ICLR 2026 Papers

Offline reinforcement learning (RL) holds the promise of training effective policies from static datasets without the need for costly online interactions. However, offline RL faces key limitations, most notably the challenge of generalizing to unseen or infrequently encountered state-action pairs. When a value function is learned from limited data in sparse-reward environments, it can become overly optimistic about parts of the space that are poorly represented, leading to unreliable value estimates and degraded policy quality. To address these challenges, we introduce a novel approach based on contrastive preference learning that bypasses direct value function estimation. Our method trains policies by contrasting successful demonstrations with failure behaviors present in the dataset, as well as synthetic behaviors generated outside the support of the dataset distribution. This contrastive formulation mitigates overestimation bias and improves robustness in offline learning. Empirical results on challenging sparse-reward offline RL benchmarks show that our method substantially outperforms existing state-of-the-art baselines in both learning efficiency and final performance.

Reinforcement learning (RL) is widely used to post-train large language models for tasks such as mathematical reasoning and coding. However, the convergence of RL training remains sensitive to batching and prompt selection strategies. We investigate the factors that affect convergence, including batch size and prompt difficulty. Through large-scale experiments across multiple models and datasets, we show that there exists an optimal batch size that balances generation time and gradient quality, and that prompts of intermediate difficulty (where the model has roughly a 50\% chance of success) are the most sample-efficient for model convergence. Motivated by these findings, we propose Prompt Curriculum Learning (PCL), a lightweight algorithm that selects intermediate-difficulty prompts using a learned value model. PCL avoids costly rollouts and efficiently guides training by focusing on the most informative samples. Empirically, PCL either achieves the highest performance or requires significantly less training time to reach comparable performance across a suite of benchmarks. Compared to using rollouts to filter, PCL is $12.1\times$ and $16.9\times$ faster on identifying intermediate-difficulty prompts when training on MATH and DeepScaleR respectively.

Ensuring that reinforcement learning (RL) controllers satisfy safety and reliability constraints in real-world settings remains challenging: state-avoidance and constrained Markov decision processes often fail to capture trajectory-level requirements or induce overly conservative behavior. Formal specification languages such as linear temporal logic (LTL) offer correct-by-construction objectives, yet their rewards are typically sparse, and heuristic shaping can undermine correctness. We introduce, to our knowledge, the first end-to-end framework that integrates LTL with differentiable simulators, enabling efficient gradient-based learning directly from formal specifications. Our method relaxes discrete automaton transitions via soft labeling of states, yielding differentiable rewards and state representations that mitigate the sparsity issue intrinsic to LTL while preserving objective soundness. We provide theoretical guarantees connecting Büchi acceptance to both discrete and differentiable LTL returns and derive a tunable bound on their discrepancy in deterministic and stochastic settings. Empirically, across complex, nonlinear, contact-rich continuous-control tasks, our approach substantially accelerates training and achieves up to twice the returns of discrete baselines. We further demonstrate compatibility with reward machines, thereby covering co-safe LTL and LTLf without modification. By rendering automaton-based rewards differentiable, our work bridges formal methods and deep RL, enabling safe, specification-driven learning in continuous domains.

Mixed-Integer Programming (MIP) solvers rely heavily on cutting planes to tighten LP relaxations, but traditional approaches generate thousands of cuts that consume gigabytes of memory while providing minimal benefit. We present an intelligent cut selection framework that achieves a 98.1\% reduction in memory usage while maintaining competitive solving with an objective gap of approximately 0.08\%. Within this RL framework, we use Proximal Policy Optimization (PPO) to learn a behavioral model that imitates the expert solver’s decisions. The adversarially imitated behavioral model drives an agent comprising these key innovations: (i) a cut-selection policy trained via curriculum learning; and (ii) adaptive inference that dynamically adjusts computational budgets. Through comprehensive evaluation across SetCover and diverse MIPLIB problems, we demonstrate consistent speedups (3.78$\times$ average on MIPLIB) and achieve a 100\% success rate on instances where traditional SCIP fails 47-53\% of the time. Our method also reduces peak memory consumption from 3.03GB to 46 MB, enabling optimization in previously inaccessible and other resource-constrained environments where traditional solvers face fundamental limitations.

