ICML 2026 论文列表

Popular offline reinforcement learning (RL) methods rely on *explicit conservatism*, penalizing out-of-dataset actions or restricting rollout horizons. We question the universality of this principle and revisit a complementary Bayesian perspective for test-time adaptation. By modeling a posterior over world models and training a history-dependent agent to maximize expected return, the Bayesian approach directly addresses epistemic uncertainty without explicit conservatism. We first illustrate in a bandit setting that Bayesianism excels on low-quality datasets where conservatism fails. Scaling to realistic tasks, we find that *long-horizon rollouts* are essential to control value overestimation once conservatism is removed. We introduce design choices that enable learning from long-horizon rollouts while mitigating compounding model errors, yielding our algorithm, NEUBAY, grounded in the neutral Bayesian principle. On D4RL and NeoRL benchmarks, NEUBAY is competitive with leading conservative algorithms, achieving new state-of-the-art on 7 datasets with rollout horizons of several hundred steps. Finally, we characterize datasets by quality and coverage to identify when NEUBAY is preferable to conservative methods.

Self-reflection is a powerful mechanism for credit assignment in human learning, converting sparse outcome feedback into actionable guidance. However, its potential for post-training Large Language Models (LLMs) remains underexplored. We propose Self-Reflective Policy Optimization (SRPO), a framework that internalizes this capability. SRPO enables LLMs to analyze their own completed trajectories, synthesize errors into concise "reflection patches," and use these reflection-conditioned rollouts as high-quality, on-policy distillation targets. This process effectively transforms sparse terminal supervision into dense, token-level learning signals without requiring external critics, separate reward models, or larger teacher models. We demonstrate that SRPO achieves state-of-the-art performance across mathematical reasoning and long-horizon agentic benchmarks with exceptional data efficiency. Using a Qwen3-8B base model, SRPO attains 73.3\% on AIME’24 using only 8\% (0.08$\times$) of the training FLOPs required by scaled supervised fine-tuning, while significantly improving success rates on WebShop (64.7\%), ALFWorld (76.8\%), and SWE-Bench-Lite (31.2\%).

We study reward poisoning attacks in reinforcement learning (RL), where an adversary manipulates rewards under a limited budget to induce a target agent to learn a policy aligned with the attacker's objectives. Most prior work focuses on *constructing* successful attacks, providing sufficient conditions under which poisoning is effective, while offering limited understanding of when such targeted attacks are fundamentally infeasible. In this paper, we provide the first characterization of reward-poisoning attackability in linear MDPs, establishing both necessary and sufficient conditions for whether a target policy can be induced within a bounded attack budget. This draws a clear boundary between the *vulnerable* RL instances and *intrinsically robust* ones, which cannot be attacked without high costs even when the learner uses standard, non-robust RL algorithms. We further demonstrate our framework beyond synthetic linear MDPs by approximating deep RL environments as linear MDPs. We show that our theoretical framework effectively distinguishes vulnerability, demonstrating how our theoretical predictions have practical significance.

A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios, and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call *early experience*: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm, we study two strategies of using such data: (1) implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. Evaluation across eight diverse environments and multiple model families shows that our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, making it a practical bridge between imitation learning and fully experience-driven agents.

Model-Based Reinforcement Learning yields sample efficiency via latent imagination, yet remains constrained by **Historical Tethering**: imagination is typically initialized from observed states. This creates a learning asymmetry, where the world model's manifold discovery outpaces the policy's sparse-reward optimization. We propose **Mind Dreamer (MD)**, a framework that instantiates **Active Causal Intervention** to transcend Markovian continuity. MD reformulates discovery as the minimization of a global Relay Expected Free Energy. Instead of initializing from historical data, it draws initial states from an adversarial generator $s_0 \sim p_{gen}(\cdot)$, creating non-continuous **latent jumps** to epistemic blind spots that are physically plausible yet cognitively challenging. We derive **Relay Value Function** and **Relay Uncertainty Function** to resolve the credit assignment paradox across these spatial ruptures. Treating synthesized anchors as interventional intermediary states, these potentials propagate pragmatic and epistemic value through Bellman-style backups. Notably, we prove that uncertainty propagation across discontinuities necessitates a quadratic discount $\gamma^2$, establishing a formal epistemic horizon. Theoretically, MD approximates a variance-minimizing importance sampler that expands the manifold's spectral gap, reducing the hitting time to critical bottleneck states. Empirically, MD achieves a **1.67$\times$ average speedup** over DreamerV3 on DeepMind Control Suite, reaching **8.8$\times$** in sparse-reward tasks.

