Jiajun Chai

Zhexin Hu, Li Wang, Xiaohan Wang et al.

Adaptive context compression is vital for scaling Large Language Models (LLMs) to complex, multi-turn agent tasks. However, rule-based compression methods may discard task-critical nuances, while Reinforcement Learning (RL) approaches usually struggle to balance information retention and token efficiency under the sparse rewards inherent to long-horizon workflows. To bridge this gap, we propose ZipRL, a novel adaptive compression framework tailored for Reinforcement Learning from Verifiable Rewards (RLVR). ZipRL features a multi-granularity compression mechanism for active, non-uniform information reduction, coupled with Hindsight Response Replay (HRR), a technique designed to densify training signals during RLVR optimization. Theoretically, we prove ZipRL's superior task-relevant utility over uniform methods. Concretely, ZipRL utilizes coarse-to-fine prompts for macro-compression and incorporates HRR into GRPO via generalized advantage reshaping. Multiple models of varying versions and parameter scales validate the effectiveness of our approach. Benchmarks on five agent tasks show ZipRL outperforms state-of-the-art approaches by 27.9% and 34.7% across Qwen3-4B and Qwen3-8B models, while maintaining exceptional token efficiency and robustness under extreme 256-turn extrapolation stress tests.

Hang He, Chuhuai Yue, Chengqi Dong et al.

Visual DeepSearch requires multimodal large reasoning model (MLRM) agents to answer complex visual queries by repeatedly inspecting image regions, grounding intermediate reasoning in visual evidence, and connecting fine-grained clues across long reasoning chains. However, existing benchmarks mainly focus on single-step visual understanding or static image-question answering, offering limited evaluation of iterative image inspection, visual-anchor grounding, and multi-hop evidence integration. In this work, we introduce VistaHop, a benchmark for evaluating vision-centric search and multi-hop visual reasoning in Visual DeepSearch. VistaHop contains 300 high-resolution images, 25 visual search scenarios, and 350 multi-hop QA tasks that require models to follow evidence chains from visual anchors or fuse information across multiple image-grounded reasoning paths. We further develop VistaArena, a unified evaluation environment that supports tool-augmented reasoning with text search, image search, image cropping, and evidence-based answer validation. Experiments on seven representative MLRMs show that current models remain far from solving VistaHop: the best model, SenseNova-MARS-32B, achieves only 24.31% Pass@1. These results reveal persistent limitations in visual grounding, evidence revisiting, long-chain reasoning, and multi-anchor information fusion, highlighting the need for stronger benchmarks and training methods for Visual DeepSearch.

Li Wang, Xiaodong Lu, Xiaohan Wang et al.

Large Language Models (LLMs) have achieved remarkable advancements in reasoning capabilities empowered by Reinforcement Learning with Verifiable Rewards (RLVR). Nonetheless, RLVR intrinsically relies on ground-truth labels for reward computation, the acquisition of which is often prohibitively expensive in real-world scenarios. While unsupervised RLVR paradigms attempt to circumvent this by training on pseudo-labels, they are notoriously susceptible to training collapse. Moreover, different samples often exhibit varying annotation values. In this paper, we propose Reinforcement Learning with Active Verifiable Rewards (RLAVR), which actively acquires ground-truth labels for a small set of selected samples and integrates them with pseudo-labels, thereby stabilizing training dynamics and improving performance under limited annotation budgets. To identify valuable samples, we propose the Corrective Advantage Gap (CAG) metric and analyze the sample-level supervision value. Building on this, we introduce Correction-Aware Reliability Estimation for RLAVR (CARE), which translates the oracle CAG criterion into a practical pre-query acquisition policy to substantially improve training stability. Extensive experiments across diverse domains, model families, and model scales demonstrate the effectiveness and generality of our approach. Our code is available at https://github.com/Lumina04/CARE.

Jiajun Chai, Wenzhang Chen, Yuanheng Zhu et al.

