强化学习

Reinforcement learning fine-tuning (RLFT) is a dominant paradigm for improving pretrained policies for downstream tasks. These pretrained policies, trained on large datasets, produce generations with a broad range of promising but unrefined behaviors. Often, a critical failure mode of RLFT arises when policies lose this diversity and collapse into a handful of easily exploitable outputs. This convergence hinders exploration, which is essential for expanding the capabilities of the pretrained policy and for amplifying the benefits of test-time compute scaling. To address this, we introduce an objective for policy gradient methods that explicitly enforces the exploration and refinement of diverse generations, which we call a polychromic objective. We then show how proximal policy optimization (PPO) can be adapted to optimize this objective. Our method (1) employs vine sampling to collect on-policy rollouts and (2) modifies the advantage function to reflect the advantage under our new objective. Experiments on BabyAI, Minigrid, and Algorithmic Creativity show that our method improves success rates by reliably solving a larger set of environment configurations and generalizes better under large perturbations. Moreover, when given multiple attempts in pass@$n$ experiments, the policy achieves substantially higher coverage, demonstrating its ability to maintain and exploit a diverse repertoire of strategies.

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

Models like OpenAI-o3 pioneer visual grounded reasoning by dynamically ref- erencing visual regions, just like human “thinking with images”. However, no benchmark exists to evaluate these capabilities holistically. To bridge this gap, we propose TreeBench (Traceable Evidence Evaluation Benchmark), a diagnostic benchmark built on three principles: (1) focused visual perception of subtle targets in complex scenes, (2) traceable evidence via bounding box evaluation, and (3) second-order reasoning to test object interactions and spatial hierarchies beyond simple object localization. Prioritizing images with dense objects, we initially sample 1K high-quality images from SA-1B, and incorporate eight LMM experts to manually annotate questions, candidate options, and answers for each image. After three stages of quality control, TreeBench consists of 405 challenging vi- sual question-answering pairs, even the most advanced models struggle with this benchmark, where none of them reach 60% accuracy, e.g., OpenAI-o3 scores only 54.87. Furthermore, we introduce TreeVGR (Traceable Evidence Enhanced Visual Grounded Reasoning), a training paradigm to supervise localization and reasoning jointly with reinforcement learning, enabling accurate localizations and explainable reasoning pathways. Initialized from Qwen2.5-VL-7B, it improves V* Bench (+16.8), MME-RealWorld (+12.6), and TreeBench (+13.4), proving traceability is key to advancing vision-grounded reasoning. The code and data will be released.

We propose QeRL, a Quantization-enhanced Reinforcement Learning framework for large language models (LLMs). While RL is essential for LLMs' reasoning capabilities, it is resource-intensive, requiring substantial GPU memory and long rollout duration. QeRL addresses these issues by combining NVFP4 quantization with Low-Rank Adaptation (LoRA), accelerating rollout phase of RL while reducing memory overhead. Beyond efficiency, our findings show that quantization noise increases policy entropy, enhancing exploration, and enabling the discovery of better strategies during RL. To further optimize exploration, QeRL introduces an Adaptive Quantization Noise (AQN) mechanism, which dynamically adjusts noise throughout training. Experiments demonstrate that QeRL delivers around a 1.2×–1.5× speedup compared to BF16 LoRA in end-to-end RL training while drastically reducing memory usage, and a 1.5×–2.0× speedup compared to QLoRA. Moreover, this is the first framework to enable RL training of a 32B LLM on a single H100 80GB GPU, while delivering overall speedups for RL training. It also achieves faster reward growth and higher final accuracy than 16-bit LoRA and QLoRA, while matching the performance of full-parameter fine-tuning on mathematical benchmarks such as GSM8K (90.8%) and MATH 500 (77.4%) in the 7B model. These results establish QeRL as an efficient and effective framework for RL training in LLMs.