Benchmarks play a central role in reinforcement learning (RL) research, yet their computational constraints often shape what is studied. Despite the motivation of lifelong learning, most continual RL papers consider only 3–10 sequential tasks, as CPU-bound environments make longer sequences impractical. Meanwhile, continual learning in cooperative multi-agent settings remains largely unexplored. To address these gaps, we introduce **MEAL** (**M**ulti-agent **E**nvironments for **A**daptive **L**earning), the first benchmark for continual multi-agent RL. By leveraging JAX and GPU acceleration, MEAL enables training on sequences of 100 tasks on a single GPU in a few hours. We find that long task sequences reveal failure modes that do not appear at smaller scales.

Extensive research has highlighted the severe threats posed by backdoor attacks to deep reinforcement learning (DRL). However, prior studies primarily focus on vanilla scenarios, while plasticity interventions have emerged as indispensable built-in components of modern DRL agents. Despite their effectiveness in mitigating plasticity loss, the impact of these interventions on DRL backdoor vulnerabilities remains underexplored, and this lack of systematic investigation poses risks in practical DRL deployments. To bridge this gap, we empirically study 14,664 cases integrating representative interventions and attack scenarios. We find that only one intervention (i.e., *SAM*) exacerbates backdoor threats, while other interventions mitigate them. Pathological analysis identifies that the exacerbation is attributed to backdoor gradient amplification, while the mitigation stems from activation pathway disruption and representation space compression. From these findings, we derive two novel insights: (1) a conceptual framework *SCC* for robust backdoor injection that deconstructs the mechanistic interplay between interventions and backdoors in DRL, and (2) abnormal loss landscape sharpness as a key indicator for DRL backdoor detection.

Socially intelligent AI systems must reason across diverse human behavioral tasks and generalize to new social contexts. However, behavioral data is inherently heterogeneous, comprising diverse modalities and prediction targets that produce uneven training signals across samples, creating imbalanced learning dynamics that challenge existing AI models. To address this, we develop OmniSapiens-7B 2.0, a foundation model for social behavior processing that explicitly addresses learning from heterogeneous behavioral data. This is enabled through Heterogeneity-Aware Relative Policy Optimization, a new RL method that rebalances learning signals across samples by approximating each sample’s contribution to the policy update and using these estimates to drive geometrically centered, inertially smoothed advantage modulation for stable training. OmniSapiens-7B 2.0 achieves the best and most consistent performance across 10 behavioral tasks, while also attaining the best performance on all five held-out benchmarks, with gains of up to +12.02% and +9.37% respectively. Furthermore, it demonstrates more consistent and interpretable reasoning traces, supporting reliable real-world behavioral applications. Our model is available at https://github.com/MIT-MI/human_behavior_atlas.