Unmanned combat air vehicle (UCAV) combat is a challenging scenario with continuous action space. In this paper, we propose a general hierarchical framework to resolve the within-vision-range (WVR) air-to-air combat problem under 6 dimensions of degree (6-DOF) dynamics. The core idea is to divide the whole decision process into two loops and use reinforcement learning (RL) to solve them separately. The outer loop takes into account the current combat situation and decides the expected macro behavior of the aircraft according to a combat strategy. Then the inner loop tracks the macro behavior with a flight controller by calculating the actual input signals for the aircraft. We design the Markov decision process for both the outer loop strategy and inner loop controller, and train them by proximal policy optimization (PPO) algorithm. For the inner loop controller, we design an effective reward function to accurately track various macro behavior. For the outer loop strategy, we further adopt a fictitious self-play mechanism to improve the combat performance by constantly combating against the historical strategies. Experiment results show that the inner loop controller can achieve better tracking performance than fine-tuned PID controller, and the outer loop strategy can perform complex maneuvers to get higher and higher winning rate, with the generation evolves.

Yaocheng Zhang, Jiajun Chai, Songjun Tu et al.

On-policy distillation (OPD) provides dense teacher feedback along rollouts generated by the student and has emerged as a promising post-training paradigm for long-horizon reasoning. However, standard OPD typically generates full rollouts during training, which is computationally expensive and may expose the student to unreliable teacher feedback at late rollout positions, especially during early training. We identify the rollout horizon as a key bottleneck in OPD that substantially impacts training efficiency. Unlike Reinforcement Learning with Verifiable Rewards (RLVR), OPD does not require a complete trajectory or a final answer reward to provide learning signals. This observation suggests that full rollouts may not always be necessary for effective OPD. Motivated by this insight, we propose two simple horizon-control strategies: Progressive OPD (POPD), which gradually expands the rollout horizon during training, and Truncated OPD (TOPD), which permanently performs distillation on reliable truncated rollouts. Experiments on mathematical reasoning show that POPD improves the training efficiency of OPD by up to 3$\times$, while TOPD matches OPD performance using only 10\% of the rollout horizon, leading to substantial wall-clock and memory reductions. These results demonstrate that controlling the rollout horizon offers a simple and practical path to more efficient OPD.

Li Wang, Xiaohan Wang, Xiaodong Lu et al.

Large language models (LLMs) have increasingly leveraged tool invocation to enhance their reasoning capabilities. However, existing approaches typically tightly couple tool invocation with immediate execution. Such immediate tool interaction may disrupt the reasoning coherence of LLMs and constrain their expressivity, ultimately degrading reasoning performance. To this end, for the first time, we propose and formalize the problem of decoupling tool invocation from execution during reasoning, and introduce delayed execution with explicit control to enhance tool-integrated reasoning (TIR). Furthermore, we propose a hierarchical control framework and theoretically derive a surrogate loss that enables an implicitly hierarchical policy to learn behavior equivalent to that of an explicit hierarchical policy, leading to the proposed IH-GRPO algorithm. Extensive experiments on IH-GRPO achieve absolute improvements of 1.87\%, 2.16\%, and 2.53\% on Qwen3-1.7B, Qwen3-4B, and Qwen3-8B across six out-of-domain mathematical reasoning benchmarks over the strongest baseline method, while also yielding consistent performance gains in other domains. Our code is available at https://github.com/Lumina04/IH-GRPO-01.

Qi Zhang, Yifei Wang, Xiaohan Wang et al.

In recent years, pre-trained large language models have achieved remarkable success across diverse tasks. Besides the pivotal role of self-supervised pre-training, their effectiveness in downstream applications also depends critically on the post-training process, which adapts models to task-specific data and objectives. However, this process inevitably introduces model shifts that can influence performance in different domains, and how such shifts transfer remains poorly understood. To open up the black box, we propose the SAE-based Transferability Score (STS), a new metric that leverages sparse autoencoders (SAEs) to forecast post-training transferability. Taking supervised fine-tuning as an example, STS identifies shifted dimensions in SAE representations and calculates their correlations with downstream domains, enabling reliable estimation of transferability \textit{before} fine-tuning. Extensive experiments across multiple models and domains show that STS accurately predicts the transferability of supervised fine-tuning, achieving Pearson correlation coefficients above 0.7 with actual performance changes. Beyond this, we take an initial step toward extending STS to reinforcement learning. We believe that STS can serve as an {\color{black} interpretable} tool for guiding post-training strategies in LLMs. Code is available at https://github.com/PKU-ML/STS.

