ICLR 2026 Papers

Retrieval-augmented generation (RAG) improves performance on knowledge-intensive tasks but can be derailed by wrong, irrelevant, or conflicting retrieved text, causing models to rely on inaccurate evidence and cascade errors. We propose Knowledgeable-R1, a reinforcement-learning framework that explicitly trains large language models to use parametric knowledge (PK) to resist contextual interference while still exploiting external context when it is reliably helpful. Knowledgeable-R1 introduces a joint sampling scheme that generates paired responses with and without retrieval, and learns both local advantages (within each decoding regime) and global advantages under the same input to quantify when to ignore misleading context versus adopt it. We employ an asymmetric advantage transformation that amplifies exploratory behaviors toward parametric knowledge. Experiments show that \method significantly improves robustness and reasoning accuracy in knowledge conflict scenarios and general RAG scenarios, outperforming SOTA baselines by 23% in counterfactual scenarios, and without degradation when the retrieved context is fully accurate.

Large Language Models (LLMs) increasingly rely on Chain-of-Thought (CoT) reasoning to solve complex problems. Contrary to the common belief that longer CoTs always improve performance, we demonstrate that **longer is not always better**. Across both real-world LLMs and theoretical models, task accuracy follows an inverted U-shaped curve with respect to CoT length: performance rises initially but declines once reasoning chains become too long. Through controlled experiments, we uncover **scaling behaviors of the optimal CoT length**: it increases with task difficulty but decreases with model capability. This exposes a significant mismatch with current practice, where supervised training often reuses the same CoT data across models and tasks without adaptivity. We further show that Reinforcement Learning (RL) can mitigate this gap by dynamically calibrating CoT length, thereby improving accuracy and offering a new perspective on differences between supervised fine-tuning and RL training. To explain these phenomena, we introduce an error-accumulation analysis that characterizes how reasoning errors propagate across steps and derives the scaling behaviors of CoT length observed empirically. Building on these insights, we show that training with optimally sized CoTs and applying length-aware filtering during inference yields substantial improvements in performance. Taken together, these findings establish a principled explanation of the ''overthinking'' effect and yield practical guidelines for calibrating CoT length in accordance with task complexity and model capability.

Diffusion Q-Learning (DQL) has established diffusion policies as a high-performing paradigm for offline reinforcement learning, but its reliance on multi-step denoising for action generation renders both training and inference slow and fragile. Existing efforts to accelerate DQL toward one-step denoising typically rely on auxiliary modules or policy distillation, sacrificing either simplicity or performance. It remains unclear whether a one-step policy can be trained directly without such trade-offs. To this end, we introduce One-Step Flow Q-Learning (OFQL), a novel framework that enables effective one-step action generation during both training and inference, without auxiliary modules or distillation. OFQL reformulates the DQL policy within the Flow Matching (FM) paradigm but departs from conventional FM by learning an average velocity field that directly supports accurate one-step action generation. This design removes the need for multi-step denoising and backpropagation-through-time updates, resulting in substantially faster and more robust learning. Extensive experiments on the D4RL benchmark show that OFQL, despite generating actions in a single step, not only significantly reduces computation during both training and inference but also outperforms multi-step DQL by a large margin. Furthermore, OFQL surpasses all other baselines, achieving state-of-the-art performance in D4RL.

Reinforcement Learning with verifiable rewards (RLVR) has emerged as a primary learning paradigm for enhancing the reasoning capabilities of multi-modal large language models (MLLMs). However, during RL training, the enormous state space of MLLM and sparse rewards often leads to entropy collapse, policy degradation, or over-exploitation of suboptimal behaviors. This necessitates an exploration strategy that maintains productive stochasticity while avoiding the drawbacks of uncontrolled random sampling, yielding inefficient exploration. In this paper, we propose CalibRL, a hybrid-policy RLVR framework that supports controllable exploration with expert guidance, enabled by two key mechanisms. First, a distribution-aware advantage weighting scales updates by group rareness to calibrate the distribution, therefore preserving exploration. Meanwhile, the asymmetric activation function (LeakyReLU) leverages the expert knowledge as a calibration baseline to moderate overconfident updates while preserving their corrective direction. CalibRL increases policy entropy in a guided manner and clarifies the target distribution by estimating the on-policy distribution through online sampling. Updates are driven by these informative behaviors, avoiding convergence to erroneous patterns. Importantly, these designs help alleviate the distributional mismatch between the model’s policy and expert trajectories, thereby achieving a more stable balance between exploration and exploitation. Extensive experiments across eight benchmarks, including both in-domain and out-of-domain settings, demonstrate consistent improvements, validating the effectiveness of our controllable hybrid-policy RLVR training. Code is available at https://github.com/zhh6425/CalibRL.

