NeurIPS 2025 Papers

The continuous nature of belief states in POMDPs presents significant computational challenges in learning the optimal policy. In this paper, we consider an approach that solves a Partially Observable Reinforcement Learning (PORL) problem by approximating the corresponding POMDP model into a finite-state Markov Decision Process (MDP) (called Superstate MDP). We first derive theoretical guarantees that improve upon prior work that relate the optimal value function of the transformed Superstate MDP to the optimal value function of the original POMDP. Next, we propose a policy-based learning approach with linear function approximation to learn the optimal policy for the Superstate MDP. Consequently, our approach shows that a POMDP can be approximately solved using TD-learning followed by Policy Optimization by treating it as an MDP, where the MDP state corresponds to a finite history. We show that the approximation error decreases exponentially with the length of this history. To the best of our knowledge, our finite-time bounds are the first to explicitly quantify the error introduced when applying standard TD learning to a setting where the true dynamics are not Markovian.

Exploration is crucial in Reinforcement Learning (RL) as it enables the agent to understand the environment for better decision-making. Existing exploration methods fall into two paradigms: active exploration, which injects stochasticity into the policy but struggles in high-dimensional environments, and passive exploration, which manages the replay buffer to prioritize under-explored regions but lacks sample diversity. To address the limitation in passive exploration, we propose Modelic Generative Exploration (MoGE), which augments exploration through the generation of under-explored critical states and synthesis of dynamics-consistent experiences. MoGE consists of two components: (1) a diffusion generator for critical states under the guidance of entropy and TD error, and (2) a one-step imagination world model for constructing critical transitions for agent learning. Our method is simple to implement and seamlessly integrates with mainstream off-policy RL algorithms without structural modifications. Experiments on OpenAI Gym and DeepMind Control Suite demonstrate that MoGE, as an exploration augmentation, significantly enhances efficiency and performance in complex tasks.

While large language models (LLMs) demonstrate strong reasoning capabilities utilizing reinforcement learning (RL) with verifiable reward, whether large vision-language models (VLMs) can directly inherit such capabilities through similar post-training strategies remains underexplored. In this work, we conduct a systematic compositional probing study to evaluate whether current VLMs trained with RL or other post-training strategies can compose capabilities across modalities or tasks under out-of-distribution conditions. We design a suite of diagnostic tasks that train models on unimodal tasks or isolated reasoning skills, and evaluate them on multimodal, compositional variants requiring skill integration. Through comparisons between supervised fine-tuning (SFT) and RL-trained models, we identify three key findings: (1) RL-trained models consistently outperform SFT on compositional generalization, demonstrating better integration of learned skills; (2) although VLMs achieve strong performance on individual tasks, they struggle to generalize compositionally under cross-modal and cross-task scenarios, revealing a significant gap in current training strategies; (3) enforcing models to explicitly describe visual content before reasoning (e.g., caption-before-thinking), along with rewarding progressive vision-to-text grounding, yields notable gains. It highlights two essential ingredients for improving compositionality in VLMs: visual-to-text alignment and accurate visual grounding. Our findings shed light on the current limitations of RL-based reasoning VLM training and provide actionable insights toward building models that reason compositionally across modalities and tasks.

Reinforcement learning (RL) systems, while achieving remarkable success across various domains, are vulnerable to adversarial attacks. This is especially a concern in vision-based environments where minor manipulations of high-dimensional image inputs can easily mislead the agent's behavior. To this end, various defenses have been proposed recently, with state-of-the-art approaches achieving robust performance even under large state perturbations. However, after closer investigation, we found that the effectiveness of the current defenses is due to a fundamental weakness of the existing $l_p$ norm-constrained attacks, which can barely alter the semantics of image input even under a relatively large perturbation budget. In this work, we propose SHIFT, a novel policy-agnostic diffusion-based state perturbation attack to go beyond this limitation. Our attack is able to generate perturbed states that are semantically different from the true states while remaining realistic and history-aligned to avoid detection. Evaluations show that our attack effectively breaks existing defenses, including the most sophisticated ones, significantly outperforming existing attacks while being more perceptually stealthy. The results highlight the vulnerability of RL agents to semantics-aware adversarial perturbations, indicating the importance of developing more robust policies.