Zili Wang, Jiajun Chai, Lin Chen et al.

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as the standard paradigm for improving reasoning capability of large language models, while Multi-Token Prediction (MTP) has been a widely adopted module in pretraining. Combining them is a natural approach, yet current RL practices detach MTP gradients because joint training degrades the performance. We revisit this failure from an optimization perspective. We show that the per-step effect of MTP on the RL objective can be decomposed into two terms: a first-order correlation and a second-order perturbation penalty. This decomposition unifies three MTP training regimes: Detach, Cross-Entropy loss, and Policy loss, and explains why each succeeds or fails. Further analysis of policy loss reveals that, although it aligns with intuition, performance still degrades: the correlation term decays while the quadratic penalty persists. Guided by the analysis, we propose Optimal Coefficient Calibration (OCC), an adaptive scheme that tracks the optimal coefficient online via a log-probability proxy at negligible cost. Across six competition-level mathematical reasoning benchmarks, OCC consistently matches or exceeds the detach baseline, delivering improved joint MTP-RL training performance.

Yaocheng Zhang, Yuanheng Zhu, Wenyue Chong et al.

Deep search agents have emerged as a promising paradigm for addressing complex information-seeking tasks, but their training remains challenging due to sparse rewards, weak credit assignment, and limited labeled data. Self-play offers a scalable route to reduce data dependence, but conventional self-play optimizes students only through sparse outcome rewards, leading to low learning efficiency. In this work, we observe that self-play naturally produces a question construction path (QCP) during task generation, an intermediate artifact that captures the reverse solution process. This reveals a new source of privileged information for self-distillation: self-play can itself provide high-quality privileged context for the teacher model in a low-cost and scalable manner, without relying on human feedback or curated privileged information. Leveraging this insight, we propose Privileged Information Self-Play ($π$-Play), a multi-agent self-evolution framework. In $π$-Play, an examiner generates tasks together with their QCPs, and a teacher model leverages QCP as privileged context to densely supervise a student via self-distillation. This design transforms conventional sparse-reward self-play into a dense-feedback self-evolution loop. Extensive experiments show that data-free $π$-Play surpasses fully supervised search agents and improves evolutionary efficiency by 2-3$\times$ over conventional self-play.

Hao Chen, Zhexin Hu, Jiajun Chai et al.

Training LLMs to invoke tools and leverage retrieved information necessitates high-quality, diverse data. However, existing pipelines for synthetic data generation often rely on tens of thousands of real API calls to enhance generalization, incurring prohibitive costs while lacking multi-hop reasoning and self-reflection. To address these limitations, we introduce ToolForge, an automated synthesis framework that achieves strong real-world tool-calling performance by constructing only a small number of virtual tools, eliminating the need for real API calls. ToolForge leverages a (question, golden context, answer) triple to synthesize large-scale tool-learning data specifically designed for multi-hop search scenarios, further enriching the generated data through multi-hop reasoning and self-reflection mechanisms. To ensure data fidelity, we employ a Multi-Layer Validation Framework that integrates both rule-based and model-based assessments. Empirical results show that a model with only 8B parameters, when trained on our synthesized data, outperforms GPT-4o on multiple benchmarks. Our code and dataset are publicly available at https://github.com/Buycar-arb/ToolForge .

Wenwen Zeng, Jinhui Zhang, Hao Chen et al.

The integration of Large Language Model (LLM) agents is transforming recommender systems from simple query-item matching towards deeply personalized and interactive recommendations. Reinforcement Learning (RL) provides an essential framework for the optimization of these agents in recommendation tasks. However, current methodologies remain limited by a reliance on single dimensional outcome-based rewards that focus exclusively on final user interactions, overlooking critical intermediate capabilities, such as instruction following and complex intent understanding. Despite the necessity for designing multi-dimensional reward, the field lacks a standardized benchmark to facilitate this development. To bridge this gap, we introduce RecRM-Bench, the largest and most comprehensive benchmark to date for agentic recommender systems. It comprises over 1 million structured entries across four core evaluation dimensions: instruction following, factual consistency, query-item relevance, and fine-grained user behavior prediction. By supporting comprehensive assessment from syntactic compliance to complex intent grounding and preference modeling, RecRM-Bench provides a foundational dataset for training sophisticated reward models. Furthermore, we propose a systematic framework for the construction of multi-dimensional reward models and the integration of a hybrid reward function, establishing a robust foundation for developing reliable and highly capable agentic recommender systems. The complete RecRM-Bench dataset is publicly available at https://huggingface.co/datasets/wwzeng/RecRM-Bench.

