ICML 2026 论文列表

Test-time reinforcement learning generates multiple candidate answers via repeated rollouts and performs online updates using pseudo-labels constructed by majority voting. To reduce overhead and improve exploration, prior work introduces tree-structured rollouts, which share reasoning prefixes and branch at key nodes to improve sampling efficiency. However, this paradigm still faces two challenges: (1) high-entropy branching can trigger rollout collapse, where the branching budget concentrates on a few trajectories with consecutive high-entropy segments, rapidly reducing the number of effective branches; (2) early pseudo-labels are noisy and biased, which can induce self-reinforcing overfitting, causing the policy to sharpen prematurely and suppress exploration. To address these issues, we propose Entropy–Confidence Hybrid Group Relative Policy Optimization (ECHO). During rollout, ECHO jointly leverages local entropy and group-level confidence to adaptively control branch width, and further introduces online confidence-based pruning to terminate persistently low-confidence branches, avoiding high-entropy traps and mitigating collapse. During policy updates, ECHO employs confidence-adaptive clipping and an entropy–confidence hybrid advantage shaping approach to enhance training robustness and mitigate early-stage bias. Experiments demonstrate that ECHO achieves consistent gains on multiple mathematical and visual reasoning benchmarks, and generalizes more effectively under a limited rollout budget.

Recent advances in reinforcement learning for code generation have made robust environments essential to prevent reward hacking. As LLMs increasingly serve as evaluators in code-based RL, their ability to detect reward hacking remains understudied. In this paper, we propose a novel taxonomy of reward exploits spanning across 54 categories and introduce TRACE (Testing Reward Anomalies in Code Environments), a synthetically curated and human-verified benchmark containing 517 testing trajectories. Unlike prior work that evaluates reward hack detection in isolated classification scenarios, we contrast these evaluations with a more realistic, contrastive anomaly detection setup on TRACE. Our experiments reveal that models capture reward hacks more effectively in contrastive settings than in isolated classification settings, with GPT-5.2 with highest reasoning mode achieving the best detection rate at 63%, up from 45% in isolated settings on TRACE. Building on this insight, we demonstrate that state-of-the-art models struggle significantly more with semantically contextualized reward hacks compared to syntactically contextualized ones. We further conduct qualitative analyses of model behaviors, as well as ablation studies showing that the ratio of benign to hacked trajectories and analysis cluster sizes substantially impact detection performance. We release the benchmark and evaluation harness to enable the community to expand TRACE and evaluate their models.

We propose a distributional framework for offline Inverse Reinforcement Learning (IRL) that jointly models uncertainty over reward functions and full distributions of returns. Unlike conventional IRL approaches that recover a deterministic reward estimate or match only expected returns, our method captures richer structure in expert behavior, particularly in learning the reward distribution, by minimizing first-order stochastic dominance (FSD) violations and thus integrating distortion risk measures (DRMs) into policy learning, enabling the recovery of both reward distributions and distribution-aware policies. This formulation is well-suited for behavior analysis and risk-aware imitation learning. Theoretical analysis show that the algorithm converge with $\mathcal{O}(\varepsilon^{-2})$ iteration complexity. Empirical results on synthetic benchmarks, real-world neurobehavioral data, and MuJoCo control tasks demonstrate that our method recovers expressive reward representations and achieves state-of-the-art imitation performance.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a central paradigm for scaling LLM reasoning, yet its optimization often suffers from training instability and suboptimal convergence. Through a systematic dissection of clipping-based GRPO-style objectives, we identify the rigid clipping decision induced by hard clipping as a key practical bottleneck in the studied RLVR setups. Specifically, our analysis suggests that informative signals can lie in the **near-boundary** region just beyond the clipping threshold, and are therefore discarded by the standard hard-clipping rule. Motivated by this diagnosis, we propose **Near-boundary Stochastic Rescue (NSR)**, a minimal, plug-and-play modification that stochastically retains these slightly out-of-bound tokens to recover lost signals. While NSR, via stochastic sampling, can be interpreted as inducing an implicit gradient decay in expectation, our ablations reveal that its stochastic, boundary-local rescue mechanism is consistently more effective than deterministic gradient decay. Validated by extensive experiments across model sizes from 7B to 30B and both dense and MoE architectures, as a plug-and-play solution, NSR substantially improves training stability and delivers consistent gains over strong baselines such as DAPO and GSPO. Our code is publicly available at https://github.com/qwenpilot/NSR.

