks (i.e., the students), we propose **REKD** (**R**ationale **E**xtraction with **K**nowledge **D**istillation) where a student RE model learns from the rationales and predictions of a teacher (i.e., a *rationalist*) in addition to the student's own RE optimization. This structural adjustment to RE aligns well with how humans could learn effectively from interpretable and verifiable knowledge. Because of the neural-model agnostic nature of the method, any black-box neural network could be integrated as a backbone model. To demonstrate the viability of REKD, we conduct experiments with multiple variants of BERT and vision transformer (ViT) models. Our experiments across language and vision classification datasets (i.e., IMDB movie reviews, CIFAR 10 and CIFAR 100) show that REKD significantly improves the predictive performance of the student RE models. The code is publicly available: https://github.com/JiayiDai/REKD.

Simultaneous speech translation requires translating source speech into a target language in real-time while handling non-monotonic word dependencies. Traditional approaches rely on supervised training with word-level aligned data, which is difficult to collect at scale and thus depends on synthetic alignments using language-specific heuristics that are suboptimal. We propose Hibiki-Zero, which eliminates the need for word-level alignments entirely. This fundamentally simplifies the training pipeline and enables seamless scaling to diverse languages with varying grammatical structures, removing the bottleneck of designing language-specific alignment heuristics. We first train on sentence-level aligned data to learn speech translation at high latency, then apply a novel reinforcement learning strategy using GRPO to optimize latency while preserving translation quality. Hibiki-Zero achieves state-of-the-art performance in translation accuracy, latency, voice transfer, and naturalness across four X-to-English tasks. Moreover, we demonstrate that our model can be adapted to support a new input language with less than 1000h of speech. We provide [examples](https://huggingface.co/spaces/kyutai/hibiki-zero-samples), [model weights](https://huggingface.co/kyutai/hibiki-zero-3b-pytorch-bf16), [inference code](https://github.com/kyutai-labs/hibiki-zero) and we release a [benchmark](https://huggingface.co/datasets/kyutai/Audio-NTREX-4L) containing 45h of multilingual data for speech translation evaluation.

We systematically evaluate Parameter-Efficient Fine-Tuning (PEFT) methods under the paradigm of Reinforcement Learning with Verifiable Rewards (RLVR). RLVR incentivizes language models to enhance their reasoning capabilities through verifiable feedback; however, while methods like LoRA are commonly used, the optimal PEFT architecture for RLVR remains unidentified. In this work, we conduct the first comprehensive evaluation of over 12 PEFT methodologies across the DeepSeek-R1-Distill families on mathematical reasoning benchmarks. Our empirical results challenge the default adoption of standard LoRA with three main findings. First, we demonstrate that structural variants, such as DoRA, AdaLoRA, and MiSS, consistently outperform LoRA. Second, we uncover a spectral collapse phenomenon in SVD-informed initialization strategies (e.g., PiSSA, MiLoRA), attributing their failure to a fundamental misalignment between principal-component updates and RL optimization. Furthermore, our ablations reveal that extreme parameter reduction (e.g., VeRA, Rank-1) severely bottlenecks reasoning capacity. We further conduct ablation studies and scaling experiments to validate our findings. This work provides a definitive guide for advocating for more exploration for parameter-efficient RL methods.

Reward design remains a significant bottleneck in applying reinforcement learning (RL) to real-world problems. A popular alternative is reward learning, where reward functions are inferred from human feedback rather than manually specified. Recent work has proposed learning reward functions from human ratings rather than traditional binary preferences, enabling richer and potentially less cognitively demanding supervision. Building on this paradigm, we introduce a new rating-based RL method, Ranked Return Regression for RL (R4). At its core, R4 uses a novel ranking mean squared error loss that learns from a dataset of trajectory–rating pairs, treating the human-provided discrete ratings (e.g., "bad," "neutral," "good") as ordinal targets. Unlike prior rating-based approaches, R4 offers formal guarantees: its solution set is provably minimal and complete under mild assumptions. Empirically, using both human-provided and simulated ratings, we demonstrate that R4 consistently matches or outperforms existing rating and preference-based RL methods on robotic benchmarks from OpenAI Gym and the DeepMind Control Suite.