Jingbo Sun, Wenyue Chong, Songjun Tu et al.

Agentic retrieval-augmented generation (RAG) systems enable large language models (LLMs) to solve complex tasks through multi-step interaction with external retrieval tools. However, such multi-step interaction often involves redundant search steps, incurring substantial computational cost and latency. Prior work limits search depth (i.e., the number of search steps) to reduce cost, but this often leads to underexploration of complex questions. To address this, we first investigate how search depth affects accuracy and find a minimal sufficient search depth that defines an accuracy-efficiency trade-off, jointly determined by question complexity and the agent's capability. Furthermore, we propose AutoSearch, a reinforcement learning (RL) framework that evaluates each search step via self-generated intermediate answers. By a self-answering mechanism, AutoSearch identifies the minimal sufficient search depth and promotes efficient search by rewarding its attainment while penalizing over-searching. In addition, reward mechanisms are introduced to stabilize search behavior and improve answer quality on complex questions. Extensive experiments on multiple benchmarks show that AutoSearch achieves a superior accuracy-efficiency trade-off, alleviating over-searching while preserving search quality.

Zihan Lin, Xiaohan Wang, Jie Cao et al.

Reinforcement Learning with Verifiable Rewards (RLVR) enhances reasoning of Large Language Models (LLMs) but usually exhibits limited generation diversity due to the over-incentivization of positive rewards. Although methods like Negative Sample Reinforcement (NSR) mitigate this issue by upweighting penalty from negative samples, they may suppress the semantic distributions shared between positive and negative responses. To boost reasoning ability without losing diversity, this paper proposes negative sample projection Residual Reinforcement Learning (ResRL) that decouples similar semantic distributions among positive and negative responses. We theoretically link Lazy Likelihood Displacement (LLD) to negative-positive head-gradient interference and derive a single-forward proxy that upper-bounds representation alignment to guide conservative advantage reweighting. ResRL then projects negative-token hidden representations onto an SVD-based low-rank positive subspace and uses projection residuals to modulate negative gradients, improving reasoning while preserving diversity and outperforming strong baselines on average across twelve benchmarks spanning Mathematics, Code, Agent Tasks, and Function Calling. Notably, ResRL surpasses NSR on mathematical reasoning by 9.4\% in Avg@16 and 7.0\% in Pass@128. Code is available at https://github.com/1229095296/ResRL.git.

Fengkai Yang, Zherui Chen, Xiaohan Wang et al.

Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.

Zhenlin Wei, Pu Jian, Yingzhuo Deng et al.

The alignment of Large Language Models (LLMs) for complex reasoning heavily relies on Reinforcement Learning with Verifiable Rewards (RLVR). However, standard algorithms like GRPO apply sequence-level rewards uniformly to all tokens, creating a severe credit-assignment bottleneck. While on-policy self-distillation attempts to resolve this by conditioning a self-teacher on privileged contexts, direct exposure to raw oracle solutions often induces over-conditioned teacher distributions, implicit answer leakage, and late-stage training collapse. To overcome these limitations, we propose Asymmetric Meta-Reflective Self-Distillation (AMR-SD). Instead of conditioning directly on raw reference traces, AMR-SD inserts a reflection bottleneck: it compresses diagnostic signals -- from verifier outcomes, peer rollouts, or reference feedback -- into concise, self-generated Socratic hints and critiques. Furthermore, we introduce Causal Information Gain (CIG) with an asymmetric, ReLU-gated threshold to translate these reflections into sparse, highly precise token-level advantage modulations. Combined with temporal annealing, this mechanism preserves the base environmental reward while filtering out distributional noise. Experiments across scientific, mathematical, and tool-use benchmarks demonstrate that AMR-SD significantly outperforms existing baselines, achieving robust long-horizon stability and successfully preventing late-stage collapse.