Large language model (LLM) agents increasingly rely on external tools such as search engines to solve complex, multi-step problems, yet their rollouts are structurally heterogeneous: variations in tool-call number, placement, and outcomes induce distinct behaviors and reward distributions. As a result, policy gradient methods with a single global baseline suffer from *cross-stratum bias*, an "apples-to-oranges'' comparison that distorts credit assignment and impedes exploration. To address this issue, we propose *Stratified GRPO*. Its core component, *Stratified Advantage Normalization* (SAN), partitions trajectories into homogeneous strata based on structural properties and computes advantages locally within each stratum, ensuring comparisons only among true peers. We show that SAN eliminates cross-stratum bias, yields conditionally unbiased unit-variance estimates within strata, and preserves the global unbiasedness and unit-variance properties of standard normalization, resulting in a more reliable learning signal. To improve robustness in finite-sample regimes, we further linearly blend SAN with the global estimator. Experiments on factual QA and deep-research agent benchmarks demonstrate that Stratified GRPO consistently outperforms GRPO by up to 12.6 points, achieving higher training rewards, improved training stability, and more effective search policies. These results establish structure-aware advantage normalization as an effective correction for RL of search agents with structurally heterogeneous rollouts.

Model-based representations recently stand out as a promising framework that embeds latent dynamics information into the representations for downstream off-policy actor-critic learning. It implicitly combines the advantages of both model-free and model-based approaches while avoiding the training costs associated with model-based methods. Nevertheless, existing model-based representation methods can fail to capture sufficient information about relevant variables and can overfit to early experiences in the replay buffer. These incur biases in representation and actor-critic learning, leading to inferior performance. To address this, we propose Debiased model-based Representations for Q-learning, tagged DR.Q algorithm. DR.Q explicitly maximizes the mutual information between the representations of the current state-action pair and the next state besides minimizing their deviations, and samples transitions with faded prioritized experience replay. We evaluate DR.Q on numerous continuous control benchmarks with a single set of hyperparameters, and the results demonstrate that DR.Q can match or surpass recent strong baselines, sometimes outperforming them by a large margin. Our code is available at https://github.com/dmksjfl/DR.Q.

Reinforcement learning (RL) has emerged as a crucial approach for enhancing the capabilities of large language models. However, in Mixture-of-Experts (MoE) models, RL training tends to be unstable and can even collapse entirely. We analyze the training-inference consistency of MoE models and identify a notable discrepancy in routing behaviors between the two phases. To address this issue, we propose Rollout Routing Replay (R3), a novel and effective method that records routing distributions from the inference engine and replays them during training. R3 significantly reduces training-inference policy KL divergence and mitigates extreme discrepancies without compromising training speed. Extensive experiments on various settings confirm that R3 succeeds in stabilizing RL training, preventing collapse and outperforming strong baselines. R3 is orthogonal to most policy optimization algorithm improvements, allowing it to be used in conjunction with them. We believe this work can offer a new solution for stabilizing RL in MoE model.

In many practical reinforcement learning (RL) environments, observations are far higher-dimensional than the variables that matter for control. In this work, we ask: can we learn representations that capture only control-relevant features of the environment? We study this question through the \emph{empowerment} objective, which maximizes an agent’s influence over the environment and is widely used for unsupervised skill learning. We show that empowerment agents induce two distinct representations --- forward and backward --- that capture complementary aspects of the state, and both of which are invariant to control-irrelevant features. Thus, empowerment maximization leads agents to learn an implicit, \emph{control-centric} model of the world. Our analysis highlights the importance of learning representations through interaction rather than from passive datasets: interaction aimed at maximizing control is essential for learning useful invariance properties, a perspective that aligns closely with the causal learning literature.

We study offline black-box optimization (BBO), aiming to discover improved designs from an offline dataset of designs and labels, a problem common in robotics, DNA, and materials science with limited labeled samples. While recent work applies autoregressive LLMs to BBO by formatting tasks as natural-language prompts, their left-to-right design generation struggles to capture the strong bidirectional dependencies inherent in design problems. To address this, we propose adapting diffusion LLMs to offline BBO to leverage their bidirectional modeling capabilities. However, a domain gap exists between the natural text pre-training of diffusion LLMs and the heterogeneous signals in BBO (prompts, designs, and labels). To bridge this gap, we construct a unified prompt–-response corpus and introduce delimiter tokens to explicitly mark field boundaries for domain adaptation. We further propose a two-stage post-training framework to align the diffusion LLM generation with high-label designs. The first stage performs supervised fine-tuning on the unified dataset via masked-response prediction, and the second stage adopts reinforcement learning with rewards defined by label improvements. Our method achieves state-of-the-art results on Design-Bench under small-data settings. Code for our work is available here: https://github.com/zpointS/DiBO.

