ICLR 2026 论文列表

Aligning Large Language Models (LLMs) with human preferences through finetuning is resource-intensive, motivating lightweight alternatives at test time. We address test-time alignment through the lens of sequential decision making, a perspective that reveals two fundamental challenges. When actions are defined at the token level, as in guided decoding, alignment suffers from the curse of horizon. Conversely, when actions are at the response level, as in traditional iterative refinement, the curse of dimensionality emerges. To resolve this trade-off, we draw inspiration from Model Predictive Control (MPC) in control theory to propose Textual Model Predictive Control (TMPC), a novel predictive planning framework adapted for aligning LLMs at inference time. A key limitation of standard MPC is its reliance on predefined, hard segment boundaries, which are often absent in text generation. TMPC overcomes this by introducing two principles inspired by hierarchical reinforcement learning: (1) Hindsight Subgoal Identification, where TMPC analyzes generation subgoals to retrospectively identify high-reward intermediate outputs as subgoals. This allows the framework to discover meaningful, task-specific planning steps (e.g., a sentence in machine translation or a bug fix in code generation.). (2) Subgoal-Conditioned Re-Generation, where these identified subgoals are used to guide subsequent planning iterations. By conditioning on these proven, high-quality subgoals, TMPC ensures stable improvement by building upon previously validated successes. TMPC is evaluated on three tasks with distinct segmentation properties: discourse-level translation, long-form response generation, and program synthesis. The results demonstrate that TMPC consistently improves performance, highlighting the generality.

Reinforcement Learning from Verifiable Rewards (RLVR) has significantly improved the performance of large language models (LLMs) on tasks with gold ground truth, such as code generation and mathematical reasoning. However, its application to open-ended question answering (QA) remains challenging, primarily due to the absence of reliable evaluation and verifiable reward signals. This difficulty is further compounded by the limitations of existing evaluation paradigms. Previous approaches typically rely on human feedback or LLM-as-judge strategies, which are costly, prone to reward hacking, and often fail to provide sufficiently discriminative or interpretable evaluation signals. To address these limitations, we introduce a schema for generating case-wise rubrics that are question-specific, content-based and stylistically sensitive, thereby evaluating both factual soundness and writing quality. Building on this schema, we propose QuRL (Open-Ended QA with Rubric-guided Reinforcement Learning), a framework that automatically mines rubrics for each question from easily accessible online sources and leverages them as reward signals. With these rubrics, QuRL employs the GRPO (Group Relative Policy Optimization) algorithm to guide the model in exploring the correct generation path. Extensive experiments show that our framework achieves significant improvements of total +17.0 points on evaluation benchmark, demonstrating the effectiveness of rubric-guided reinforcement learning for open-ended QA.

Reasoning-based image quality assessment (IQA) models trained through reinforcement learning (RL) exhibit exceptional generalization, yet the underlying mechanisms and critical factors driving this capability remain underexplored in current research. Moreover, despite their superior performance, these models incur inference energy usage and latency orders of magnitude higher than their earlier counterparts, restricting their deployment in specific scenarios. Through extensive experiments, this paper verifies and elaborates that through RL training, MLLMs leverage their reasoning capability to convert redundant visual representations into compact, cross-domain aligned text representations. This conversion is precisely the source of the generalization exhibited by these reasoning-based IQA models. Building on this fundamental insight, we propose a novel algorithm, RALI, which employs contrastive learning to directly align images with these generalizable text representations learned by RL. This approach eliminates the reliance on reasoning processes and even obviates the need to load an LLM. For the quality scoring task, this framework achieves generalization performance comparable to reasoning-based models while requiring less than 5% of their model parameters and inference time.

Large reasoning models (LRMs) exhibit unprecedented capabilities in solving complex problems through Chain-of-Thought (CoT) reasoning. However, recent studies reveal that their final answers often contradict their own reasoning traces. We hypothesize that this inconsistency stems from two competing mechanisms for generating answers: CoT reasoning and memory retrieval. To test this hypothesis, we conduct controlled experiments that challenge LRMs with misleading cues during reasoning and/or corrupted answers during retrieval. Our results across models and datasets confirm that both mechanisms operate simultaneously, with their relative dominance influenced by multiple factors: problem domains, model scales, and fine-tuning approaches (e.g., reinforcement learning vs. distillation). The findings reveal a critical limitation in current reasoning fine-tuning paradigms: models can exploit the retrieval mechanism as a shortcut, effectively "hacking" the reward signal and undermining genuine reasoning development. To address this challenge, we introduce FARL, a novel fine-tuning framework that integrates memory unlearning with reinforcement learning. By carefully suppressing retrieval shortcuts during the fine-tuning process, FARL promotes reasoning-dominant behavior and enhances generalizable reasoning capabilities. The code is available at https://anonymous.4open.science/r/FARL-EF56.