Token-based transformer world models have shown strong performance in visual reinforcement learning, but often suffer from temporal inconsistency in long-horizon rollouts, including object duplication, disappearance, and transmutation. A key reason is that most existing approaches treat next-frame prediction purely as a token generation problem, without considering the persistence of tokens across time. We introduce Identifiable Token Correspondence (ITC), a decoding step for token-based transformer world models that formulates next-frame prediction as a structured assignment problem with latent token correspondence variables: each next-frame token is explained either by copying a token from the previous frame or by generating a new one. ITC leaves the transformer architecture and training procedure unchanged and can be added on top of existing backbones. Our experiments show state-of-the-art performance on 4 challenging benchmarks. The proposed method achieves a return of 72.5% and a score of 35.6% on the Craftax-classic benchmark, significantly surpassing the previous best of 67.4% and 27.9%. We release our source code at https://github.com/snu-mllab/Identifiable-Token-Correspondence.

Robust reinforcement learning (RRL) aims to tackle unexpected environmental changes by optimizing policies against the worst case. However, RRL remains impractical due to the cost of the Max-Min optimization, where it suffers from the exhaustive query complexity for finding the worst-case (dubbed 'Min') within the environmental uncertainty set $\mathcal{U}$, i.e., $\mathcal{O}(|\mathcal{U}|)$. By viewing this via a lens of quantum perspective, we raise a pivotal question: *If we can query from the environment with quantum superpositions, is it possible to accelerate the Max-Min optimization of RRL?* Our answer is 'Yes'. Our method, called quantum robust inner minimization (QRIM), encodes the uncertainty set with quantum superposition and amplifies low-return cases, thus enabling RL for solving the robust (i.e., worst-case) Bellman equation. Importantly, QRIM achieves a quadratic speed-up in query complexity without altering the outer RL pipeline, i.e., $\mathcal{O}(\sqrt{|\mathcal{U}|})$. Validated through classical simulations to real quantum hardware execution, QRIM learns more robust policies with quadratically reduced queries than classical RL.

Floorplanning determines the coordinate and shape of each module in Integrated Circuit. With the advancement of technology node, in floorplanning stage especially 3D scenarios with multiple stacked layers, it has become increasingly challenging to adhere to complex hardware design rules. Current methods are only capable of handling specific and limited design rules, while violations of other rules require manual and meticulous adjustment. This leads to labor-intensive and time-consuming post-processing for expert engineers. In this paper, we propose an all-in-one deep reinforcement learning-based approach to tackle these challenges, and design novel representations for real-world IC design rules that have not been addressed by previous approaches. Specifically, the processing of various hardware design rules is unified into a single framework with three key components: 1) novel matrix representations to model the design rules, 2) constraints on the action space to filter out invalid actions that cause rule violations, and 3) quantitative analysis of constraint satisfaction as reward signals. Experiments on public benchmarks demonstrate the effectiveness and validity of our approach. Furthermore, transferability are well shown on unseen circuits. Our framework is extensible to accommodate new design rules, thus providing flexibility to address emerging challenges in future chip design.

Medical vision-language models (VLMs) have rapidly advanced as general-purpose multimodal assistants, yet their deployment in 3D Computed Tomography (CT) analysis remains constrained by a persistent mismatch between optimization objectives and clinical rigor. Current Reinforcement Learning (RL) paradigms still rely on lexical proxy signals that induce Evaluation Hallucinations, where models optimize linguistic fluency rather than factual clinical correctness, leading to diagnostically critical errors. To bridge this gap, we introduce the Clinical Abnormality Benchmarking Substrate (CABS), a structured system that decomposes radiology reports into verifiable clinical semantic units. Using CABS, we identify a Mechanistic Divergence in standard RL, where surface-similarity rewards drive policy gradients to bypass medical facts. We therefore propose Trajectory-Integral Feedback GRPO (TIF-GRPO), a novel framework integrating control-theoretic principles into policy optimization. By formulating clinical reasoning as a pseudo-temporal trajectory for anomaly discovery, TIF-GRPO regulates anatomy-aware rewards via an integral feedback loop that penalizes persistent omissions as cumulative state errors and suppresses hallucinations as excessive control effort. Experiments on 3D CT benchmarks demonstrate that our approach significantly enhances abnormality detection and clinical faithfulness, establishing a new paradigm for fine-grained regulation in medical VLMs. Our project is available at https://github.com/ZJU4HealthCare/TIF-GRPO.