We introduce Deep Flow Networks (DFNs), a new class of discrete function approximators. DFNs are inspired by and generalize minimum-cost flow value functions that map node imbalances on a subset of nodes to the optimal flow cost. Such functions are known to be M-convex (Murota2003) and admit efficient optimization. On the theoretical side, we prove that DFNs are universal approximators for discrete functions on $\mathbb{Z}^d$ that admit convex extensions to $\mathbb{R}^d$, and characterize their optimization complexity in terms of their deviation from the M-convex regime. Guided by these results, we develop a practical DFN implementation for learning from data. Finally, we evaluate our implementation empirically on data from different ground-truth functions, showing that DFNs achieve strong approximation accuracy while being substantially faster to optimize than benchmark approaches.

Autonomous 3D indoor scene synthesis breaks down in non-convex rooms with tightly coupled spatial constraints. Data-driven generators lack topological priors for long-horizon planning, while iterative agents fragment semantics and become geometrically brittle. We present \textbf{ZoneMaestro}, a unified framework that shifts the paradigm from object-centric synthesis to Zone-Graph Orchestration. By internalizing a novel zone-based logic, ZoneMaestro translates high-level semantic intent into functional zones and topological constraints, enabling robust adaptation to diverse architectural forms. To support this, we construct \textit{Zone-Scene-10K}, a large-scale dataset enriched with explicit Zone-Graph annotations. We further introduce an \textit{Alternating Alignment Strategy} that cycles between reasoning internalization and Zone-Aware Group Relative Policy Optimization (\textit{Z-GRPO}), effectively reconciling the tension between semantic richness and geometric validity without relying on external physics engines. To rigorously evaluate spatial intelligence beyond convex primitives, we formally define the task of \textbf{Intricate Spatial Orchestration} and release SCALE, a stress-test benchmark for irregular indoor scenarios with complex, dense spatial relations. Extensive experiments demonstrate that ZoneMaestro resolves the density-safety dichotomy, significantly outperforming state-of-the-art baselines in both structural coherence and intent adherence.

This paper studies a constrained linear best arm identification problem with covariate selection in the fixed-confidence setting, where each arm is evaluated across multiple performance metrics. The mean performance of each metric depends linearly on the feature vectors of both arms and covariates. The goal is to identify the arm with the highest expected value of one targeted metric while ensuring that the means of the remaining metrics stay below specified thresholds for each covariate. We first establish an instance-dependent lower bound on the sample complexity, formulated as a multi-level optimization problem that captures both feasibility and optimality. We then prove that this bound is tight by designing an algorithm that asymptotically matches it. Since the original algorithm is computationally intensive, we develop a relaxed version of the bound through a surrogate optimization problem and derive its convex dual. Using this bound, we propose a duality-based decomposition algorithm that is computationally efficient, updating only two coordinates and performing a single gradient step per iteration. We further show that the algorithm achieves the relaxed bound in theory and demonstrates its practical effectiveness through numerical experiments.

Humans often specify tasks incompletely, so assistants must know when and how to ask clarifying questions. However, effective clarification remains challenging in software engineering tasks as not all missing information is equally valuable, and questions must target information users can realistically provide. We study clarification in real software engineering tasks by quantifying which types of information most affect task success and which questions elicit useful responses from simulated users. Using Shapley attribution and distributional comparisons, we identify two key properties of effective clarification: task relevance (which information predicts success) and user answerability (what users can realistically provide). We operationalize these properties as multi-stage reinforcement learning rewards to train CLARITI, an 8B-parameter clarification module, that matches GPT-5’s resolution rate on underspecified issues while generating 41% fewer questions. Our results suggest that grounding reward design in empirical analysis of information impact and user answerability improves clarification efficiency.

Diffusion Large Language Models (dLLMs) often exhibit severe instability during Group Relative Policy Optimization (GRPO) training, limiting the effectiveness of reinforcement learning for improving reasoning capabilities. In dLLMs, the importance ratios used by GRPO are derived from finite-sample estimates rather than exact likelihoods, making them inherently noisy. In this paper, we show that GRPO is highly sensitive to this noise, which drives training instability. Through theoretical analysis and empirical evidence, we identify a self-reinforcing instability loop in which noisy importance ratios induce gradient spikes and policy drift, further amplifying future importance ratio estimation variance. To address this issue, we propose StableDRL, a novel reinforcement learning framework for dLLMs. StableDRL stabilizes training via (i) unconditional clipping to suppress outlier-induced gradient spikes, and (ii) self-normalization to constrain gradients within the convex hull of per-sample updates. We further extend StableDRL to block-wise diffusion models via a staircase attention mechanism. StableDRL is the first method that enables stable, full-parameter reinforcement learning for dLLMs. It achieves the state-of-the-art performance, outperforming prior best full-attention baselines by 6% on MATH500 and block-diffusion baselines by 25.6% on AIME.