Self-improvement training enables the large reasoning models (LRMs) to improve themselves by self-generating reasoning trajectories as training data without external supervision. However, we find that this method often falls short in complex reasoning tasks and even leads to model collapse. Through a series of preliminary analyses, we reveal two problems: (1) data imbalance, where most training samples are simple, but the challenging yet crucial samples are scarce; (2) overthinking, where many undesired samples with redundant reasoning steps are used for self-training. To this end, we propose HSIR, which effectively Harnesses Self-Improvement in large Reasoning models via two simple-yet-effective approaches. Specifically, HSIR introduces a verify-then-exit sampling strategy to mitigate data imbalance by efficiently collecting more accurate solutions for difficult queries, and designs an Intrinsic Diversity score to quantify overthinking and filter out the undesired solutions. We apply HSIR to various post-training paradigms, among which we further propose H-GRPO, an enhanced GRPO algorithm that leverages the intrinsic diversity as an external reward to encourage concise and diverse reasoning via reinforcement learning. Extensive results show that HSIR not only effectively enhances the reasoning performance, i.e., bringing up to +10.9% average performance gains, but also significantly improves the reasoning efficiency by reducing up to 42.4% relative inference overhead.

Programming-by-Example (PBE), as a typical few-shot inductive reasoning paradigm, aims to synthesize corresponding algorithms from a set of input-output examples. Although Large Language Models (LLMs) have demonstrated strong program synthesis potential, they still remain ineffective when handling complex PBE tasks. Specifically, LLMs often struggle to accurately grasp the underlying intent of examples, resulting in synthesized programs that either partially satisfy the examples or completely deviate from the target. To address these limitations, we introduce a process-supervised reinforcement learning method that provides fine-grained feedback during the synthesis process, improving the ability of LLMs to capture the intended behavior of provided examples. Firstly, we develop a reasoning tree construction method that is used to build a PBE process supervision dataset. Subsequently, we train a process reward model through preference learning to evaluate the effectiveness of reasoning steps. Finally, we introduce a curriculum learning strategy based on the difficulty of PBE tasks, using Proximal Policy Optimization (PPO) to optimize the model. Experimental results on representative PBE benchmarks show that our approach achieves an average pass rate of 56.61\%, significantly outperforming the state-of-the-art baseline by 8.73\%.

Existing LLMs-post-training techniques are broadly categorized into supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). Each paradigm presents a distinct trade-off: (1) SFT excels at mimicking demonstration data, but can lead to problematic generalization as a form of behavior cloning. (2) Conversely, RFT can significantly enhance a model's performance but is prone to learning unexpected behaviors, and its performance is sensitive to the initial policy. In this paper, we propose a unified view of these methods and introduce Prefix-RFT, a hybrid approach that synergizes learning from both demonstration and exploration. Using mathematical reasoning problems as a test bed, we empirically demonstrate that Prefix-RFT is simple yet effective. Not only does it surpass the performance of standalone SFT and RFT, but it also outperforms parallel mixed-policy RFT methods. Our analysis highlights the complementary nature of SFT and RFT, validating that Prefix-RFT effectively harmonizes them. Further ablation studies confirm the method's robustness to variations in the quality and quantity of demonstration data.

Group Relative Policy Optimization (GRPO), a prominent algorithm within the Reinforcement Learning from Verifiable Rewards (RLVR) framework, has achieved strong results in improving the reasoning capabilities of large language models (LLMs). However, GRPO is prone to advantage collapse, a failure mode where homogeneous rewards within a group (e.g., all correct or all incorrect answers) yield near-zero advantages and vanishing gradients. To address this, we introduce the Advantage Collapse Rate (ACR), the first diagnostic metric quantifying the proportion of training batches with ineffective gradients. Across models from 0.5B to 14B parameters on mathematical reasoning benchmarks, we show that ACR strongly predicts training stagnation and final performance. We then propose Adaptive Virtual Sample Policy Optimization (AVSPO), a lightweight extension of GRPO that injects virtual reward samples, guided by real-time ACR monitoring, to enable learning from homogeneous groups without additional model rollouts. AVSPO reduces advantage collapse by 58–63% relative to GRPO and yields consistent accuracy gains of 4–6 percentage points across all model scales, while maintaining generalization on the evaluated out-of-domain task. Code and datasets are available at https://github.com/hexixiang/Advantage-Collapse-Rate.