Reinforcement learning with verifiable rewards (RLVR) has emerged as a promising paradigm for post-training large language models (LLMs) on complex reasoning tasks. Yet, the conditions under which RLVR yields robust generalization remain poorly understood. This paper provides an empirical study of RLVR generalization in the setting of probabilistic inference over causal graphical models. This setting offers two natural axes along which to examine generalization: (i) the level of the probabilistic query—associational, interventional, or counterfactual—and (ii) the structural complexity of the query, measured by the size of its relevant subgraph. We construct datasets of causal graphs and queries spanning these difficulty axes and fine-tune Qwen-2.5-Instruct models using RLVR or supervised fine-tuning (SFT). We vary both the model scale (3B–32B) and the query level included in training. We find that RLVR yields stronger within-level and across-level generalization than SFT, but only for specific combinations of model size and training query level. Further analysis shows that RLVR’s effectiveness depends on the model’s initial reasoning competence. With sufficient initial competence, RLVR improves an LLM’s marginalization strategy and reduces errors in intermediate probability calculations, producing substantial accuracy gains, particularly on more complex queries. These findings show that RLVR can improve specific causal reasoning subskills, with its benefits emerging only when the model has sufficient initial competence.

Diffusion models are widely used for generative tasks across domains. While pre-trained diffusion models effectively capture the training data distribution, it is often desirable to shape these distributions using reward functions to align with downstream applications. Policy gradient methods, such as Proximal Policy Optimization (PPO), are widely used in the context of autoregressive generation. However, the marginal likelihoods required for such methods are intractable for diffusion models, leading to alternative proposals and relaxations. In this context, we unify variants of Rejection sAmpling based Fine-Tuning (RAFT) as GRAFT, and show that this induces same solution as PPO objective with reshaped rewards. We then introduce P-GRAFT to shape distributions at intermediate noise levels and demonstrate empirically that this can lead to more effective fine-tuning. We mathematically explain this via a bias-variance tradeoff. Motivated by this, we propose inverse noise correction to improve flow models without leveraging explicit rewards. We empirically evaluate our methods on text-to-image(T2I) generation, layout generation, molecule generation and unconditional image generation. Notably, our framework, applied to Stable Diffusion 2, improves over policy gradient methods on popular T2I benchmarks in terms of VQAScore and shows an 8.81% relative improvement over the base model. For unconditional image generation, inverse noise correction improves FID of generated images at lower FLOPs/image.

The role of reasoning in Audio Large Language Models remains widely underexplored, as introducing a reasoning process often degrades rather than improves performance during inference, a phenomenon we term test-time inverse scaling, where longer reasoning chains yield progressively worse results. We demonstrate that this stems not from fundamental limitations of reasoning itself, but from inadequate training: models without proper guidance for the reasoning process produce hallucinatory, inconsistent reasoning that accumulates errors over longer chains. To address these challenges, we introduce CESAR (Consistent, Effective, and Scalable Audio Reasoners), shifting from outcome verification to rewarding the reasoning process. Our online reinforcement learning framework employs Group Relative Policy Optimization with a multi-faceted reward suite that incentivizes not only correctness and format but also consistency, structured analytical patterns, causal reasoning, domain-knowledge integration, and calibrated reasoning depth. CESAR resolves test-time inverse scaling, transforming reasoning from detriments into gains while revealing model-specific "reasoning sweet spots", where performance peaks during test-time scaling. We achieve state-of-the-art results on MMAU Test-mini, substantially outperforming Gemini 2.5 Pro and GPT-4o Audio, and near-human-level performance on MMSU reasoning tasks. Through AI-as-judge evaluations and qualitative comparisons, we provide both quantitative and qualitative validation of our improved reasoning quality. Importantly, enhanced reasoning creates synergistic effects, simultaneously improving multimodal reasoning and perception capabilities. Overall, CESAR establishes a principled method for developing robust and scalable reasoning in Audio LLMs.