Given a training dataset, the goal of dataset distillation is to derive a synthetic dataset such that models trained on the latter perform as well as those trained on the training dataset. In this work, we develop and analyze an efficient dataset distillation algorithm for supervised learning, specifically regression in $\mathbb{R}^d$, based on matching the losses on the training and synthetic datasets with respect to a fixed set of randomly sampled regressors without any model training. Our first key contribution is a novel performance guarantee proving that our algorithm needs only $\tilde{O}(d^2)$ sampled regressors to derive a synthetic dataset on which the MSE loss of any bounded linear model is approximately the same as its MSE loss on the given training data. In particular, the model optimized on the synthetic data has close to minimum loss on the training data, thus performing nearly as well as the model optimized on the latter. Complementing this, we also prove a matching lower bound of $\Omega(d^2)$ for the number of sampled regressors showing the tightness of our analysis. Our second contribution is to extend our algorithm to offline RL dataset distillation by matching the Bellman loss, unlike previous works which used a behavioral cloning objective. This is the first such method which leverages both, the rewards and the next state information, available in offline RL datasets, without any policy model optimization. We show similar guarantees: our algorithm generates a synthetic dataset whose Bellman loss with respect to any linear action-value predictor is close to the latter’s Bellman loss on the offline RL training dataset. Therefore, a policy associated with an action-value predictor optimized on the synthetic dataset performs nearly as well as that derived from the one optimized on the training data. We conduct extensive experiments to validate our theoretical guarantees and observe performance gains on real-world RL environments with offline training datasets and supervised regression datasets.

In vision–language models (VLMs), misalignment between textual descriptions and visual coordinates often induces hallucinations. This issue becomes particularly severe in dense prediction tasks such as spatial–temporal video grounding (STVG). Prior approaches typically focus on enhancing visual–textual alignment or attaching auxiliary decoders. However, these strategies inevitably introduce additional trainable modules, leading to significant annotation costs and computational overhead. In this work, we propose a novel visual prompting paradigm that avoids the difficult problem of aligning coordinates across modalities. Specifically, we reformulate per-frame coordinate prediction as a compact instance-level identification problem by assigning each object a unique, temporally consistent ID. These IDs are embedded into the video as visual prompts, providing explicit and interpretable inputs to the VLMs. Furthermore, we introduce STVG-R1, the first reinforcement learning framework for STVG, which employs a task-driven reward to jointly optimize temporal accuracy, spatial consistency, and structural format regularization. Extensive experiments on six benchmarks demonstrate the effectiveness of our approach. STVG-R1 surpasses the baseline Qwen2.5-VL-7B by a remarkable margin of 20.9% on m_IoU on the HCSTVG-v2 benchmark, establishing a new state of the art (SOTA). Surprisingly, STVG-R1 also exhibits strong zero-shot generalization to multi-object referring video object segmentation task, achieving a SOTA 47.3% J&F on MeViS.

Group relative policy optimization (GRPO) has demonstrated significant potential in improving the reasoning capabilities of large language models (LLMs) via reinforcement learning. However, its practical deployment is impeded by an excessively slow training process, primarily attributed to the computationally intensive autoregressive generation of multiple responses per query, which makes the generation phase the primary performance bottleneck. Although speculative decoding presents a promising direction for acceleration, its direct application in GRPO achieves limited speedup under high-concurrency training conditions. To overcome this limitation, we propose a concurrency-aware speculative decoding framework that dynamically adjusts the drafting and verification strategy according to real-time concurrency levels, thereby maximizing the acceleration of the generation process. Furthermore, to address performance degradation arising from distributional drift between the evolving target model and the fixed draft model during training, we introduce an online draft learning mechanism that enables the draft model to continuously adapt using feedback signals from the target model. Experimental results across multiple mathematical reasoning datasets and models demonstrate that the proposed method achieves end-to-end speedups of 2.35x to 2.72x, significantly surpassing baseline approaches in efficiency. The code is available at https://github.com/yedaotian9/FastGRPO.