Xiaodong Lu, Xiaohan Wang, Jiajun Chai et al.

Reinforcement Learning with Verifiable Rewards (RLVR) is an effective paradigm for improving the reasoning capabilities of large language models. However, existing RLVR methods utilize rollouts in an indiscriminate and short-horizon manner: responses of heterogeneous quality within each prompt are treated uniformly, and historical rollouts are discarded after a single use. This leads to noisy supervision, poor sample efficiency, and suboptimal policy updates. We address these issues by formulating rollout scheduling in RLVR as a contextual bandit problem and proposing a unified neural scheduling framework that adaptively selects high-value rollouts throughout training. Each rollout is treated as an arm whose reward is defined by the induced performance gain between consecutive optimization steps. The resulting scheduler supports both noise-aware intra-group selection and adaptive global reuse of historical rollouts within a single principled framework. We provide theoretical justification by deriving sublinear regret bounds and showing that enlarging the rollout buffer improves the achievable performance upper bound. Experiments on six mathematical reasoning benchmarks demonstrate consistent gains in performance and training efficiency across multiple RLVR optimization methods.

Jiajun Chai, Guojun Yin, Zekun Xu et al.

Large language models excel at basic reasoning but struggle with tasks that require interaction with external tools. We present RLFactory, a plug-and-play reinforcement learning post-training framework for multi-round tool use. RLFactory tackles (i) tool-call stability and adaptability amid tool heterogeneity and interface issues via an asyncio-based asynchronous caller and a decoupled tool/training architecture, and (ii) diverse evaluation needs via a reward layer supporting rule-based, model-judgment, and tool-verification signals. It reconstructs the MDP by introducing observation markers from tool feedback, closing the loop among model, tools, and environment, and implements a generate-parse-invoke-update workflow for dynamic policy optimization. On Search-R1 with Qwen3-4B, RLFactory achieves a 0.486 test score on the Natural Questions (NQ) dataset, surpassing larger models trained with similar techniques (e.g., Qwen2.5-7B-Instruct-GRPO at 0.473), and increases training throughput by 6.8x. RLFactory provides a low-barrier, highly adaptable framework for strengthening multi-round tool use of LLMs in real-world scenarios. Code: https://github.com/Simple-Efficient/RL-Factory.

Siyu Xia, Zekun Xu, Jiajun Chai et al.

Large Language Models (LLMs) based agents have demonstrated remarkable potential in autonomous task-solving across complex, open-ended environments. A promising approach for improving the reasoning capabilities of LLM agents is to better utilize prior experiences in guiding current decisions. However, LLMs acquire experience either through implicit memory via training, which suffers from catastrophic forgetting and limited interpretability, or explicit memory via prompting, which lacks adaptability. In this paper, we introduce a novel agent-centric, trainable, multi-layered graph memory framework and evaluate how context memory enhances the ability of LLMs to utilize parametric information. The graph abstracts raw agent trajectories into structured decision paths in a state machine and further distills them into high-level, human-interpretable strategic meta-cognition. In order to make memory adaptable, we propose a reinforcement-based weight optimization procedure that estimates the empirical utility of each meta-cognition based on reward feedback from downstream tasks. These optimized strategies are then dynamically integrated into the LLM agent's training loop through meta-cognitive prompting. Empirically, the learnable graph memory delivers robust generalization, improves LLM agents' strategic reasoning performance, and provides consistent benefits during Reinforcement Learning (RL) training.

Zihan Lin, Xiaohan Wang, Hexiong Yang et al.