Segmentation in pathology image requires handling out-of-domain tissue morphologies and new pathologies beyond training distributions, where traditional closed-set segmentation approaches fail to generalize. Reasoning segmentation enables zero-shot generalization via prompting with text queries. However, existing reasoning segmentation models face three barriers when applied to pathology: (1) the vision encoder lack pathology-specific knowledge and robustness to staining variations, (2) the large language model (LLM) backbone for reasoning fails to identify whether it has gathered sufficient semantic context to trigger the segmentation output, and (3) no reasoning segmentation benchmarks and datasets exist for pathology analysis. Consequently, we introduce PathChat-SegR1, a reasoning segmentation model built upon pathology-specific vision encoders trained with a novel stain-invariant self-distillation for robust pathology image representations. Moreover, we propose Segmentation-Optimized GRPO (SO-GRPO), a reinforcement learning method specifically for reasoning segmentation that learns to determine optimal segmentation timing based on accumulated reasoning context. Finally, we construct a pathology-specific reasoning segmentation benchmark of 118,667 triplets of pathology image, ground-truth mask, query, and reasoning chain including both public and private pathology images. Zero-shot evaluation on pathology images with out-of-domain morphologies/pathologies shows 61\% improvement over state-of-the-art segmentation models.

Positional encodings, a mechanism for incorporating sequential information into the Transformer model, are central to contemporary research on neural architectures. Previous work has largely focused on understanding their function through the principle of distance attenuation, where proximity dictates influence. However, the interaction between positional and semantic information remains insufficiently explored, and the complexity of mainstream corpora hinders systematic, comparative studies of these methods. This paper addresses these challenges through a deconstruction of the attention-logit computation and a structured analysis of all mainstream positional encodings. A key focus is placed on Rotary Positional Embedding (RoPE), whose product-based structure uniquely facilitates a direct interaction between position and content. To probe this characteristic, we designed a novel synthetic task that explicitly demands a strong synthesis of positional and semantic information. As theoretically predicted, RoPE demonstrates a significant performance advantage over other encodings on this specialized task. Concurrently, this targeted evaluation uncovers an implicit training issue: a hidden bias manifesting as a distinct information aggregation phenomenon in the model's shallow layers, which we term the "single-head deposit pattern." Through subsequent ablation studies, we analyze this pattern and identify a method for its mitigation. These findings highlight the need for a deeper investigation into the training dynamics of positional encodings to bridge the gap between their theoretical design and practical implementation.

Why do reinforcement learning (RL) policies fail or succeed? This is a challenging question due to the complex, high-dimensional nature of agent-environment interactions. We take a causal perspective on explaining the global behavior of RL policies by viewing the states, actions, and rewards as variables in a low-level causal model. We introduce random perturbations to policy actions during execution and observe their effects on the cumulative reward, learning a simplified high-level causal model that explains these relationships. To this end, we develop a nonlinear Causal Model Reduction framework that ensures approximate interventional consistency, i.e., the simplified high-level model responds to interventions in a way consistent with the original complex system. \todo{make more convincing}We prove that for a class of nonlinear causal models, there exists a unique solution that achieves exact interventional consistency, ensuring learned explanations reflect meaningful causal patterns. Experiments on both synthetic causal models and practical RL tasks~-~including pendulum control and robot table tennis~-~demonstrate that our approach can uncover important behavioral patterns, biases, and failure modes in trained RL policies.

We propose ERA, a new paradigm for entropy-constrained policy via output activation. It guarantees minimum sampling entropy by transforming the outputs of the last layer. Our approach demonstrates broad effectiveness across different domains: 1) for large language models~(LLMs), boosting the AIME 2025 score for Qwen2.5-Math-7B by 37.4%; 2) for continuous control reinforcement learning agents, improving performance by more than 30% over strong baselines such as SAC on the challenging HumanoidBench; 3) for image classification, enhancing ImageNet top-1 accuracy by 0.69% for ResNet-50. These gains are achieved with a computational overhead of less than 7%. Our work validates output activation as a powerful tool for entropy control, opening a new direction for designing simpler and more robust algorithms.