Value decomposition has long been a fundamental technique in multi-agent reinforcement learning and dynamic programming. Specifically, the value function of a global state $(s_1,s_2,\ldots,s_N)$ is often approximated as the sum of local functions: $V(s_1,s_2,\ldots,s_N)\approx\sum_{i=1}^N V_i(s_i)$. This approach has found various applications in modern RL systems. However, the theoretical justification for why this decomposition works so effectively remains underexplored. In this paper, we uncover the underlying mathematical structure that enables value decomposition. We demonstrate that a Markov decision process (MDP) permits value decomposition *if and only if* its transition matrix is not "entangled"—a concept analogous to quantum entanglement in quantum physics. Drawing inspiration from how physicists measure quantum entanglement, we introduce how to measure the "Markov entanglement" and show that this measure can be used to bound the decomposition error in general multi-agent MDPs. Using the concept of Markov entanglement, we proved that a widely-used class of policies, the index policy, is weakly-entangled and enjoys a sublinear $\mathcal O(\sqrt{N})$ scale of decomposition error for $N$-agent systems. Finally, we show Markov entanglement can be efficiently estimated, guiding practitioners on the feasibility of value decomposition.

Large language models (LLMs) have significantly advanced in reasoning tasks through reinforcement learning (RL) optimization, achieving impressive capabilities across various challenging benchmarks. However, our empirical analysis reveals a critical drawback: reasoning-oriented RL fine-tuning significantly increases the prevalence of hallucinations. We theoretically analyze the RL training dynamics, identifying high-variance gradient, entropy-induced randomness, and susceptibility to spurious local optima as key factors leading to hallucinations. To address this drawback, we propose Factuality-aware Step-wise Policy Optimization (FSPO), an innovative RL fine-tuning algorithm incorporating explicit factuality verification at each reasoning step. FSPO leverages automated verification against given evidence to dynamically adjust token-level advantage values, incentivizing factual correctness throughout the reasoning process. Experiments across mathematical reasoning and hallucination benchmarks using Qwen2.5 and Llama models demonstrate that FSPO effectively reduces hallucinations while enhancing reasoning accuracy, substantially improving both reliability and performance.

Protein fitness optimization involves finding a protein sequence that maximizes desired quantitative properties in a combinatorially large design space of possible sequences. Recent advances in steering protein generative models (e.g., diffusion models and language models) with labeled data offer a promising approach. However, most previous studies have optimized surrogate rewards and/or utilized large amounts of labeled data for steering, making it unclear how well existing methods perform and compare to each other in real-world optimization campaigns where fitness is measured through low-throughput wet-lab assays. In this study, we explore fitness optimization using small amounts (hundreds) of labeled sequence-fitness pairs and comprehensively evaluate strategies such as classifier guidance and posterior sampling for guiding generation from different discrete diffusion models of protein sequences. We also demonstrate how guidance can be integrated into adaptive sequence selection akin to Thompson sampling in Bayesian optimization, showing that plug-and-play guidance strategies offer advantages over alternatives such as reinforcement learning with protein language models. Overall, we provide practical insights into how to effectively steer modern generative models for next-generation protein fitness optimization.