While reinforcement learning (RL) shows promise in training tool-use large language models (LLMs) using verifiable outcome rewards, existing methods largely overlook the potential of explicit reasoning rewards to bolster reasoning and tool utilization. Furthermore, natively combining reasoning and outcome rewards may yield suboptimal performance or conflict with the primary optimization objective. To address this, we propose advantage-weighted policy optimization (AWPO) -- a principled RL framework that effectively integrates explicit reasoning rewards to enhance tool-use capability. AWPO incorporates variance-aware gating and difficulty-aware weighting to adaptively modulate advantages from reasoning signals based on group-relative statistics, alongside a tailored clipping mechanism for stable optimization. Extensive experiments demonstrate that AWPO achieves state-of-the-art performance across standard tool-use benchmarks, significantly outperforming strong baselines and leading closed-source models in challenging multi-turn scenarios. Notably, with exceptional parameter efficiency, our 4B model surpasses Grok-4 by 16.0 percent in multi-turn accuracy while preserving generalization capability on the out-of-distribution MMLU-Pro benchmark.

Hang He, Chuhuai Yue, Chengqi Dong et al.

Recent advances in large reasoning models (LRMs) have enabled agentic search systems to perform complex multi-step reasoning across multiple sources. However, most studies focus on general information retrieval and rarely explores vertical domains with unique challenges. In this work, we focus on local life services and introduce LocalSearchBench, which encompass diverse and complex business scenarios. Real-world queries in this domain are often ambiguous and require multi-hop reasoning across merchants and products, remaining challenging and not fully addressed. As the first comprehensive benchmark for agentic search in local life services, LocalSearchBench includes over 150,000 high-quality entries from various cities and business types. We construct 300 multi-hop QA tasks based on real user queries, challenging agents to understand questions and retrieve information in multiple steps. We also developed LocalPlayground, a unified environment integrating multiple tools for agent interaction. Experiments show that even state-of-the-art LRMs struggle on LocalSearchBench: the best model (DeepSeek-V3.1) achieves only 34.34% correctness, and most models have issues with completeness (average 77.33%) and faithfulness (average 61.99%). This highlights the need for specialized benchmarks and domain-specific agent training in local life services. Code, Benchmark, and Leaderboard are available at localsearchbench.github.io.

Chengqi Dong, Chuhuai Yue, Hang He et al.

Recent VLM-based agents aim to replicate OpenAI O3's ``thinking with images" via tool use, but most open-source methods limit input to a single image, falling short on real-world multi-image QA tasks. To address this, we propose IMAgent, an open-source vision agent trained via end-to-end reinforcement learning dedicated for complex multi-image tasks. By leveraging a multi-agent system, we generate challenging and visually-rich multi-image QA pairs to fully activate the tool-use potential of the base VLM. Through manual verification, we obtain MIFG-QA, comprising 10k samples for training and evaluation. With deeper reasoning steps, VLMs may increasingly ignore visual inputs. We therefore develop two specialized tools for visual reflection and confirmation, allowing the model to proactively reallocate its attention to image content during inference. Benefiting from our well-designed action-trajectory two-level mask strategy, IMAgent achieves stable tool use behavior via pure RL training without requiring costly supervised fine-tuning data. Extensive experiments demonstrate that IMAgent maintains strong performance on existing single-image benchmarks while achieving substantial improvements on our proposed multi-image dataset, with our analysis providing actionable insights for the research community. Codes and data will be released soon.

Yuqian Fu, Tinghong Chen, Jiajun Chai et al.

Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet the optimal integration of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from entropy-based perspectives, we reveal key differences between these paradigms: SFT induces coarse-grained global changes to LLM policy distributions, while RL performs fine-grained selective optimizations, with entropy serving as a critical indicator of training effectiveness. Building on these observations, we propose Supervised Reinforcement Fine-Tuning (SRFT), a single-stage method that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. Our approach simultaneously applies SFT and RL to directly optimize the LLM using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT achieves 59.1% average accuracy, outperforming zero-RL methods by 9.0% on five mathematical reasoning benchmarks and 10.9% on three out-of-distribution benchmarks.

Yuqian Fu, Yuanheng Zhu, Jiajun Chai et al.