Large language models (LLMs) have achieved significant advancements in reasoning capabilities through reinforcement learning (RL) via environmental exploration. As the intrinsic properties of the environment determine the abilities that LLMs can learn, the environment plays a important role in the RL finetuning process. An ideal LLM reasoning environment should possess three core characteristics: scalability, generalizable reasoning, and verifiability. However, existing mathematical and coding environments are difficult to scale due to heavy reliance on expert annotation, while the skills learned in game-based environments are too specialized to generalize. To bridge this gap, we introduce the \textbf{S}tructured \textbf{I}n-context \textbf{E}nvironment (SIE) framework. SIE achieves scalability by automatically constructing reasoning environments from large-scale structured data, where the rich compositional patterns naturally support generalizable reasoning. Moreover, the explicit schemas and reasoning chains in structured data provide a foundation for rule-based verifiability. Experimental results show that SIE framework not only achieves substantial improvements in in-domain structured reasoning, but also enables the learned compositional reasoning skills to generalize effectively to out-of-domain mathematical and logical reasoning tasks. We further explored learning in information-limited partial SIEs and found that LLMs can infer the missing information through exploring the environment, leading to robust reasoning improvements and generalization performance. Our code can be available at \url{https://anonymous.4open.science/r/SIE_ICLR-EE6F}

Deep neural networks trained on nonstationary data must balance stability (i.e., retaining prior knowledge) and plasticity (i.e., adapting to new tasks). Standard reinitialization methods, which reinitialize weights toward their original values, are widely used but difficult to tune: conservative reinitializations fail to restore plasticity, while aggressive ones erase useful knowledge. We propose FIRE, a principled reinitialization method that explicitly balances the stability–plasticity tradeoff. FIRE quantifies stability through Squared Frobenius Error (SFE), measuring proximity to past weights, and plasticity through Deviation from Isometry (DfI), reflecting weight isotropy. The reinitialization point is obtained by solving a constrained optimization problem, minimizing SFE subject to DfI being zero, which is efficiently approximated by Newton–Schulz iteration. FIRE is evaluated on continual visual learning (CIFAR-10 with ResNet-18), language modeling (OpenWebText with GPT-0.1B), and reinforcement learning (HumanoidBench with SAC and Atari games with DQN). Across all domains, FIRE consistently outperforms both naive training without intervention and standard reinitialization methods, demonstrating effective balancing of the stability–plasticity tradeoff.

Speech-to-Speech (S2S) models have shown promising dialogue capabilities, but their ability to handle paralinguistic cues—such as emotion, tone, and speaker attributes—and to respond appropriately in both content and style remains underexplored. Progress is further hindered by the scarcity of high-quality and expressive demonstrations. To address this, we introduce a novel reinforcement learning (RL) framework for paralinguistic-aware S2S, ParaS2S, which evaluates and optimizes both content and speaking style directly at the waveform level. We first construct ParaS2SBench, a benchmark comprehensively evaluates S2S models’ output for content and style appropriateness from diverse and challenging input queries. It scores the fitness of input-output pairs and aligns well with human judgments, serving as an automatic judge for model outputs. With this scalable scoring feedback, we enable the model to explore and learn from diverse unlabeled speech via Group Relative Policy Optimization (GRPO). Experiments show that existing S2S models fail to respond appropriately to paralinguistic attributes, performing no better than pipeline-based baselines. Our RL-based strong baseline achieves a 11% relative improvement in response content and style’s appropriateness on ParaS2SBench over supervised fine-tuning (SFT), surpassing all prior models while requiring substantially fewer warm-up annotations than pure SFT.

Replication of experimental results has been a challenge faced by many scientific disciplines, including the field of machine learning. Recent work on the theory of machine learning has formalized replicability as the demand that an algorithm produce identical outcomes when executed twice on different samples from the same distribution. Provably replicable algorithms are especially interesting for reinforcement learning (RL), where algorithms are known to be unstable in practice. While replicable algorithms exist for tabular RL settings, extending these guarantees to more practical function approximation settings has remained an open problem. In this work, we make progress by developing replicable methods for linear function approximation in RL. We first introduce two efficient algorithms for replicable random design regression and uncentered covariance estimation, each of independent interest. We then leverage these tools to provide the first provably efficient replicable RL algorithms for linear Markov decision processes in both the generative model and episodic settings. Finally, we evaluate our algorithms experimentally and show how they can inspire more consistent neural policies.