Graphical User Interface (GUI) agents have made substantial strides in understanding and executing user instructions across diverse platforms. Yet, grounding these instructions to precise interface elements remains challenging—especially in complex, high-resolution, professional environments. Traditional supervised fine-tuning (SFT) methods often require large volumes of diverse data and exhibit weak generalization. To overcome these limitations, we introduce a reinforcement learning (RL)-based framework that incorporates three core strategies: (1) seed data curation to ensure high-quality training samples, (2) a dense policy gradient that provides continuous feedback based on prediction accuracy, and (3) a self-evolutionary reinforcement finetuning mechanism that iteratively refines the model using attention maps. With only 3k training samples, our 7B-parameter model achieves state-of-the-art results among similarly sized models on three grounding benchmarks. Notably, it attains 47.3\% accuracy on the ScreenSpot-Pro dataset—outperforming much larger models, such as UI-TARS-72B, by a margin of 24.2\%. These findings underscore the effectiveness of RL-based approaches in enhancing GUI agent performance, particularly in high-resolution, complex environments.

Paucity of medical data severely limits the generalizability of diagnostic ML models, as the full spectrum of disease variability can not be represented by a small clinical dataset. To address this, diffusion models (DMs) have been considered as a promising avenue for synthetic image generation and augmentation. However, they frequently produce _medically inaccurate_ images, deteriorating the model performance. Expert domain knowledge is critical for synthesizing images that correctly encode clinical information, especially when data is scarce and quality outweighs quantity. Existing approaches for incorporating human feedback, such as reinforcement learning (RL) and Direct Preference Optimization (DPO), rely on robust reward functions or demand labor-intensive expert evaluations. Recent progress in Multimodal Large Language Models (MLLMs) reveals their strong visual reasoning capabilities, making them adept candidates as evaluators. In this work, we propose a novel framework, coined MAGIC (**M**edically **A**ccurate **G**eneration of **I**mages through AI-Expert **C**ollaboration), that synthesizes clinically accurate skin disease images for data augmentation. Our method creatively translates expert-defined criteria into actionable feedback for image synthesis of DMs, significantly improving clinical accuracy while reducing the direct human workload. Experiments demonstrate that our method greatly improves the clinical quality of synthesized skin disease images, with outputs aligning with dermatologist assessments. Additionally, augmenting training data with these synthesized images improves diagnostic accuracy by +9.02% on a challenging 20-condition skin disease classification task, and by +13.89% in the few-shot setting. Beyond image synthesis, MAGIC illustrates a task-centric alignment paradigm: instead of adapting MLLMs to niche medical tasks, it adapts tasks to the evaluative strengths of general-purpose MLLMs by decomposing domain knowledge into attribute-level checklists. This design offers a scalable and reliable path for leveraging foundation models in specialized domains.

Neural scaling laws are observed in a range of domains, to date with no universal understanding of why they occur. Recent theories suggest that loss power laws arise from Zipf's law, a power law observed in domains like natural language. One theory suggests that language scaling laws emerge when Zipf-distributed task quanta are learned in descending order of frequency. In this paper we examine power-law scaling in AlphaZero, a reinforcement learning algorithm, using a model of language-model scaling. We find that game states in training and inference data scale with Zipf's law, which is known to arise from the tree structure of the environment, and examine the correlation between scaling-law and Zipf's-law exponents. In agreement with the quanta scaling model, we find that agents optimize state loss in descending order of frequency, even though this order scales inversely with modelling complexity. We also find that inverse scaling, the failure of models to improve with size, is correlated with unusual Zipf curves where end-game states are among the most frequent states. We show evidence that larger models shift their focus to these less-important states, sacrificing their understanding of important early-game states.

Policy optimization (PO) is a cornerstone of modern reinforcement learning (RL), with diverse applications spanning robotics, healthcare, and large language model training. The increasing deployment of PO in sensitive domains, however, raises significant privacy concerns. In this paper, we initiate a theoretical study of differentially private policy optimization, focusing explicitly on its sample complexity. We first formalize an appropriate definition of differential privacy (DP) tailored to PO, addressing the inherent challenges arising from on-policy learning dynamics and the subtlety involved in defining the unit of privacy. We then systematically analyze the sample complexity of widely-used PO algorithms, including policy gradient (PG), natural policy gradient (NPG) and more, under DP constraints and various settings, via a unified framework. Our theoretical results demonstrate that privacy costs can often manifest as lower-order terms in the sample complexity, while also highlighting subtle yet important observations in private PO settings. These offer valuable practical insights for privacy-preserving PO algorithms.