Training expressive flow-based policies with off-policy reinforcement learning is notoriously unstable due to gradient pathologies in the multi-step action sampling process. We trace this instability to a fundamental connection: the flow rollout is algebraically equivalent to a residual recurrent computation, making it susceptible to the same vanishing and exploding gradients as RNNs. To address this, we reparameterize the velocity network using principles from modern sequential models, introducing two stable architectures: Flow-G, which incorporates a gated velocity, and Flow-T, which utilizes a decoded velocity. We then develop a practical SAC-based algorithm, enabled by a noise-augmented rollout, that facilitates direct end-to-end training of these policies. Our approach supports both from-scratch and offline-to-online learning and achieves state-of-the-art performance on continuous control and robotic manipulation benchmarks, eliminating the need for common workarounds like policy distillation or surrogate objectives. Anonymized code is available at \url{https://anonymous.4open.science/r/SAC-FLOW}

Variance-dependent regret bounds for linear contextual bandits, which improve upon the classical $\tilde{O}(d\sqrt{K})$ regret bound to $\tilde{O}(d\sqrt{\sum_{k=1}^K\sigma_k^2})$, where $d$ is the context dimension, $K$ is the number of rounds, and $\sigma^2_k$ is the noise variance in round $k$, has been widely studied in recent years. However, most existing works focus on the regret upper bounds instead of lower bounds. To our knowledge, the only lower bound is from Jia et al. (2024), which proved that for any eluder dimension $d_{\textbf{elu}}$ and total variance budget $\Lambda$, there exists an instance with $\sum_{k=1}^K\sigma_k^2\leq \Lambda$ for which any algorithm incurs a variance-dependent lower bound of $\Omega(\sqrt{d_{\textbf{elu}}\Lambda})$. However, this lower bound has a $\sqrt{d}$ gap with existing upper bounds. Moreover, it only considers a fixed total variance budget $\Lambda$ and does not apply to a general variance sequence $\{\sigma_1^2,\ldots,\sigma_K^2\}$. In this paper, to overcome the limitations of Jia et al. (2024), we consider the general variance sequence under two settings. For a prefixed sequence, where the entire variance sequence is revealed to the learner at the beginning of the learning process, we establish a variance-dependent lower bound of $\Omega(d \sqrt{\sum_{k=1}^K\sigma_k^2 }/\log K)$ for linear contextual bandits. For an adaptive sequence, where an adversary can generate the variance $\sigma_k^2$ in each round $k$ based on historical observations, we show that when the adversary must generate $\sigma_k^2$ before observing the decision set $D_k$, a similar lower bound of $\Omega(d\sqrt{ \sum_{k=1}^K\sigma_k^2} /\log^6(dK))$ holds. In both settings, our results match the upper bounds of the SAVE algorithm (Zhao et al. 2023) up to logarithmic factors. Furthermore, if the adversary can generate the variance $\sigma_k$ after observing the decision set $D_k$, we construct a counter-example showing that it is impossible to construct a variance-dependent lower bound if the adversary properly selects variances in collaboration with the learner. Our lower bound proofs use a novel peeling technique that groups rounds by variance magnitude. For each group, we construct separate instances and assign the learner distinct decision sets. We believe this proof technique may be of independent interest.

Tool use represents a critical capability for AI agents, with recent advances focusing on leveraging reinforcement learning (RL) for test-time scaling to achieve better performance through more deliberate reasoning. However, there are some key challenges in current RL-based scaling approaches: (a) direct RL training often struggles to scale up thinking length sufficiently to solve complex problems, and (b) scaled-up models tend to overthink simpler problems, resulting in substantial token inefficiency. To address these challenges, we propose a novel training paradigm that first employs warm-up supervised fine-tuning to help models distinguish between simple and complex problems, followed by RL that enable models to automatically determine appropriate reasoning trajectories. Furthermore, to tackle the issue of automatic thinking-length scaling, we discover that entropy-based optimization objectives effectively maintain model diversity while successfully unlocking the model's scaling capabilities. Based on this insight, we introduce an entropy-based long-short reasoning fusion RL strategy. Our experiments on three benchmarks demonstrate that model successfully achieves auto-scaling for efficient tool use, achieving significant 9.8\% accuracy improvements while reducing computational overhead by ~81\%.