Ensembling large language models (LLMs) can effectively combine diverse strengths of different models, offering a promising approach to enhance performance across various tasks. However, existing methods typically rely on fixed weighting strategies that fail to adapt to the dynamic, context-dependent characteristics of LLM capabilities. In this work, we propose Reinforcement Learning-Assisted Ensemble for LLMs (RLAE), a novel framework that reformulates LLM ensemble through the lens of a Markov Decision Process (MDP). Our approach introduces a RL agent that dynamically adjusts ensemble weights by considering both input context and intermediate generation states, with the agent being trained using rewards that directly correspond to the quality of final outputs. We implement RLAE using both single-agent and multi-agent reinforcement learning algorithms ($\text{RLAE}_\text{PPO}$ and $\text{RLAE}_\text{MAPPO}$ ), demonstrating substantial improvements over conventional ensemble methods. Extensive evaluations on a diverse set of tasks show that RLAE outperforms existing approaches by up to $3.3\%$ accuracy points, offering a more effective framework for LLM ensembling. Furthermore, our method exhibits superior generalization capabilities across different tasks without the need for retraining, while simultaneously achieving lower time latency.

Chuhuai Yue, Chengqi Dong, Yinan Gao et al.

Large reasoning models (LRMs) have recently achieved significant progress in complex reasoning tasks, aided by reinforcement learning with verifiable rewards. However, LRMs often suffer from overthinking, expending excessive computation on simple problems and reducing efficiency. Existing efficient reasoning methods typically require accurate task assessment to preset token budgets or select reasoning modes, which limits their flexibility and reliability. In this work, we revisit the essence of overthinking and identify that encouraging effective steps while penalizing ineffective ones is key to its solution. To this end, we propose a novel rule-based verifiable stepwise reward mechanism (VSRM), which assigns rewards based on the performance of intermediate states in the reasoning trajectory. This approach is intuitive and naturally fits the step-by-step nature of reasoning tasks. We conduct extensive experiments on standard mathematical reasoning benchmarks, including AIME24 and AIME25, by integrating VSRM with PPO and Reinforce++. Results show that our method achieves substantial output length reduction while maintaining original reasoning performance, striking an optimal balance between efficiency and accuracy. Further analysis of overthinking frequency and pass@k score before and after training demonstrates that our approach in deed effectively suppresses ineffective steps and encourages effective reasoning, fundamentally alleviating the overthinking problem. All code will be released upon acceptance.

Kaixuan Xu, Jiajun Chai, Sicheng Li et al.

Diplomacy is a complex multiplayer game that requires both cooperation and competition, posing significant challenges for AI systems. Traditional methods rely on equilibrium search to generate extensive game data for training, which demands substantial computational resources. Large Language Models (LLMs) offer a promising alternative, leveraging pre-trained knowledge to achieve strong performance with relatively small-scale fine-tuning. However, applying LLMs to Diplomacy remains challenging due to the exponential growth of possible action combinations and the intricate strategic interactions among players. To address this challenge, we propose DipLLM, a fine-tuned LLM-based agent that learns equilibrium policies for Diplomacy. DipLLM employs an autoregressive factorization framework to simplify the complex task of multi-unit action assignment into a sequence of unit-level decisions. By defining an equilibrium policy within this framework as the learning objective, we fine-tune the model using only 1.5% of the data required by the state-of-the-art Cicero model, surpassing its performance. Our results demonstrate the potential of fine-tuned LLMs for tackling complex strategic decision-making in multiplayer games.

Dengcan Liu, Fengkai Yang, Xiaohan Wang et al.

Reward modeling is essential for aligning Large Language Models(LLMs) with human preferences, yet conventional reward models suffer from poor interpretability and heavy reliance on costly expert annotations. While recent rubric-based approaches enhance evaluation transparency, they lack systematic quality control, yielding noisy and redundant criteria, failing to mitigate persistent biases (e.g., verbosity, position) in LLM evaluators, and creating a scalability-reliability trade-off. To address these limitations, we propose CDRRM (Contrast-Driven Rubric Reward Model), a framework built on a novel Contrast-then-Synthesis paradigm for high-quality rubric generation and guided preference judgment. CDRRM first conducts multi-dimensional contrastive profiling on preference pairs to identify causal discriminative factors, then synthesizes these insights into compact, context-aware rubrics to guide preference judg- ments. Extensive experiments on three authoritative benchmarks (RewardBench, RMBench, RMB) demonstrate that CDRRM achieves state-of-the-art performance across diverse domains and effectively mitigates aforementioned evaluation biases. Notably, our approach delivers exceptional data efficiency: training the rubric generator on only 3k high-quality samples empowers a frozen pre-trained judge model to outperform fully fine-tuned baselines. This work offers a scalable, interpretable, and data-efficient path for reward modeling.

