ICLR 2026 论文列表

In this work, we present a simple yet theoretically motivated improvement to Supervised Fine-Tuning (SFT) for the Large Language Model (LLM), addressing its limited generalization compared to reinforcement learning (RL). Through mathematical analysis, we reveal that standard SFT gradients implicitly encode a problematic reward structure that may severely restrict the generalization capabilities of model compared to RL. To rectify this, we propose Dynamic Fine-Tuning (DFT), stabilizing gradient updates for each token by dynamically rescaling the objective function with the probability of this token. With just a single-line change, the method outperforms standard SFT on multiple difficult benchmarks and base models, from math reasoning to code generation and multi-modal tasks, demonstrating improved generalization. Additionally, DFT achieves competitive results in offline RL settings, and further boosts the effectiveness of subsequent RL training, providing an effective yet streamlined alternative. The experiments further demonstrate that DFT not only strengthens SFT performance but also consistently improves the effectiveness of subsequent RL training. By bridging theoretical insights with practical solutions, this work advances the state of SFT. The source code will be publicly released.

For RL algorithms, appropriate entropy control is crucial to their effectiveness. To control the policy entropy, a commonly used method is entropy regularization, which is adopted in various popular RL algorithms including PPO, SAC and A3C. Although entropy regularization proves effective in robotic and games RL conventionally, studies found that it gives weak to no gains in LLM-RL training. In this work, we study the issues of entropy bonus in LLM-RL setting. Specifically, we first argue that the conventional entropy regularization suffers from the LLM's extremely large response space and the sparsity of the optimal outputs. As a remedy, we propose AEnt, an entropy control method that utilizes a new clamped entropy bonus with an automatically adjusted coefficient. The clamped entropy is evaluated with the re-normalized policy defined on certain smaller token space, which encourages exploration within a more compact response set. In addition, the algorithm automatically adjusts entropy coefficient according to the clamped entropy value, effectively controlling the entropy-induced bias while leveraging the entropy's benefits. AEnt is tested in math-reasoning tasks under different base models and datasets, and it is observed that AEnt outperforms the baselines consistently across multiple benchmarks.

Effective exploration in reinforcement learning requires not only tracking where an agent has been, but also understanding how the agent perceives and represents the world. To learn powerful representations, an agent should actively explore states that contribute to its knowledge of the environment. Temporal representations can capture the information necessary to solve a wide range of potential tasks while avoiding the computational cost associated with full state reconstruction. In this paper, we propose an exploration method that leverages temporal contrastive representations to guide exploration, prioritizing states with unpredictable future outcomes. We demonstrate that such representations can enable the learning of complex exploratory behaviors in locomotion, manipulation, and embodied-AI tasks, revealing capabilities and behaviors that traditionally require extrinsic rewards. Unlike approaches that rely on explicit distance learning or episodic memory mechanisms (e.g., quasimetric-based methods), our method builds directly on temporal similarities, yielding a simpler yet effective strategy for exploration.

Reinforcement learning with verifiable reward has recently emerged as a central paradigm for post-training large language models (LLMs); however, prevailing mean-based methods, such as Group Relative Policy Optimization (GRPO), suffer from entropy collapse and limited reasoning gains. We argue that these issues stem from overemphasizing high-probability output sequences while neglecting rare but informative reasoning paths. To address these challenges, we propose Risk-based Policy Optimization (RiskPO), which substitutes classical mean-based objectives with principled risk measures. Specifically, we introduce a Mixed Value-at-Risk objective that integrates weighted attention over multiple regions of the reward distribution, thereby amplifying gradient signals on challenging instances and preventing overconfident convergence. We further design a bundling scheme that aggregates multiple questions into bundles, thus enriching the feedback signal and yielding more stable and informative training dynamics. Theoretically, we prove that the risk-averse update alleviates entropy collapse and promotes exploration. Numerically, RiskPO achieves consistent and significant improvements in mathematical reasoning, multi-modal reasoning, and code generation benchmarks, surpassing GRPO and its variants on both Pass@1 and Pass@k metrics. Our results demonstrate that risk-based optimization provides a rigorous and effective paradigm for enhancing LLM reasoning capabilities.