Most value-based and actor-critic reinforcement learning methods rely on Bellman-style recursions, yet these recursions collapse under non-exponential discounting common in human preferences and survival processes. We show the breakdown is structural: exponential discounting sits at a fragile intersection of multiplicativity and time homogeneity, and violating either property breaks standard dynamic programming. To overcome this, we propose **Pontryagin-Guided Direct Policy Optimization (PG-DPO)**, a variational framework that abandons recursion and couples the Pontryagin Maximum Principle with Monte Carlo rollouts via an *Adjoint-MC projection* enforcing pointwise Hamiltonian maximization. Across multidimensional hyperbolic and survival-discount benchmarks, PG-DPO improves accuracy and stability where equation-driven solvers and critic-based baselines diverge.

State-of-the-art model-based Reinforcement Learning (RL) approaches either use gradient-free, population-based methods for planning, learned policy networks, or a combination of policy networks and planning. Hybrid approaches that combine Model Predictive Control (MPC) with a learned model and a policy prior to leverage the advantages of both paradigms have shown promising results. However, these approaches typically rely on gradient-free optimization methods, which can be computationally expensive for high-dimensional control tasks. While gradient-based methods are a promising alternative, recent works have empirically shown that gradient-based methods often perform worse than their gradient-free counterparts. We propose Dream-MPC, a novel approach that generates few candidate trajectories from a rolled-out policy and optimizes each trajectory by gradient ascent using a learned world model, uncertainty regularization and amortization of optimization iterations over time by reusing previously optimized actions. Our results on 24 continuous control tasks show that Dream-MPC can significantly improve the performance of the underlying policy and can outperform gradient-free MPC and state-of-the-art baselines. Code and videos are available at https://dream-mpc.github.io.

Token-level credit assignment remains a key obstacle for reinforcement learning (RL) in large language models (LLMs), where RL recipes typically treat all tokens equally, failing to distinguish decisive reasoning steps from routine formatting or fluent filler. Recent attempts leverage model-internal signals to assign finer-grained credit, but these are often point-wise heuristics that ignore the global structure of information propagation. We propose FlowTracer, an RL framework that traces answer-targeted reasoning flow on an attention-induced directed acyclic graph in which nodes correspond to tokens and edge capacities come from aggregated attention weights and derives token credit from this global structure. The edge capacities are reweighted to retain only the influence that can reach the answer region, while enforcing local flow conservation so intermediate tokens neither lose nor gain effective mass due to path length or irrelevant branches. On this graph, FlowTracer extracts an information-flow backbone connecting the question to the answer and scores tokens by flow throughput, revealing high-impact hubs and aggregation checkpoints that mediate long-range dependencies. These derived importances are used to shape token-level rewards, enabling learning signals to focus precisely on the tokens that route information toward (or away from) correct answers and delivering consistent performance gains across a range of reasoning tasks.

Long-term fairness algorithms aim to satisfy fairness beyond static and short-term notions by accounting for the dynamics between decision-making policies and population behavior. Most previous approaches evaluate performance and fairness measures from observable features and a label, which is assumed to be fully observed. However, in scenarios such as hiring or lending, the labels (e.g., ability to repay the loan) are _selective labels_ as they are only revealed based on positive decisions (e.g., when a loan is granted). In this paper, we study long-term fairness in the selective labels setting and analytically show that naive solutions do not guarantee fairness. To address this gap, we then introduce a novel framework that leverages both the observed data and a label predictor model to estimate the true fairness measure value by decomposing it into the observed fairness and bias from label predictions. This allows us to derive sufficient conditions to satisfy true fairness from observable quantities by using the confidence in the predictor model. Finally, we rely on our theoretical results to propose a novel reinforcement learning algorithm for effective long-term fair decision-making with selective labels. In semisynthetic environments, the proposed algorithm reached comparable fairness and performance to an agent with oracle access to the true labels.

Reinforcement Learning (RL) has shown promise for aligning Large Language Models (LLMs) to follow instructions with various constraints. Despite the encouraging results, RL improvement inevitably relies on sampling successful, high-quality responses; however, the initial model often struggles to generate responses that satisfy all constraints due to its limited capabilities, yielding sparse or indistinguishable rewards that impede learning. In this work, we propose ***H**indsight **i**nstruction **R**eplay* (HiR), a novel sample-efficient RL framework for complex instruction following tasks, which employs a *select*-then-*rewrite* strategy to *replay failed attempts as successes* based on the constraints that have been satisfied in hindsight. We perform RL on these replayed samples as well as the original ones, theoretically framing the objective as dual-preference learning at both the instruction- and response-level to enable efficient optimization using only a binary reward signal. Extensive experiments demonstrate that the proposed HiR yields promising results across different instruction following tasks, while requiring less computational budget. Our code and dataset are available at https://github.com/sastpg/HIR.