Xiaojun Guo, Runyu Zhou, Yifei Wang et al.

Vision-language models (VLMs) have shown remarkable abilities by integrating large language models with visual inputs. However, they often fail to utilize visual evidence adequately, either depending on linguistic priors in vision-centric tasks or resorting to textual shortcuts during reasoning. Although reinforcement learning (RL) can align models with desired behaviors, its application to VLMs has been hindered by the lack of scalable and reliable reward mechanisms. To overcome this challenge, we propose SSL4RL, a novel framework that leverages self-supervised learning (SSL) tasks as a source of verifiable rewards for RL-based fine-tuning. Our approach reformulates SSL objectives-such as predicting image rotation or reconstructing masked patches-into dense, automatic reward signals, eliminating the need for human preference data or unreliable AI evaluators. Experiments show that SSL4RL substantially improves performance on both vision-centric and vision-language reasoning benchmarks. Furthermore, through systematic ablations, we identify key factors-such as task difficulty, model scale, and semantic alignment with the target domain-that influence the effectiveness of SSL4RL tasks, offering new design principles for future work. We also demonstrate the framework's generality by applying it to graph learning, where it yields significant gains. SSL4RL establishes a versatile and effective paradigm for aligning multimodal models using verifiable, self-supervised objectives.

Zihan Lin, Xiaohan Wang, Jie Cao et al.

Large language models (LLMs) transcend passive generation and act as goal-directed agents by invoking external tools. Reinforcement learning (RL) offers a principled framework for optimizing these emergent tool-use policies, yet the prevailing paradigm relies exclusively on sparse outcome rewards and lacks consideration of the particularity of tool-use tasks, inflating policy-gradient variance and resulting in inefficient training. To better understand and address these challenges, we first establish a theoretical link between policy entropy and training stability of tool-use tasks, which reveals that structured, low-entropy tokens are primary determinants of rewards. Motivated by this insight, we propose \textbf{Res}haped \textbf{T}oken-level policy gradients (\textbf{ResT}) for tool-use tasks. ResT reshapes the policy gradient through entropy-informed token reweighting, progressively upweighting reasoning tokens as training proceeds. This entropy-aware scheme enables a smooth shift from structural correctness to semantic reasoning and stabilizes convergence in multi-turn tool-use tasks. Evaluation on BFCL and API-Bank shows that ResT achieves state-of-the-art results, outperforming prior methods by up to $8.76\%$. When fine-tuned on a 4B base LLM, ResT further surpasses GPT-4o by $4.11\%$ on single-turn tasks and $1.50\%$ on multi-turn base tasks.

Zekun Xu, Siyu Xia, Chuhuai Yue et al.

As large language models (LLMs) are increasingly used in Text-to-SQL tasks, Reinforcement Learning (RL) has become a common method for improving performance. Existing methods primarily rely on static execution feedback, which restricts real-time error correction. However, integrating multi-turn tool invocation along with dynamic feedback could significantly improve adaptability and robustness, ultimately enhancing model performance. To address these issues, we propose MTIR-SQL, an innovative Multi-turn Tool-Integrated Reasoning reinforcement learning framework for Text-to-SQL. Our approach introduces an execution-aware multi-turn reasoning paradigm that seamlessly incorporates database execution feedback at each reasoning step, enabling context-sensitive query generation and progressive refinement throughout the reasoning process. The framework extends the GRPO algorithm to accommodate complex multi-turn interaction scenarios. Considering the training instability characteristics of MTIR and the potential for significant Deviation of model distribution from the initial model, we enhance the GRPO algorithm by adding a trajectory filtering mechanism and removing KL loss constraints. Experimental results demonstrate that MTIR-SQL, with 4B parameters, achieves \textbf{64.4}\% accuracy in the BIRD Dev and 84.6% execution accuracy in the SPIDER Dev, significantly outperforming existing approaches.