Aligning Large Language Models (LLMs) with specific personas typically relies on Supervised Fine-Tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF); however, these methods are resource-intensive, requiring expensive data collection and distinct model training for each target personality. In this work, we propose a parameter-efficient framework for continuous, multi-dimensional personality control via inference-time activation steering. Our approach addresses the challenge of combining multiple interventions by iteratively retraining probes on the residual stream modified by prior traits, ensuring compatibility. Once established, these steering vectors function as modular, reusable primitives; users can instantly synthesize novel, complex personality profiles by simply adjusting steering coefficients (α) without any additional training. To support this, we introduce an automated pipeline that identifies optimal intervention layers via activation separation analysis and calibrates coefficients via hyperparameter optimization to maximize alignment while constraining perplexity. Empirical evaluations validate individual trait shifts using an LLM-as-a-judge framework and demonstrate, via the Big Five inventory, that our method effectively modulates the model's holistic personality profile without updating base model parameters.

Planning with a learned model remains a key challenge in model-based reinforcement learning (RL). In decision-time planning, state representations are critical as they must support local cost computation while preserving long-horizon structure. In this paper, we show that the Laplacian representation provides an effective latent space for planning by capturing state-space distances at multiple time scales. This representation preserves meaningful distances and naturally decomposes long-horizon problems into subgoals, also mitigating the compounding errors that arise over long prediction horizons. Building on these properties, we introduce ALPS, a hierarchical planning algorithm, and demonstrate that it outperforms commonly used baselines on a selection of offline goal-conditioned RL tasks from OGBench, a benchmark previously dominated by model-free methods.

Reinforcement learning from human feedback (RLHF) has evolved to be one of the main methods for fine-tuning large language models (LLMs). However, existing RLHF methods are non-robust, and their performance deteriorates if the downstream task differs significantly from the preference dataset used in fine-tuning. In order to mitigate this problem, we introduce a distributionally robust RLHF for fine-tuning LLMs. In particular, our goal is to ensure that a fine-tuned model retains its performance even when the distribution of prompts significantly differs from the distribution encountered during fine-tuning. We formulate distributionally robust optimization (DRO) version of two popular fine-tuning methods -- (1) reward-based RLHF and (2) reward-free DPO (direct preference optimization). We propose a minibatch gradient descent based algorithms for both of them, and theoretically prove convergence guarantees for the algorithms. Subsequently, we evaluate our algorithms on an out-of-distribution (OOD) task by first training the model on the Unified-Feedback dataset and evaluating its performance on two different datasets. The experimental results show that our robust training improves the accuracy of the learned reward models on average, and markedly on some tasks, such as reasoning. Furthermore, we show that the robust versions of policy optimization methods, similarly improve performance on OOD tasks.

Although reinforcement learning (RL) enhances the reasoning capabilities of large language models (LLMs), it is primarily learned from the model's self-generated distribution, limiting its ability to acquire reasoning skills beyond its initial knowledge. To overcome this, we propose a Difficulty-Aware Learning Strategy Allocation (DALSA) framework, which adaptively assigns appropriate learning strategies to samples based on their difficulty signals. DALSA is built on the key insight that samples beyond models' knowledge scope are better addressed through supervised fine-tuning (SFT), while those within the boundary but insufficiently mastered benefit more from RL, and well-learned samples are discarded to avoid redundant updates. To realize this principle, we extract a series of difficulty-aware training characteristics and employ a learnable strategy allocator to dynamically determine the optimal learning strategy for each sample based on its training dynamics. The allocator and the LLM are alternately optimized, enabling adaptive strategy allocation. Furthermore, two regularization techniques, anti-curriculum weighting and adversarial label smoothing, are integrated to alleviate the inherent limitations of RL and SFT, backed by comprehensive theoretical analyses. Extensive experiments on ten LLMs ranging from 1.5B to 70B across various tasks indicate that DALSA consistently outperforms baselines under both full and parameter-efficient fine-tuning settings.