Graphical user interface visual grounding (GUI-VG)—a core capability for GUI agents—has primarily relied on supervised fine-tuning (SFT) of multimodal large language models (MLLMs), demanding extensive data curation and significant training costs. However, as MLLMs continue to advance and even cover GUI domains during pretraining, the necessity of exhaustive SFT post-training becomes increasingly questionable. Meanwhile, the recent successes of rule-based reinforcement fine-tuning (RFT) suggest a more efficient alternative. However, despite its promise, the optimal manner of RFT for GUI-VG remains unexplored. To bridge this gap, we introduce GuirlVG, a reinforcement learning–based GUI-VG method built on a systematic empirical study and a novel stabilization technique. Preliminarily, we find that naive application of RFT underperforms the SFT baseline, motivating a deeper exploration of RFT. First, we decompose RFT into its core components and analyze the optimal formulation of each. Second, as part of this exploration, we propose a novel Adversarial KL Factor that dynamically stabilizes training to mitigate reward over-optimization. Third, we further explore the training configurations of RFT to enhance the effectiveness. Extensive experiments show that GuirlVG, with only 5.2K training samples, outperforms SFT methods trained on over 10M samples, achieving a +7.7% improvement on ScreenSpot, a +17.2% improvement on ScreenSpotPro and 91.9% accuracy on ScreenSpotV2.

Large language model based multi-agent systems (MAS) have unlocked significant advancements in tackling complex problems, but their increasing capability introduces a structural fragility that makes them difficult to debug. A key obstacle to improving their reliability is the severe scarcity of large-scale, diverse datasets for error attribution, as existing resources rely on costly and unscalable manual annotation. To address this bottleneck, we introduce *Aegis*, a novel framework for **A**utomated **e**rror **g**eneration and attr**i**bution for multi-agent **s**ystems. *Aegis* constructs a large dataset of **9,533** trajectories with annotated faulty agents and error modes, covering diverse MAS architectures and task domains. This is achieved using a LLM-based manipulator that can adaptively inject context-aware errors into successful execution trajectories. Leveraging fine-grained labels and the structured arrangement of positive-negative sample pairs, *Aegis* supports three different learning paradigms: Supervised Fine-Tuning, Reinforcement Learning, and Contrastive Learning. We develop learning methods for each paradigm. Comprehensive experiments show that trained models consistently achieve substantial improvements in error attribution. Notably, several of our fine-tuned LLMs demonstrate performance competitive with or superior to proprietary models an order of magnitude larger, validating our automated data generation framework as a crucial resource for developing more robust and interpretable multi-agent systems.

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

Computing worst-case robust strategies in pursuit-evasion games (PEGs) is time-consuming, especially when real-world factors like partial observability are considered. While important for general security purposes, real-time applicable pursuit strategies for graph-based PEGs are currently missing when the pursuers only have imperfect information about the evader's position. Although state-of-the-art reinforcement learning (RL) methods like Equilibrium Policy Generalization (EPG) and Grasper provide guidelines for learning graph neural network (GNN) policies robust to different game dynamics, they are restricted to the scenario of perfect information and do not take into account the possible case where the evader can predict the pursuers' actions. This paper introduces the first approach to worst-case robust real-time pursuit strategies (R2PS) under partial observability. We first prove that a traditional dynamic programming (DP) algorithm for solving Markov PEGs maintains optimality under the asynchronous moves by the evader. Then, we propose a belief preservation mechanism about the evader's possible positions, extending the DP pursuit strategies to a partially observable setting. Finally, we embed the belief preservation into the state-of-the-art EPG framework to finish our R2PS learning scheme, which leads to a real-time pursuer policy through cross-graph reinforcement learning against the asynchronous-move DP evasion strategies. After reinforcement learning, our policy achieves robust zero-shot generalization to unseen real-world graph structures and consistently outperforms the policy directly trained on the test graphs by the existing game RL approach.

As humans increasingly share environments with diverse agents powered by RL, LLMs, and beyond, the ability to explain agent policies in natural language is vital for reliable coexistence. We introduce a general-purpose framework that trains explanation-generating LLMs via reinforcement learning from AI feedback, with distributional rewards generated by generative continuous normalizing flows (CNFs). CNFs capture the pluralistic and probabilistic nature of human judgments about explanations. Moreover, under mild assumptions, CNFs provably bound deviations from true human reward distributions when trained on noisy proxy rewards from LLMs. We design a specialized CNF architecture that selectively attends to linguistic cues in decision context and explanations when generating rewards. Human and LLM evaluators find that our method delivers explanations that enable more accurate predictions of true agent decisions, exhibit greater logical soundness and actionability, and impose lower cognitive load than explanations trained with proxy LLM rewards or state-of-the-art RLHF and RLAIF baselines.