We study offline reinforcement learning (RL) with trajectory preferences, where the RL agent does not receive explicit rewards at each step but instead receives human-provided preferences over pairs of trajectories. Despite growing interest in preference-based reinforcement learning (PbRL), contemporary works cannot robustly learn policies in offline settings with poor data coverage and often lack algorithmic tractability. We propose a novel **M**odel-based **C**onservative **P**lanning (MCP) algorithm for offline PbRL, which leverages a general function class and uses a tractable conservative learning framework to improve the policy upon an arbitrary reference policy. We prove that, MCP can compete with the best policy within data coverage when the reference policy is supported by the data. To the best of our knowledge, MCP is the first provably sample-efficient and computationally tractable offline PbRL algorithm under partial data coverage, without requiring known transition dynamics. We further demonstrate that, with certain structural properties in PbRL dynamics, our algorithm can effectively exploit these structures to relax the partial data coverage requirement and improve regret guarantees. We evaluate MCP on a comprehensive suite of human-in-the-loop benchmarks in Meta-World. Experimental results show that our algorithm achieves competitive performance compared to state-of-the-art offline PbRL algorithms.

Model-based offline reinforcement learning (RL) aims to enhance offline RL with a dynamics model that facilitates policy exploration. However, model exploitation could occur due to inevitable model errors, which degrades algorithm performance. Adversarial model learning offers a theoretical framework to mitigate model exploitation by solving a maximin formulation, and RAMBO provides a practical implementation with model gradient. However, we empirically observe that severe Q-value underestimation and gradient explosion can occur in RAMBO with only slight hyperparameter tuning, suggesting that it tends to be overly conservative and suffers from unstable model updates. To address these issues, we propose RObust value-aware Model learning via Implicitly differentiable adaptive weighting (ROMI). Instead of updating the dynamics model with model gradient, ROMI introduces a novel robust value-aware model learning approach. This approach requires the dynamics model to predict future states with values close to the minimum Q-value within a scale-adjustable state uncertainty set, enabling controllable conservatism and stable model updates. To further improve out-of-distribution (OOD) generalization during multi-step rollouts, we propose implicitly differentiable adaptive weighting, a bi-level optimization scheme that adaptively achieves dynamics- and value-aware model learning. Empirical results on D4RL and NeoRL datasets show that ROMI significantly outperforms RAMBO and achieves competitive or superior performance compared to state-of-the-art methods on datasets where RAMBO typically underperforms.

Reinforcement learning (RL) has emerged as a powerful post-training paradigm for large language models (LLMs), yet its effectiveness varies significantly across base models. While incorporating a pre-RL supervised fine-tuning (SFT) phase can enhance RL training, key questions remain: how long should the SFT cold-start phase last, and is the SFT objective truly aligned with the requirements for effective RL preparation? In our analysis of cold-start dynamics, we uncover a key limitation: the SFT checkpoint with the highest evaluation performance often fails to maximize RL potential due to distributional forgetting—a phenomenon where the model drifts excessively away from the base model’s distribution even before traditional overfitting occurs. We identify diversity metrics, such as the entropy and self-BLEU, as more reliable early-stopping criteria than the standard performance-based checkpoint selection. Our findings show that SFT checkpoints with peak diversity consistently lead to superior post-RL results. Building on these insights, we introduce Adaptive Early-Stop Loss (AESL), a lightweight and dynamic cold-start method that balances the acquisition of new patterns with the preservation of the base model's distribution. AESL operates at both the token and subsequence levels, providing finer-grained control over the cold-start process. Experimental results on mathematical reasoning benchmarks demonstrate that diversity-based early stopping surpasses traditional performance-based SFT, while AESL further enhances RL preparation. By steering LLMs toward better initialization points for RL, AESL consistently achieves superior final performance compared to existing SFT and cold-start strategies. The code is publicly available at \url{https://github.com/LXXXXR/AESL}.

A safe and trustworthy use of Large Language Models (LLMs) requires an accurate expression of confidence in their answers. We propose a novel Reinforcement Learning approach that allows to directly fine-tune LLMs to express calibrated confidence estimates alongside their answers to factual questions. Our method optimizes a reward based on the logarithmic scoring rule, explicitly penalizing both over- and under-confidence. This encourages the model to align its confidence estimates with the actual predictive accuracy. The optimal policy under our reward design would result in perfectly calibrated confidence expressions. Unlike prior approaches that decouple confidence estimation from response generation, our method integrates confidence calibration seamlessly into the generative process of the LLM. Empirically, we demonstrate that models trained with our approach exhibit substantially improved calibration and generalize to unseen tasks without further fine-tuning, suggesting the emergence of general confidence awareness. We provide our training and evaluation code in the supplementary and will make it publicly available upon acceptance.