Recent advances in applying reinforcement learning (RL) to large language models (LLMs) have led to substantial progress. In particular, a series of remarkable yet often counterintuitive phenomena have been reported in LLMs, exhibiting patterns not typically observed in traditional RL settings. For example, notable claims include that a single training example can match the performance achieved with an entire dataset, that the reward signal does not need to be very accurate, and that training solely with negative samples can match or even surpass sophisticated reward-based methods. However, the precise conditions under which these observations hold—and, critically, when they fail—remain unclear. In this work, we identify a key factor that differentiates RL observations: whether the pretrained model already exhibits strong *Model-Task Alignment*, as measured by pass@k accuracy on the evaluated task. Through a systematic and comprehensive examination of a series of counterintuitive claims, supported by rigorous experimental validation across different model architectures and task domains, our findings show that while standard RL training remains consistently robust across settings, many of these counterintuitive results arise only when the model and task already exhibit strong model-task alignment. In contrast, these techniques fail to drive substantial learning in more challenging regimes, where standard RL methods remain effective.

Reinforcement Learning with Verifiable Rewards (RLVR) is a powerful paradigm for enhancing the reasoning ability of Large Language Models (LLMs). Yet current RLVR methods often explore poorly, leading to premature convergence and entropy collapse. Moreover, they tend to produce poorly calibrated policies that remain confident in their generations regardless of correctness. To address this challenge, we introduce **Curiosity-Driven Exploration (CDE)**, a framework that leverages the model's intrinsic sense of curiosity to guide exploration. We formalize curiosity with signals from both the actor and the critic: for the actor, we use perplexity over its generated response, and for the critic, we use the variance of value estimates from a multi-head critic architecture. Both signals serve as an exploration bonus within the RLVR framework to guide the model. Our theoretical analysis shows that the actor-wise bonus inherently penalizes overconfident errors and promotes diversity among correct responses; moreover, we connect the critic-wise bonus to the well-established count-based exploration bonus in RL. Empirically, our method achieves an approximate **+3** point improvement over standard RLVR using GRPO/PPO on AIME benchmarks.

Recent advances in deep reinforcement learning (RL) have achieved strong results on high-dimensional control tasks, but applying RL to reachability problems raises a fundamental mismatch: reachability seeks to maximize the set of states from which a system remains safe indefinitely, while RL optimizes expected returns over a user-specified distribution. This mismatch can result in policies that perform poorly on low-probability states that are still within the safe set. A natural alternative is to frame the problem as a robust optimization over a set of initial conditions that specify the initial state, dynamics and safe set, but whether this problem has a solution depends on the feasibility of the specified set, which is unknown a priori. We propose Feasibility-Guided Exploration (FGE), a method that simultaneously identifies a subset of feasible initial conditions under which a safe policy exists, and learns a policy to solve the reachability problem over this set of initial conditions. Empirical results demonstrate that FGE learns policies with over 50% more coverage than the best existing method for challenging initial conditions across tasks in the MuJoCo simulator and the Kinetix simulator with pixel observations.

The rapid progress of large language models (LLMs) has laid the foundation for multimodal models. Nevertheless, visual language models (VLMs) still face significant computational overhead when scaled from images to the video domain. When video data is too large (due to high frame rates and long durations), the inference cost of models increases sharply. This severely hinders their deployment and application in environments that require rapid responses and have limited computation resources. Token compression for input videos is one of the promising directions, as effective compression schemes can significantly reduce computational overhead. Most existing compression methods are based on training-free token merging strategies in either the spatial or temporal dimension. Although these methods reduce computational overhead, their training-free nature inevitably leads to information loss during token compression, resulting in a significant performance drop. To address these challenges, we propose a Memory-Augmented Reinforcement Learning-based Token Compression (MARC) method for efficient video understanding that integrates structured retrieval with RL-based distillation. Our proposed MARC is a retrieve-then-compress method, which employs a Visual Memory Retriever (VMR) tool and a Compression Group Relative Policy Optimization (C-GRPO) training strategy. The Visual Memory Retriever first segments videos into event-level fragments and selects query-relevant clips. The C-GRPO distills reasoning ability from a Teacher Network to a Student Network by encouraging the output of the student network to match the performance of the teacher network. Extensive experiments on six video benchmarks demonstrate that our compression method achieves nearly identical accuracy to the 64-frame Qwen2.5-VL-3B baseline while using only one frame’s worth of tokens as input, resulting in a 95% reduction in visual tokens. Moreover, our approach reduces GPU memory usage by 72% and generation latency by 23.9%. These results demonstrate the strong potential of our compression method as a robust solution for RL-based post-training compression of large-scale models, enabling practical deployment in latency-sensitive and resource-constrained applications such as real-time video question answering, surveillance, and autonomous driving.