Reinforcement finetuning (RFT) has garnered significant attention in recent years, particularly for enhancing large reasoning models such as OpenAI o1 and Deepseek R1. The appeal of RFT largely stems from its ability to refine model knowledge, better align outputs with user intent, and address challenges associated with limited finetuning data. Despite these advantages, the application of RFT in large Transformer-based generative agents remains relatively underexplored. Although these agents are designed to address multiple tasks through large-scale autoregressive pretraining and share many properties with large reasoning models, current adaptation strategies predominantly rely on supervised finetuning (SFT). In this work, we conduct a systematic investigation of several RFT techniques across a variety of finetuning parameter configurations and meta-reinforcement learning (meta-RL) environments, employing few-shot offline datasets. We provide a comprehensive analysis of RFT algorithm performance under diverse experimental conditions and, based on our empirical findings, introduce a lightweight enhancement to existing RFT methods. This enhancement consistently improves outcomes by combining the strengths of both SFT and RFT. Our findings provide valuable insights for advancing the effectiveness of RFT approaches and broadening their applicability to meta-RL tasks with large Transformer-based generative agents, motivating further research in broader domains.

Recent advances in reasoning domains with neural networks have primarily been enabled by a training recipe that optimizes Large Language Models, previously trained to predict the next-token in a sequence, with reinforcement learning algorithms. We introduce a framework to study the success of this paradigm, and we theoretically expose the optimization mechanisms by which reinforcement learning improves over next-token prediction in this setting. We study learning from mixture distributions of short and long “chain-of-thought” sequences encoding a single task. In particular, when the task consists of predicting the parity of $d$ bits and long sequences are rare, we show how reinforcement learning after next-token prediction enables autoregressive transformers to generalize, whereas mere next-token prediction requires extreme statistical or computational resources to do so. We further explain how reinforcement learning leverages increased test-time computation, manifested in longer responses, to facilitate this learning process. In a simplified setting, we theoretically prove that autoregressive linear models following this training recipe can efficiently learn to predict the parity of $d$ bits as long as the proportion of long demonstrations in the data mix is not exponentially small in the input dimension $d$. Finally, we demonstrate these same phenomena in other settings, including the post-training of Llama-series models on mixture variations of common mathematical reasoning benchmarks.

Reinforcement Learning with Verifiable Rewards (RLVR) for large language models (LLMs) has achieved remarkable progress in enhancing LLMs’ reasoning capabilities on tasks with clear correctness criteria, such as mathematical reasoning tasks. Several training metrics, such as entropy or response length, have been observed to correlate with different reasoning behaviors in reinforcement learning. Prior approaches incorporate such priors through reward or advantage shaping, which often relies on hand-crafted penalties and preferences (e.g., higher-is-better or lower-is-better). However, without careful hyper-parameter tuning, these directional priors can be overly biased and may lead to failure. To this end, we introduce ***C****onditional adv****AN****tage estimati****ON*** (***CANON***), amplifying the impact of the target metric without presuming its direction. Specifically, *CANON* regroups the sampled responses into two groups based on the higher or lower value of a target metric, measures which metric trend contributes to better performance through inter-group comparison, and identifies the better response within the same group. In summary, *CANON* based on entropy consistently outperforms prior methods across three LLMs on both math reasoning and high-complexity logic tasks. When applied to response length, *CANON* further improves token efficiency, yielding a more favorable Pareto frontier in the performance–cost trade-off.

Trajectory Optimization (TO) models have achieved remarkable success in offline reinforcement learning (offline RL). However, their vulnerability to backdoor attacks remains largely unexplored. We find that existing backdoor attacks in RL, which typically rely on reward manipulation throughout training, are largely ineffective against TO models due to their inherent sequence modeling nature and large network size. Moreover, the complexities introduced by high-dimensional continuous action further compound the challenge of injecting effective backdoors. To address these gaps, we propose TrojanTO, the first action-level backdoor attack against TO models. TrojanTO is a post-training attack and employs alternating training to forge a strong connection between triggers and target actions, ensuring high attack effectiveness. To maintain attack stealthiness, it utilizes trajectory filtering to preserve the benign performance and batch poisoning for trigger consistency. Extensive evaluations demonstrate that TrojanTO effectively implants backdoors across diverse tasks and attack objectives with a low attack budget (0.3\% of trajectories). Furthermore, TrojanTO exhibits broad applicability to DT, GDT, and DC, underscoring its scalability across diverse TO model architectures.