We present GRL-SNAM, a geometric reinforcement learning framework for Simultaneous Navigation and Mapping in unknown environments. GRL-SNAM differs from traditional SLAM and other reinforcement learning methods by relying exclusively on local sensory observations without constructing a global map. Our approach formulates navigation and mapping as coupled dynamics on generalized Hamiltonian manifolds: sensory inputs are translated into local energy landscapes that encode reachability, obstacle barriers, and deformation constraints, while policies for sensing, planning, and reconfiguration evolve stagewise under Differential Policy Optimization (DPO). A reduced Hamiltonian serves as an adaptive score function, updating kinetic/potential terms, embedding barrier constraints, and continuously refining trajectories as new local information arrives. We evaluate GRL-SNAM on 2D deformable navigation tasks, where a hyperelastic robot learns to squeeze through narrow gaps, detour around obstacles, and generalize to unseen environments. We evaluate GRL-SNAM on procedurally generated 2D deformable–robot tasks comparing against local reactive baselines (PF, CBF, staged DWA) and global A* references (rigid, clearance-aware) under identical stagewise sensing constraints. GRL-SNAM shows superior path quality while using the minimal map coverage, preserves clearance, generalizes to unseen layouts, and demonstrates that Hamiltonian-structured RL enables high-quality navigation through minimal exploration via local energy refinement rather than global mapping.

We prove the stability and global convergence of a coupled actor-critic gradient flow for infinite-horizon and entropy-regularised Markov decision processes (MDPs) in continuous state and action space with linear function approximation under Q-function realisability. We consider a version of the actor critic gradient flow where the critic is updated using temporal difference (TD) learning while the policy is updated using a policy mirror descent method on a separate timescale. For general action spaces, the relative entropy regularizer is unbounded and thus it is not clear a priori that the actor-critc flow does not suffer from finite-time blow-up. Therefore we first demonstrate stability which in turn enables us obtain a convergence rate of the actor critic flow to the optimal regularised value function. The arguments presented show that timescale separation is crucial for stability and convergence in this setting.

The extension of diffusion models to language tasks has shown promising results, but their post-training methods remain largely unexplored. We highlight the importance of aligning a diffusion language model’s preference-inference trajectory with its post-training objective. To this end, we propose TraceRL, a trajectory-aware reinforcement learning framework for DLMs that incorporates information from inference trajectories into post-training and is applicable to both full-attention and block-attention diffusion models. We also introduce a diffusion-based value model that enhances training stability and naturally accommodates process rewards. We demonstrate TraceRL’s superiority in enhancing a model’s reasoning ability on complex math and coding tasks, as well as its applicability in scaling block diffusion models to larger block sizes. Employing TraceRL, we derive a series of state-of-the-art diffusion language models, namely TraDo. Although smaller than Qwen2.5-7B-Instruct, TraDo-4B-Instruct consistently outperforms it on complex math reasoning tasks. TraDo-8B-Instruct achieves 4.5% higher accuracy on MATH500 than Qwen2.5-7B-Instruct and 6.6% higher accuracy on LiveCodeBench-V2 than Llama3.1-8B-Instruct. Through curriculum learning, we also develop the first 8B-scale long-CoT diffusion language model.

We aim to improve the reasoning capabilities of language models via reinforcement learning with verifiable rewards (RLVR). Recent RLVR post-trained models like DeepSeek-R1 have demonstrated reasoning abilities on mathematical and coding tasks. However, prior studies suggest that using RLVR alone to improve reasoning on inherently difficult tasks is less effective due to sparse rewards. Here, we draw inspiration from curriculum learning and propose to schedule tasks from easy to hard (E2H), allowing LLMs to build reasoning skills gradually. Our method is termed E2H Reasoner. Empirically, we observe that, although easy tasks are important initially, fading them out through appropriate scheduling is essential in preventing overfitting. Theoretically, we establish convergence guarantees for E2H Reasoner within an approximate policy iteration framework. We derive finite-sample complexity bounds and show that when tasks are appropriately decomposed and conditioned, learning through curriculum stages requires fewer total samples than direct learning. Experiments across diverse datasets and models demonstrate that E2H Reasoner substantially enhances LLM reasoning.