We introduce ExRec, a general framework for personalized exercise recommendation with semantically-grounded knowledge tracing. Our method builds on the observation that existing exercise recommendation approaches simulate student performance via knowledge tracing (KT) but they often overlook two key aspects: (a) the semantic content of questions and (b) the sequential, structured progression of student learning. To address this, our ExRec presents an end-to-end pipeline, from annotating the KCs of questions and learning their semantic representations to training KT models and optimizing several reinforcement learning (RL) methods. Moreover, we improve standard Q-learning-based continuous RL methods via a tailored model-based value estimation (MVE) approach that directly leverages the components of KT model in estimating cumulative knowledge improvement. We validate the effectiveness of our ExRec using various RL methods across four real-world tasks with different educational goals in online math learning. We further show that ExRec generalizes robustly to new, unseen questions and that it produces interpretable student learning trajectories. Together, our findings highlight the promise of KT-guided RL for effective personalization in education.

Transformers have demonstrated exceptional performance across a wide range of domains. While their ability to perform reinforcement learning in-context has been established both theoretically and empirically, their behavior in non-stationary environments remains less understood. In this study, we address this gap by showing that transformers can achieve nearly optimal dynamic regret bounds in non-stationary settings. We prove that transformers are capable of approximating strategies used to handle non-stationary environment, and can learn the approximator in the in-context learning setup. Our experiments further show that transformers can match or even outperform existing expert algorithms in such environments.

In financial applications, reinforcement learning (RL) agents are commonly trained on historical data, where their actions do not influence prices. However, during deployment, these agents trade in live markets where their own transactions can shift asset prices, a phenomenon known as market impact. This mismatch between training and deployment environments can significantly degrade performance. Traditional robust RL approaches address this model misspecification by optimizing the worst-case performance over a set of uncertainties, but typically rely on symmetric structures that fail to capture the directional nature of market impact. To address this issue, we develop a novel class of elliptic uncertainty sets. We establish both implicit and explicit closed-form solutions for the worst-case uncertainty under these sets, enabling efficient and tractable robust policy evaluation. Experiments on single-asset and multi-asset trading tasks demonstrate that our method achieves superior Sharpe ratio and remains robust under increasing trade volumes, offering a more faithful and scalable approach to RL in financial markets.

We show that reinforcement learning with verifiable reward using one training example (1-shot RLVR) is effective in incentivizing the math reasoning capabilities of large language models (LLMs). Applying RLVR to the base model Qwen2.5-Math-1.5B, we identify a single example that elevates model performance on MATH500 from 36.0\% to 73.6\% (8.6\% improvement beyond format correction), and improves the average performance across six common mathematical reasoning benchmarks from 17.6\% to 35.7\% (7.0\% non-format gain). This result matches the performance obtained using the 1.2k DeepScaleR subset (MATH500: 73.6\%, average: 35.9\%), which contains the aforementioned example. Furthermore, RLVR with only two examples even slightly exceeds these results (MATH500: 74.8\%, average: 36.6\%). Similar substantial improvements are observed across various models (Qwen2.5-Math-7B, Llama3.2-3B-Instruct, DeepSeek-R1-Distill-Qwen-1.5B), RL algorithms (GRPO and PPO), and different math examples. In addition, we identify some interesting phenomena during 1-shot RLVR, including cross-category generalization, increased frequency of self-reflection, and sustained test performance improvement even after the training accuracy has saturated, a phenomenon we term \textit{post-saturation generalization}. Moreover, we verify that the effectiveness of 1-shot RLVR primarily arises from the policy gradient loss, distinguishing it from the "grokking" phenomenon. We also show the critical role of promoting exploration (e.g., by incorporating entropy loss with an appropriate coefficient) in 1-shot RLVR training. We also further discuss related observations about format correction, label robustness and prompt modification. These findings can inspire future work on RLVR efficiency and encourage a re-examination of recent progress and the underlying mechanisms in RLVR. Our code, models, and data are open source at https://github.com/ypwang61/One-Shot-RLVR.