Large Language Models (LLMs) have been widely adopted as task planners for AI agents in sequential decision-making problems, leveraging their extensive world knowledge. However, the gap between their general knowledge and environment-specific requirements often leads to inaccurate plans. To address this, existing approaches rely on frequent LLM queries to iteratively refine plans based on immediate environmental feedback, which incurs substantial query costs. However, this refinement is typically guided by short-term environmental feedback, limiting LLMs from developing plans aligned with long-term rewards. We propose **Co**de Driven **P**lanning w**i**th Domain-Adaptive Sele**C**tor (CoPiC). Instead of relying on frequent queries, CoPiC employs LLMs to generate a diverse set of high-level planning programs, which iteratively produce and refine candidate plans. A trained domain-adaptive selector then evaluates these candidates and selects the one most aligned with long-term rewards for execution. Using high-level planning programs as planner and domain-adaptive selector as estimator, CoPiC improves planning while significantly reducing query costs. Results in ALFWorld, NetHack, and StarCraft II Unit Building show that CoPiC outperforms advanced LLM-based baselines, achieving an average (1) 19.14\% improvement in success rate and (2) 79.39\% reduction in token costs.

Self-supervised feature learning and pretraining methods in reinforcement learning (RL) often rely on information-theoretic principles, termed mutual information skill learning (MISL). These methods aim to learn a representation of the environment while also incentivizing exploration thereof. However, the role of the representation and mutual information parametrization in MISL is not yet well understood theoretically. Our work investigates MISL through the lens of identifiable representation learning by focusing on the Contrastive Successor Features (CSF) method. We prove that CSF can provably recover the environment's ground-truth features up to a linear transformation due to the inner product parametrization of the features and skill diversity in a discriminative sense. This first identifiability guarantee for representation learning in RL also helps explain the implications of different mutual information objectives and the downsides of entropy regularizers. We empirically validate our claims in MuJoCo and DeepMind Control and show how CSF provably recovers the ground-truth features both from states and pixels.

Imitation Learning (IL) enables agents to mimic expert behavior by learning from demonstrations. However, traditional IL methods require large amounts of medium-to-high-quality demonstrations as well as actions of expert demonstrations, both of which are often unavailable. To reduce this need, we propose Latent Wasserstein Adversarial Imitation Learning (LWAIL), a novel adversarial imitation learning framework that focuses on state-only distribution matching. It benefits from the Wasserstein distance computed in a dynamics-aware latent space. This dynamics-aware latent space differs from prior work and is obtained via a pre-training stage, where we train the Intention Conditioned Value Function (ICVF) to capture a dynamics-aware structure of the state space using a small set of randomly generated state-only data. We show that this enhances the policy's understanding of state transitions, enabling the learning process to use only one or a few state-only expert episodes to achieve expert-level performance. Through experiments on multiple MuJoCo environments, we demonstrate that our method outperforms prior Wasserstein-based IL methods and prior adversarial IL methods, achieving better results across various tasks.

Critic-free methods like GRPO reduce memory demands by estimating advantages from multiple rollouts but tend to converge slowly, as critical learning signals are diluted by an abundance of uninformative samples and tokens. To tackle this challenge, we propose the **Dynamic Dual-Level Down-Sampling (D$^3$S)** framework that prioritizes the most informative samples and tokens across groups to improve the efficiency of policy optimization. D$^3$S operates along two levels: (1) the sample-level, which selects a subset of rollouts to maximize advantage variance ($\text{Var}(A)$). We theoretically proved that this selection is positively correlated with the upper bound of the policy gradient norms, yielding higher policy gradients. (2) the token-level, which prioritizes tokens with a high product of advantage magnitude and policy entropy ($|A_{i,t}|\times H_{i,t}$), focusing updates on tokens where the policy is both uncertain and impactful. Moreover, to prevent overfitting to high-signal data, D$^3$S employs a dynamic down-sampling schedule inspired by curriculum learning. This schedule starts with aggressive down-sampling to accelerate early learning and gradually relaxes to promote robust generalization. Extensive experiments on Qwen2.5 and Llama3.1 demonstrate that integrating D$^3$S into advanced RL algorithms achieves state-of-the-art performance with generalization while requiring fewer samples and tokens across diverse reasoning benchmarks.