At the boundary between the known and the unknown, an agent inevitably confronts the dilemma of whether to explore or to exploit. Epistemic uncertainty reflects such boundaries, representing systematic uncertainty due to limited knowledge. In this paper, we propose a Bayesian reinforcement learning (RL) algorithm, $\texttt{EUBRL}$, which leverages epistemic guidance to achieve principled exploration. This guidance adaptively reduces per-step regret arising from estimation errors. We establish nearly minimax-optimal regret and sample complexity guarantees for a specific class of priors in infinite-horizon discounted MDPs. Empirically, we evaluate $\texttt{EUBRL}$ on tasks characterized by sparse rewards, long horizons, and stochasticity. Results demonstrate that $\texttt{EUBRL}$ achieves superior sample efficiency, scalability, and consistency.

Large Reasoning Models (LRMs) have demonstrated impressive capabilities but suffer from cognitive inefficiencies like ''overthinking'' simple problems and ''underthinking'' complex ones. While existing methods that use supervised fine-tuning (SFT) or reinforcement learning (RL) with token-length rewards can improve efficiency, they often do so at the cost of accuracy. This paper introduces DeepCompress, a novel framework that simultaneously enhances both the accuracy and efficiency of LRMs. We challenge the prevailing approach of consistently favoring shorter reasoning paths, showing that longer responses can contain a broader range of correct solutions for difficult problems. DeepCompress employs an adaptive length reward mechanism that dynamically classifies problems as "Simple" or "Hard" in real-time based on the model's evolving capability. It encourages shorter, more efficient reasoning for "Simple" problems while promoting longer, more exploratory thought chains for "Hard" problems. This dual-reward strategy enables the model to autonomously adjust its Chain-of-Thought (CoT) length, compressing reasoning for well-mastered problems and extending it for those it finds challenging. Experimental results on challenging mathematical benchmarks show that DeepCompress consistently outperforms baseline methods, achieving superior accuracy while significantly improving token efficiency.

Generalizing policies across environments with altered dynamics remains a key challenge in reinforcement learning, particularly in offline settings where direct interaction or fine-tuning is impractical. We introduce Masked Skill Token Training (MSTT), a fully offline hierarchical RL framework that enables policy transfer using observation-only demonstrations. MSTT constructs a discrete skill space via unsupervised trajectory tokenization and trains a skill-conditioned value function using masked Bellman updates, which simulate dynamics shifts by selectively disabling skills. A diffusion-based trajectory generator, paired with feasibility-based filtering, enables the agent to execute valid, temporally extended actions without requiring action labels or access to the target environment. Our results in both discrete and continuous domains demonstrate the potential of mask-guided planning for robust generalization under dynamics shifts. To our knowledge, MSTT is the first work to explore masking as a mechanism for simulating and generalizing across off-dynamics environments. It marks a promising step toward scalable, structure-aware transfer and opens avenues to explore multi-goal conditioning, and extensions to more complex, real-world scenarios.

A central challenge in image-based Model-Based Reinforcement Learning (MBRL) is to learn representations that distill essential information from irrelevant visual details. While promising, reconstruction-based methods often waste capacity on large task-irrelevant regions. Decoder-free methods instead learn robust representations by leveraging Data Augmentation (DA), but reliance on such external regularizers limits versatility. We propose R2-Dreamer, a decoder-free MBRL framework with a self-supervised objective that serves as an internal regularizer, preventing representation collapse without resorting to DA. The core of our method is a \emph{redundancy-reduction} objective inspired by Barlow Twins, which can be easily integrated into existing frameworks. On DeepMind Control Suite and Meta-World, R2-Dreamer is competitive with strong baselines such as DreamerV3 and TD-MPC2 while training 1.59$\times$ faster than DreamerV3, and yields substantial gains on DMC-Subtle with tiny task-relevant objects. These results suggest that an effective internal regularizer can enable versatile, high-performance decoder-free MBRL. Code is available at https://github.com/NM512/r2dreamer.