Training robust reasoning vision-language models (VLMs) in rare domains (such as geospatial) is fundamentally constrained by supervision scarcity. While raw geospatial imagery is abundant, the amount of task-direct supervision falls far behind that of common domains. In this work, we validate an important conclusion: indirect verifiable rewards, derived from seemingly unrelated metadata, are sufficient to induce sophisticated and generalizable geospatial reasoning across a wide range of downstream tasks (25+). We present Geo-R1 as one empirical instantiation of this paradigm. Rather than relying on limited task-specific annotations (i.e., direct rewards), Geo-R1 utilizes scalable, verifiable indirect proxy rewards based on cross-view alignment with metadata (geolocation information) to drive reinforcement learning at scale. Such indirect rewards successfully motivate the model to discover and internalize zero-shot geospatial reasoning across diverse tasks, achieving extraordinary zero-shot transfer on out-of-distribution benchmarks and even surpassing fully supervised specialists on certain benchmarks. These findings indicate that optimizing for indirect verifiable rewards may provide a scalable pathway to unlock generalized reasoning capabilities in rare domains with massive unlabeled data archives. Our code is available at: https://github.com/miniHuiHui/Geo-R1.

Alignment of large language models (LLMs) with human values has recently garnered significant attention, with prominent examples including the canonical yet costly Reinforcement Learning from Human Feedback (RLHF) and the simple Direct Preference Optimization (DPO). In this work, we demonstrate that both RLHF and DPO can be interpreted from the perspective of mutual information (MI) maximization, uncovering a profound connection to contrastive learning. Within this framework, both RLHF and DPO can be interpreted as methods that performing contrastive learning based on the positive and negative samples derived from base model, leveraging the Donsker–Varadhan (DV) lower bound on MI (equivalently, the MINE estimator). Such paradigm further illuminates why RLHF may not intrinsically incentivize reasoning capacities in LLMs beyond what is already present in the base model. Building on the perspective, we replace the DV/MINE bound with the Jensen–Shannon (JS) MI estimator and propose the Mutual Information Optimization (MIO). Comprehensive theoretical analysis and extensive empirical evaluations demonstrate that MIO mitigates the late-stage decline in chosen-likelihood observed in DPO, achieving competitive or superior performance across various challenging reasoning and mathematical benchmarks

Offline goal-conditioned RL (GCRL) learns goal-reaching policies from static datasets, but real-world environments are often partially observable, so the collected trajectories are only partly consistent with the Markov assumption while other segments remain history-dependent. History-aware sequence models such as Decision Transformer (DT) are a natural fit for long-term dependency modeling, yet pure attention is inefficient and brittle when handling local Markovian structure and long-range context simultaneously. Although recent hybrid architectures (e.g., LSDT) introduce local extractors, their fixed-window extraction cannot adapt the effective memory to varying dependency lengths, often truncating long-range context instead of compressing it. Moreover, under sparse rewards, return-to-go (RTG) becomes non-discriminative across sub-trajectories, offering little guidance for stitching goal-reaching behaviors from diverse demonstrations. To address these limitations, we propose QHyer (Q-conditioned Hybrid Attention-Mamba Transformer), which replaces RTG with a Normalizing Flows (NFs) parameterized goal-reaching Q-estimator used directly as conditioning tokens, and a gated Hybrid Attention-Mamba backbone whose selective state-space dynamics enable content-adaptive history compression while attention captures global goal-directed dependencies. Extensive experiments on OGBench and D4RL demonstrate that QHyer achieves state-of-the-art performance on both non-Markovian and Markovian datasets, validating its effectiveness for diverse scenarios.