Thompson Sampling (TS) has emerged as a powerful algorithm for sequential decision-making, with strong empirical success and theoretical guarantees. However, it has been shown that its behavior under stringent safety and robustness criteria --- such as safety of cumulative regret distribution and robustness to model mis-specification --- can sometimes perform poorly. In this work, we try to address these aspects through the lens of adaptive variance inflation for Gaussian Thompson Sampling. Our one-line change introduces a time- and arm-dependent inflation factor into the sampling variance, and yields several compelling benefits. The resulting policy achieves provably worst-case optimal expected regret and worst-case optimal fast-decaying regret tail bounds, even in the presence of heavy-tailed (sub-exponential) noise or mis-specified environments. The policy is also robust to mis-specified noise variances. Beyond cumulative regret, we further demonstrate that our method ensures strong post-experiment guarantees: simple regret and estimation error per arm exhibit fast-decaying tail probabilities, contributing to more reliable and robust downstream decisions. Finally, we extend our policy to incorporate settings with unknown arm-specific variances and empirically validate the consistent performance of our approach across a range of environments.

Visual reinforcement learning has shown promise in various real-world applications. However, deploying policies in complex real-world environments with visual perturbations remains a significant challenge. We notice that humans tend to filter information at the object level prior to decision-making, facilitating efficient skill transfer across different contexts. Inspired by this, we introduce Focus-Then-Reuse (FTR), a method utilizing a novel object selection mechanism to focus on task-relevant objects, and directly reuse the simulation-trained policy on them. The training of the object selection mechanism integrates prior knowledge from a vision-language model and feedback from the environment. Experimental results on challenging tasks based on DeepMind Control Suite and Franka Emika Robotics demonstrate that FTR enables rapid adaptation in visual perturbation environments and achieves state-of-the-art performance. The source code is available at https://github.com/LAMDA-RL/FTR.

Planning with pretrained diffusion models has emerged as a promising approach for solving test-time guided control problems. Standard gradient guidance typically performs optimally under convex, differentiable reward landscapes. However, it shows substantially reduced effectiveness in real-world scenarios with non-convex objectives, non-differentiable constraints, and multi-reward structures. Furthermore, recent supervised planning approaches require task-specific training or value estimators, which limits test-time flexibility and zero-shot generalization. We propose a Tree-guided Diffusion Planner (TDP), a zero-shot test-time planning framework that balances exploration and exploitation through structured trajectory generation. We frame test-time planning as a tree search problem using a bi-level sampling process: (1) diverse parent trajectories are produced via training-free particle guidance to encourage broad exploration, and (2) sub-trajectories are refined through fast conditional denoising guided by task objectives. TDP addresses the limitations of gradient guidance by exploring diverse trajectory regions and harnessing gradient information across this expanded solution space using only pretrained models and test-time reward signals. We evaluate TDP on three diverse tasks: maze gold-picking, robot arm block manipulation, and AntMaze multi-goal exploration. TDP consistently outperforms state-of-the-art approaches on all tasks. The project page can be found at: https://tree-diffusion-planner.github.io

This paper studies reinforcement learning from human feedback (RLHF) for aligning large language models with human preferences. While RLHF has demonstrated promising results, many algorithms are highly sensitive to misspecifications in the underlying preference model (e.g., the Bradley-Terry model), the reference policy, or the reward function, resulting in undesirable fine-tuning. To address model misspecification, we propose a doubly robust preference optimization algorithm that remains consistent when either the preference model or the reference policy is correctly specified (without requiring both). Our proposal demonstrates superior and more robust performance than state-of-the-art algorithms, both in theory and in practice. The code is available at https://github.com/DRPO4LLM/DRPO4LLM