Large Language Models (LLMs) often hallucinate, generating content inconsistent with the input. Retrieval-Augmented Generation (RAG) and Reinforcement Learning with Human Feedback (RLHF) can mitigate hallucinations but require resource-intensive retrieval or large-scale fine-tuning. Decoding-based methods are lighter yet lack explicit hallucination control. To address this, we present \textbf{Token-Guard}, a token-level hallucination control method based on self-checking decoding. Token-Guard performs internal verification at each reasoning step to detect hallucinated tokens before they propagate. Candidate fragments are further evaluated in a latent space with explicit hallucination risk scoring, while iterative pruning and regeneration dynamically correct detected errors. Experiments on HALU datasets show Token-Guard substantially reduces hallucinations and improves generation accuracy, offering a scalable, lightweight solution for reliable LLM outputs. Our code is publicly available\footnote{\url{https://github.com/rhq945/Token-Guard}}.

Goal-Conditioned Reinforcement Learning (GCRL) mitigates the difficulty of reward design by framing tasks as goal reaching rather than maximizing hand-crafted reward signals. In this setting, the optimal goal-conditioned value function naturally forms a quasimetric, motivating Quasimetric RL (QRL), which constrains value learning to quasimetric mappings and enforces local consistency through discrete, trajectory-based constraints. We propose Eikonal-Constrained Quasimetric RL (Eik-QRL), a continuous-time reformulation of QRL based on the Eikonal Partial Differential Equation (PDE). This PDE-based structure makes Eik-QRL trajectory-free, requiring only sampled states and goals, while improving out-of-distribution generalization. We provide theoretical guarantees for Eik-QRL and identify limitations that arise under complex dynamics. To address these challenges, we introduce Eik-Hierarchical QRL (Eik-HiQRL), which integrates Eik-QRL into a hierarchical decomposition. Empirically, Eik-HiQRL achieves state-of-the-art performance in offline goal-conditioned navigation and yields consistent gains over QRL in manipulation tasks, matching temporal-difference methods.

Reinforcement learning (RL) has become an increasingly important paradigm for improving large language models (LLMs) on alignment, reasoning, and coding tasks, yet it remains extremely costly. The majority of training time is spent on rollouts. Allowing actor and policy distributions to differ could unlock substantial scalability and efficiency benefits, such as supporting large-batch or asynchronous training, and even enabling a lightweight rollout model. However, existing importance sampling–based corrections for distribution mismatch suffer from an inherent trade-off between stability and training performance. To tackle this problem, we propose Jackpot, which leverages Optimal Budget Rejection Sampling to directly reduce the gap between actor and policy distributions. For efficiency and stability in practical training, We introduce an efficient probability estimation strategy based on Top-$K$ logits with batch bias correction, and designs a stabilized Jackpot-PPO loss that jointly accounts for both the importance sampling ratio and the trust-region constraint in PPO. Empirically, our method achieves stable improvements in large-batch and asynchronous training, and in extreme off-policy training it substantially delays the onset of collapse and delivers competitive performance. Specifically, we achieve 20\% improvement on AMC benchmarks and ~8\% AIME benchmarks over the off-policy baseline under 128$\times$ actor-policy update ratio for Qwen3-4B-Base and 64$\times$ for Qwen3-8B-Base, while achieving greater stability and better performance than prior off-policy RL methods under extreme settings.