A central paradox in fine-tuning Large Language Models (LLMs) with Reinforcement Learning with Verifiable Reward (RLVR) is the frequent degradation of multi-attempt performance (Pass@k) despite improvements in single-attempt accuracy (Pass@1). This is often accompanied by catastrophic forgetting, where models lose previously acquired skills. Despite numerous proposed methods, the community's focus on the standard reverse KL-divergence has led to a surprising oversight: the potential of alternative f-divergences as a proactive solution has been largely unexamined. We argue that standard RLVR objectives—both those using the mode-seeking reverse KL-divergence and those forgoing a divergence term entirely—lack a crucial mechanism for knowledge retention. The reverse-KL actively accelerates this decay by narrowing the policy, while its absence provides no safeguard against the model drifting from its diverse knowledge base. We propose a fundamental shift in perspective: using the divergence term itself as the solution. Our framework, Diversity-Preserving Hybrid RL (DPH-RL), leverages mass-covering f-divergences (like forward-KL and JS-divergence) to function as a 'rehearsal mechanism'. By continuously referencing the initial policy, this approach forces the model to maintain broad solution coverage. Math and SQL generation experiments show that DPH-RL both improves in-domain Pass@1 and Pass@k scores and effectively prevents catastrophic forgetting on out-of-domain tasks. Additionally, DPH-RL is more training-efficient because it computes f-divergence using generator functions, requiring only sampling from the initial policy and no online reference model. Our work highlights a crucial, overlooked axis for improving RLVR, demonstrating that the proper selection of a divergence measure is a powerful tool for building more general and diverse reasoning models.

When there exists an unlearnable source of randomness (noisy-TV) in the environment, a naively intrinsic reward driven exploring agent gets stuck at that source of randomness and fails at exploration. Intrinsic reward based on uncertainty estimation or distribution similarity, while eventually escapes noisy-TVs as time unfolds, suffers from poor sample efficiency and high computational cost. Inspired by recent findings from neuroscience that humans monitor their improvements during exploration, we propose a novel method for intrinsically-motivated exploration, named Learning Progress Monitoring (LPM). During exploration, LPM rewards model improvements instead of prediction error or novelty, effectively rewards the agent for observing learnable transitions rather than the unlearnable transitions. We introduce a dual-network design that uses an error model to predict the expected prediction error of the dynamics model in its previous iteration, and use the difference between the model errors of the current iteration and previous iteration to guide exploration. We theoretically show that the intrinsic reward of LPM is zero-equivariant and a monotone indicator of Information Gain (IG), and that the error model is necessary to achieve monotonicity correspondence with IG. We empirically compared LPM against state-of-the-art baselines in noisy environments based on MNIST, 3D maze with 160x120 RGB inputs, and Atari. Results show that LPM's intrinsic reward converges faster, explores more states in the maze experiment, and achieves higher extrinsic reward in Atari. This conceptually simple approach marks a shift-of-paradigm of noise-robust exploration.

The dataset distributions in offline reinforcement learning (RL) often exhibit complex and multi-modal distributions, necessitating expressive policies to capture such distributions beyond widely-used Gaussian policies. To handle such complex and multi-modal datasets, in this paper, we propose Flow Actor-Critic, a new actor-critic method for offline RL, based on recent flow policies. The proposed method not only uses the flow model for actor as in previous flow policies but also exploits the expressive flow model for conservative critic acquisition to prevent Q-value explosion in out-of-data regions. To this end, we propose a new form of critic regularizer based on the flow behavior proxy model obtained as a byproduct of flow-based actor design. Leveraging the flow model in this joint way, we achieve new state-of-the-art performance for test datasets of offline RL including the D4RL and recent OGBench benchmarks.

Diffusion models have demonstrated remarkable success in high-fidelity image generation, yet aligning them with human preferences remains challenging. Direct Preference Optimization (DPO) offers a promising framework, but its effectiveness is critically hindered by noisy data arising from mislabeled preference pairs and individual preference pairs. We theoretically show that existing DPO objectives are equivalent to minimizing the Forward Kullback–Leibler (KL) divergence, whose mass-covering nature makes it intrinsically sensitive to such noise. To address this limitation, we propose $\alpha$-DPO, which reformulates preference alignment through the lens of $\alpha$-divergence. This formulation promotes mode-seeking behavior and bounds the influence of outliers, thereby enhancing robustness. Furthermore, we introduce a dynamic scheduling mechanism that adaptively adjusts $\alpha$ according to the observed preference distribution, providing data-aware noise tolerance during training. Extensive experiments on synthetic and real-world datasets validate that $\alpha$-DPO consistently outperforms existing baselines, achieving superior robustness and preference alignment.