Although artificial intelligence (AI) has become deeply integrated into various stages of the research workflow and achieved remarkable advancements, academic rebuttal remains a significant and underexplored challenge. This is because rebuttal is a complex process of strategic communication under severe information asymmetry rather than a simple technical debate. Consequently, current approaches struggle as they largely imitate surface-level linguistics, missing the essential element of perspective-taking required for effective persuasion. In this paper, we introduce RebuttalAgent, the first framework to ground academic rebuttal in Theory of Mind (ToM), operationalized through a ToM-Strategy-Response (TSR) framework that models reviewer mental state, formulates persuasion strategy, and generates evidence-based response. To train our agent, we construct RebuttalBench, a large-scale dataset synthesized via a novel critique-and-refine approach. Our training process consists of two stages, beginning with a supervised fine-tuning phase to equip the agent with ToM-based analysis and strategic planning capabilities, followed by a reinforcement learning phase leveraging the self-reward mechanism for scalable self-improvement. For reliable and efficient automated evaluation, we further develop Rebuttal-RM, a specialized evaluator trained on over 100K samples of multi-source rebuttal data, which achieves scoring consistency with human preferences surpassing powerful judge GPT-4.1. Extensive experiments show RebuttalAgent significantly outperforms the base model by an average of 18.3% on automated metrics, while also outperforming advanced proprietary models across both automated and human evaluations.

Process Reward Models (PRMs) have recently emerged as a powerful framework for enhancing the reasoning capabilities of large reasoning models (LRMs), particularly in the context of test-time scaling (TTS). However, their potential for supervising LRMs on tabular reasoning domains remains underexplored. Through detailed empirical analyses, we identify that existing PRMs, though widely adopted for supervising text-only reasoning steps, struggle with table-specific operations such as sub-table retrieval and schema interaction, leading to critical performance bottlenecks. To address this limitation, we propose TaTToo, a novel table-grounded PRM framework that (i) reasons explicitly over tabular reasoning steps and (ii) integrates tool-based verification to provide precise reward supervision. Concretely, we first design a scalable data curation pipeline that constructs over 60k high-quality step-level annotations by integrating table verification rationales with tool-based executions. Building on the collected data, we train TaTToo with a dual-stage paradigm: cold-start supervised fine-tuning to capture tool-use reasoning patterns, followed by reinforcement learning with tool-grounded reward shaping to align our model with table-based verification. We provide a comprehensive evaluation of the policy improvement induced by our newly designed PRM. Across 5 challenging tabular reasoning benchmarks covering numerical reasoning, fact-checking, and data analysis, TaTToo improves downstream policy LRMs by 30.9\% at inference, surpasses strong PRM baselines such as Qwen-2.5-Math-PRM-72B with only 8B parameters, and demonstrates strong generalizability across diverse TTS strategies.

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

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

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

The automated mining of predictive signals, or alphas, is a central challenge in quantitative finance. While Reinforcement Learning (RL) has emerged as a promising paradigm for generating formulaic alphas, existing frameworks are fundamentally hampered by a triad of interconnected issues. First, they suffer from reward sparsity, where meaningful feedback is only available upon the completion of a full formula, leading to inefficient and unstable exploration. Second, they rely on semantically inadequate sequential representations of mathematical expressions, failing to capture the structure that determine an alpha's behavior. Third, the standard RL objective of maximizing expected returns inherently drives policies towards a single optimal mode, directly contradicting the practical need for a diverse portfolio of non-correlated alphas. To overcome these challenges, we introduce **AlphaSAGE** (**S**tructure-**A**ware Alpha Mining via **G**enerative Flow Networks for Robust **E**xploration), a novel framework is built upon three cornerstone innovations: (1) a structure-aware encoder based on Relational Graph Convolutional Network (RGCN); (2) a new framework with Generative Flow Networks (GFlowNets); and (3) a dense, multi-faceted reward structure. Empirical results demonstrate that AlphaSAGE outperforms existing baselines in mining a more diverse, novel, and highly predictive portfolio of alphas, thereby proposing a new paradigm for automated alpha mining. Our code is available at https://anonymous.4open.science/r/AlphaSAGE-3BA9.

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

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