Image captioning is a fundamental task that bridges the visual and linguistic domains, playing a critical role in pre-training Large Vision-Language Models (LVLMs). Current state-of-the-art captioning models are typically trained with Supervised Fine-Tuning (SFT), a paradigm that relies on expensive, non-scalable data annotated by humans or proprietary models. This approach often leads to models that memorize specific ground-truth answers, limiting their generality and ability to generate diverse, creative descriptions. To overcome the limitation of SFT, we propose applying the Reinforcement Learning with Verifiable Rewards (RLVR) paradigm to the open-ended task of image captioning. A primary challenge, however, is designing an objective reward function for the inherently subjective nature of what constitutes a "good" caption. We introduce Captioning Reinforce- ment Learning (CapRL), a novel training framework that redefines caption quality through its utility: a high-quality caption should enable a non-visual language model to accurately answer questions about the corresponding image. CapRL employs a decoupled two-stage pipeline where an LVLM generates a caption, and the objective reward is derived from the accuracy of a separate, vision-free LLM answering Multiple-Choice Questions based solely on that caption. As the first study to apply RLVR to the subjective image captioning task, we demonstrate that CapRL significantly enhances multiple settings. Pretraining on the CapRL- 5M caption dataset annotated by CapRL-3B results in substantial gains across 12 benchmarks. Moreover, within the Prism Framework for caption quality evaluation, CapRL achieves performance comparable to Qwen2.5-VL-72B, while exceeding the baseline by an average margin of 8.4%. Results validate that our CapRL effec- tively trains models to produce a more general and accurate image descriptions, moving beyond the limitations of traditional SFT-based image captioning models.

We study behavior-regularized reinforcement learning (RL), which encompasses offline RL and RL from human feedback (RLHF). In both settings, regularization toward a reference distribution (offline data in offline RL or the supervised-finetuned policy in RLHF) is essential to prevent value over-optimization caused by erroneous out-of-distribution extrapolation. Existing methods typically add distance or divergence penalties on the learning objective, which introduces optimization challenges and over-conservatism. In this paper, we propose Value Gradient Flow (VGF), a new paradigm for behavior-regularized RL. VGF formulates an optimal transport problem from the reference distribution to the optimal policy distribution induced by the value function. This problem is solved via discrete gradient flow, where value gradients guide particles sampled from the reference distribution. Our theoretical analysis shows that an implicit behavior regularization is imposed by controlling the transport budget. This formulation avoids unnecessary restrictions on the optimization problem, enabling better reward maximization. Moreover, VGF operates without explicit policy parameterization while remaining expressive and flexible, allowing adaptively test-time scaling by adjusting the transport budget. Extensive experiments demonstrate that VGF significantly outperforms prior methods, achieving state-of-the-art results on offline RL benchmarks (D4RL, OGBench) and challenging RLHF tasks.

Safe exploration is a key requirement for reinforcement learning agents to learn and adapt online, beyond controlled (e.g. simulated) environments. In this work, we tackle this challenge by utilizing suboptimal yet conservative policies (e.g., obtained from offline data or simulators) as priors. Our approach, SOOPER, uses probabilistic dynamics models to optimistically explore, yet pessimistically fall back to the conservative policy prior if needed. We prove that SOOPER guarantees safety throughout learning, and establish convergence to an optimal policy by bounding its cumulative regret. Extensive experiments on key safe RL benchmarks and real-world hardware demonstrate that SOOPER is scalable, outperforms the state-of-the-art and validate our theoretical guarantees in practice.

Generalizing deep reinforcement learning agents to unseen environments remains a significant challenge. One promising solution is Unsupervised Environment Design (UED), a co‑evolutionary framework in which a teacher adaptively generates tasks with high learning potential, while a student learns a robust policy from this evolving curriculum. Existing UED methods typically measure learning potential via regret, the gap between optimal and current performance, approximated solely by value‑function loss. Building on these approaches, we introduce the transition-prediction error as an additional term in our regret approximation. To capture how training on one task affects performance on others, we further propose a lightweight metric called Co-Learnability. By combining these two measures, we present Transition‑aware Regret Approximation with Co‑learnability for Environment Design (TRACED). Empirical evaluations show that TRACED produces curricula that improve zero-shot generalization over strong baselines across multiple benchmarks. Ablation studies confirm that the transition-prediction error drives rapid complexity ramp‑up and that Co‑Learnability delivers additional gains when paired with the transition-prediction error. These results demonstrate how refined regret approximation and explicit modeling of task relationships can be leveraged for sample-efficient curriculum design in UED.