Multi-Agent Systems (MAS) offer a powerful paradigm for solving complex problems through distributed reasoning and collaboration. However, their effectiveness is often hindered by the challenge of optimizing interactions among agents. To address this, we introduce AgentPO, a novel framework that directly optimizes agent collaboration. AgentPO employs reinforcement learning to train a specialized Collaborator agent, which refines its interaction policy to enhance overall system performance within a fixed multi-agent topology. We evaluated AgentPO on multiple mathematical reasoning tasks, where it consistently outperformed strong baselines. With Llama-3.2-3B-Instruct as the actor model, AgentPO achieves accuracy improvements of +1.8\% and +7.2\% over strong baselines like Role Assignment and EvoAgent, respectively. When using the larger Llama-3.1-8B-Instruct model, these gains increase to +5.6\% and +11.3\%. Crucially, AgentPO achieves these results with remarkable efficiency: it requires only 500 training samples and operates at just 7.8\% of EvoAgent's inference cost, highlighting its superior scalability and practicality.

Large Language Models (LLMs) trained with reinforcement learning and verifiable rewards have achieved strong results on complex reasoning tasks. Recent work extends this paradigm to a multi-agent setting, where a meta-thinking agent proposes plans and monitors progress while a reasoning agent executes subtasks through sequential conversational turns. Despite promising performance, we identify a critical limitation: lazy agent behavior, in which one agent dominates while the other contributes little, undermining collaboration and collapsing the setup to an ineffective single agent. In this paper, we first provide a theoretical analysis showing why lazy behavior naturally arises in multi-agent reasoning. We then introduce a stable and efficient method for measuring causal influence, helping mitigate this issue. Finally, as collaboration intensifies, the reasoning agent risks getting lost in multi-turn interactions and trapped by previous noisy responses. To counter this, we propose a verifiable reward mechanism that encourages deliberation by allowing the reasoning agent to discard noisy outputs, consolidate instructions, and restart its reasoning process when necessary. Extensive experiments demonstrate that our framework alleviates lazy agent behavior and unlocks the full potential of multi-agent framework for complex reasoning tasks.

Long-horizon sparse-reward tasks, such as goal-conditioned or robot manipulation tasks, remain challenging in offline reinforcement learning due to the credit assignment problem. Hierarchical methods have been proposed to tackle this problem by introducing sub-goal planning guided by value functions, which in principle can shorten the effective planning horizon for both high-level and low-level planners, and thereby avoiding the credit assignment problem. However, we demonstrate that the sub-goal selection mechanism is unreliable, as it relies on value functions suffering from generalization noise, which misguides value estimation and thus leads to sub-optimal sub-goals. In this work, to provide more reliable sub-goals, we novelly introduce a reliability-driven decision mechanism, and propose Reliability-Driven HRL (RD-HRL) as the solution. The reliability-driven decision mechanism provide decision-level targets from transition regions, thereby providing noise-immune decision spaces for high-level policy, ensuring the reliability of sub-goals (which are termed as action-level targets in this paper). Comprehensive experimental results demonstrate that our approach RD-HRL outperforms baseline methods across multiple benchmarks, highlighting the competitive advantages of RD-HRL. Our code is anonymously available at \url{https://anonymous.4open.science/r/RD-HRL-243D }.

Deep reinforcement learning (RL) suffers from plasticity loss severely due to the nature of non-stationarity, which impairs the ability to adapt to new data and learn continually. Unfortunately, our understanding of how plasticity loss arises, dissipates, and can be dissolved remains limited to empirical findings, leaving the theoretical end underexplored. To address this gap, we study the plasticity loss problem from the theoretical perspective of network optimization. By formally characterizing the two culprit factors in online RL process: the non-stationarity of data distributions and the non-stationarity of targets induced by bootstrapping, our theory attributes the loss of plasticity to two mechanisms: the rank collapse of the Neural Tangent Kernel (NTK) Gram matrix and the Θ(1/k) decay of gradient magnitude. The first mechanism echoes prior empirical findings from the theoretical perspective and sheds light on the effects of existing methods, e.g., network reset, neuron recycle, and noise injection. Against this backdrop, we focus primarily on the second mechanism and aim to alleviate plasticity loss by addressing the gradient attenuation issue, which is orthogonal to existing methods. We propose Sample Weight Decay (SWD) --- a lightweight method to restore gradient magnitude, as a general remedy to plasticity loss for deep RL methods based on experience replay. In experiments, we evaluate the efficacy of SWD upon TD3, SAC with SimBa architecture in MuJoCo and DeepMind Control Suite tasks. The results demonstrate that SWD effectively alleviates plasticity loss and consistently improves learning performance across various configurations of deep RL algorithms, UTD, network architectures, and environments, achieving SOTA performance on challenging DMC Humanoid tasks.