Generative posterior sampling using diffusion models has emerged as a dominant paradigm for solving inverse problems in imaging, which usually consists of three main components: data consistency (DC) guidance, classifier-free guidance (CFG) and stochasticity. While prior arts have focused on how to develop each or all components, less attention has given to how to schedule them, leading to heuristically fixed or partially adjusted suboptimal schedules. In this work, we argue that the interactions among all three components in terms of scheduling are crucial for significantly improved performance in solving inverse problems in imaging. Our analysis shows that aggressive CFG early in sampling conflict with DC guidance, while stochasticity brings the trajectory back to higher-probability regions. Based on these findings, we propose Triadic Dynamics Aware Posterior Sampling (TriPS), which reformulates posterior sampling as a time-varying control problem and optimizes schedules following a triadic trend of decreasing DC and stochasticity scales alongside increasing CFG scale. TriPS achieves this through two strategies: template-based search over functional priors for reliable baseline schedules, and Group Relative Policy Optimization (GRPO)-based reinforcement learning for more flexible temporal curves. Experiments demonstrate TriPS outperforms state-of-the-art baselines in data fidelity and perceptual realism.

Domain adaptation typically follows a two-stage pipeline: Supervised Fine-Tuning (SFT) then Reinforcement Learning (RL). However, *does RL necessarily require a pre-SFT phase for domain adaptation?* SFT confines the model to an imitation distribution, limiting RL exploration, while the two-stage transition causes capability regression and extra engineering. We propose **One-stage Policy Optimization (OnePO)**, an SFT-free paradigm that adapts pretrained LLMs to target domains in a single RL stage. OnePO uses teacher outputs as transient guidance to overcome the slow convergence of pure RL, while avoiding two failures of naive teacher-output integration: gradient starvation for low-probability teacher tokens and distribution anchoring from persistent teacher signals. It introduces two mechanisms: (1) *Adaptive Objective Evolution*, reshaping the RL objective for rapid absorption of teacher-provided knowledge; and (2) *Teacher Retirement*, automatically discarding teacher outputs once the model surpasses them. On medical adaptation, OnePO achieves 67.2 on HealthBench with only 20K training samples, outperforming SFT+RL by +2.7 and pure RL by +7.4 points. Scaling the same recipe produces *HuatuoGPT-3*, an open-source medical LLM series whose 32B variant reaches 70.3 on HealthBench. Additional writing and legal-domain experiments show that OnePO extends beyond medicine. Models and code are available at https://github.com/FreedomIntelligence/HuatuoGPT-3.

Reinforcement learning with verifiable rewards (RLVR) trains a policy by verifying sampled completions and reinforcing higher-scoring outputs, but practical verifiers (e.g., incomplete unit tests or noisy judges) are prone to false positives and false negatives. We ask when such noise merely slows learning and when it reverses it. Modeling GRPO-style RLVR as a bandit over recurring \emph{reasoning modes}, we derive mean-field replicator-style (natural-selection) flow on the probability simplex. The dynamics decouples into within-correct-mode competition and a one-dimensional evolution for the mass on incorrect modes, whose drift is determined solely by Youden's index $J=\mathrm{TPR}-\mathrm{FPR}$. This yields a sharp phase transition: when $J>0$, the incorrect mass is driven toward extinction (learning); when $J=0$, the process is neutral; and when $J<0$, incorrect modes amplify until they dominate (anti-learning and collapse). In the learning regime $J>0$, noise primarily rescales convergence time (``rate, not fate''). Experiments on verifiable programming tasks under synthetic noise reproduce the predicted $J=0$ boundary. Beyond noise, the framework offers a general lens for analyzing RLVR stability, convergence, and algorithmic interventions.

We propose a data augmentation method for offline reinforcement learning, motivated by active positioning problems. Particularly, our approach enables the training of off-policy models from a limited number of suboptimal trajectories. We introduce a trajectory-based augmentation technique that exploits task structure and the geometric relationship between rewards, value functions, and mathematical properties of logging policies. During data collection, our augmentation supports suboptimal logging policies, leading to higher data quality and improved offline reinforcement learning performance. We provide theoretical justification for these strategies and validate them empirically across positioning tasks of varying dimensionality and under partial observability.