Many sequential decision-making tasks involve optimizing multiple conflicting objectives, requiring policies that adapt to different user preferences. In multi-objective reinforcement learning (MORL), one widely studied approach addresses this by training a single policy network conditioned on preference-weighted rewards. In this paper, we explore a novel algorithmic perspective: leveraging reward-free reinforcement learning (RFRL) for MORL. While RFRL has historically been studied independently of MORL, it learns optimal policies for any possible reward function, making it a natural fit for MORL's challenge of handling unknown user preferences. We propose using the RFRL's training objective as an auxiliary task to enhance MORL, enabling more effective knowledge sharing beyond the multi-objective reward function given at training time. To this end, we adapt a state-of-the-art RFRL algorithm to the MORL setting and introduce a preference-guided exploration strategy that focuses learning on relevant parts of the environment. Through extensive experiments and ablation studies, we demonstrate that our approach significantly outperforms the state-of-the-art MORL methods across diverse MO-Gymnasium tasks, achieving superior performance and data efficiency. This work provides the first systematic adaptation of RFRL to MORL, demonstrating its potential as a scalable and empirically effective solution to multi-objective policy learning.

While supervised fine-tuning (SFT) and preference learning (PL) are widely used to enhance the instruction-following ability of large language models (LLMs), they often struggle to generalize to novel or complex instructions and may compromise the models' general capabilities. In-context learning (ICL) emerges as a promising alternative due to its strong generalization without modifying the model's parameters, but its effectiveness is constrained by the reliance on high-quality, manually curated demonstration pools. To overcome this limitation, we propose ContextIF, a reinforcement learning (RL) framework for automatic context generation. Guided by comprehensive context reward, ContextIF is optimized by Group Relative Policy Optimization (GRPO). It aims to generate precise constraint summaries and optimal context demonstrations tailored to given instructions, thereby improving the instruction-following performance of target LLMs. We evaluate ContextIF on multiple representative instruction-following benchmarks using popular open-source LLMs. Experimental results demonstrate that ContextIF achieves substantial performance gains over existing SFT and ICL methods, while also generalizing effectively to unseen constraint conditions. Moreover, ContextIF preserves the parameters and general capabilities of the target models, offering strong adaptability and scalability. The code is provided in the Supplementary Materials.

Recent years have witnessed increasing interests in tackling heteroscedastic noise in bandits and reinforcement learning \citep{zhou2021nearly, zhao2023variance, jia2024does, pacchiano2025second}. In these works, the cumulative variance of the noise $\Lambda = \sum_{t=1}^T \sigma_t^2$, where $\sigma_t^2$ is the variance of the noise at round $t$, has been used to characterize the statistical complexity of the problem, yielding simple regret bounds of order $\tilde{\mathcal O}(d \sqrt{\Lambda / T^2})$ for linear bandits with heteroscedastic noise \citep{zhou2021nearly, zhao2023variance}. However, with a closer look, $\Lambda$ remains the same order even if the noise is close to zero at half of the rounds, which indicates that the $\Lambda$-dependence is not optimal. In this paper, we revisit the linear bandit problem with heteroscedastic noise. We consider the setting where the action set is fixed throughout the learning process. We propose a novel variance-adaptive algorithm VAEE (Variance-Aware Exploration with Elimination) for large action set, which actively explores actions that maximizes the information gain among a candidate set of actions that are not eliminated. With the active-exploration strategy, we show that VAEE achieves a *simple regret* with a nearly *harmonic-mean* dependent rate, i.e. $\tilde{\mathcal O}\Big(d\Big[\sum_{t = 1}^T \frac{1}{\sigma_t^2} - \sum_{i = 1}^{\tilde{O}(d)} \frac{1}{[\sigma^{(i)}]^2} \Big]^{-\frac{1}{2}}\Big)$ where $d$ is the dimension of the feature space and $\sigma^{(i)}$ is the $i$-th smallest variance among $\\{\sigma_t\\}_{t=1}^T$. For finitely many actions, we propose a variance-aware variant of G-optimal design based exploration, which achieves a $\tilde {\mathcal O}$ $\bigg(\sqrt{d \log |\mathcal A| }\Big[ \sum\_{t = 1}\^T \frac{1}{\sigma\_t\^2}- \sum\_{i = 1}^{\tilde{O}(d)} \frac{1}{[\sigma^{(i)}]^2} \Big]^{-\frac{1}{2}}\bigg)$ simple regret bound. We also establish a nearly matching lower bound for the fixed action set setting indicating that \emph{harmonic-mean} dependent rate is unavoidable. To the best of our knowledge, this is the first work that breaks the $\sqrt{\Lambda}$ barrier for linear bandits with heteroscedastic noise.