The foundation of reproducible science lies in protocols that are precise, logically ordered, and executable. The autonomous generation of these protocols through natural language queries could greatly improve the efficiency of the reproduction process. However, current leading large language models (LLMs) often generate incomplete or inconsistent protocols, limiting their utility. To address this limitation, we first introduce SciRecipe, a large-scale dataset of over 12K structured protocols spanning 27 biological subfields and encompassing both comprehension and problem-solving tasks. To further improve protocol generation, we propose the "Sketch-and-Fill" paradigm, which separates analysis, structuring, and expression to ensure each step is explicit and verifiable. Complementing this, the structured component-based reward mechanism evaluates step granularity, action order, and semantic fidelity, aligning model optimization with experimental reliability. Building on these components, we develop Thoth, trained through a staged Knowledge-to-Action process that progresses from knowledge acquisition to operational reasoning and ultimately to robust, executable protocol generation. Across multiple benchmarks, Thoth consistently surpasses both proprietary and open-source LLMs, achieving significant improvements in step alignment, logical sequencing, and semantic accuracy. Our approach paves the way for reliable scientific assistants that bridge knowledge with experimental execution.

Diffusion large language models (dLLMs) are compelling alternatives to autoregressive (AR) models because their denoising models operate over the entire sequence. The global planning and iterative refinement features of dLLMs are particularly useful for code generation. However, current training and inference mechanisms for dLLMs in coding are still under-explored. To demystify the decoding behavior of dLLMs and unlock their potential for coding, we systematically investigate their denoising processes and reinforcement learning (RL) methods. We train a 7B dLLM, DiffuCoder, on 130B tokens of code. Using this model as a testbed, we analyze its decoding behavior, revealing how it differs from that of AR models: (1) dLLMs can decide how causal their generation should be without relying on semi-AR decoding, and (2) increasing the sampling temperature diversifies not only token choices but also their generation order. This diversity creates a rich search space for RL rollouts. For RL training, to reduce the variance of token log-likelihood estimates and maintain training efficiency, we propose coupled-GRPO, a novel sampling scheme that constructs complementary mask noise for completions used in training. In our experiments, coupled-GRPO significantly improves DiffuCoder's performance on code generation benchmarks (+4.4\% on EvalPlus) and reduces reliance on AR bias during decoding. Our work provides deeper insight into the machinery of dLLM generation and offers an effective, diffusion-native RL training framework.

Reward models (RMs) are crucial in reinforcement learning from human feedback (RLHF) to align large language models (LLMs) with human values. However, RM training data is commonly recognized as low-quality, always containing preference conflicts and inductive biases, such as response length or speaking style, which can easily lead to reward overfitting and hacking. A few recent RM debiasing methods either target merely a single specific type of preference bias or only address simple linear bias relations such as Pearson coefficients. To mitigate more complicated inductive bias of reward modeling, inspired by the information bottleneck, we introduce a novel information-theoretic debiasing method called **D**ebiasing via **I**nformation optimization for **R**M (DIR). More specifically, our method trains RMs by maximizing the mutual information (MI) between preference prediction and input response pairs, while minimizing the MI between RM outputs and biased attributes of preference inputs. With the theoretical justification of information theory, DIR can handle different types of bias with more comprehensive non-linear correlations, enlarging its real-world application scenarios. In experiments, we verify the effectiveness of DIR with three types of inductive biases: response length, sycophancy, and format. Based on the numerical results, we discover that DIR can not only effectively diminish target inductive biases but also improve RLHF performances on various benchmarks with better generalization abilities.

Learning-based methods for routing have gained significant attention in recent years, both in single-objective and multi-objective contexts. Yet, existing methods are unsuitable for routing on multigraphs, which feature multiple edges with distinct attributes between node pairs, despite their strong relevance in real-world scenarios. In this paper, we propose two graph neural network-based methods to address multi-objective routing on multigraphs. Our first approach operates directly on the multigraph by autoregressively selecting edges until a tour is completed. The second model, which is more scalable, first simplifies the multigraph via a learned pruning strategy and then performs autoregressive routing on the resulting simple graph. We evaluate both models empirically, across a wide range of problems and graph distributions, and demonstrate their competitive performance compared to strong heuristics and neural baselines.