Time-series reasoning remains a significant challenge in multimodal large language models (MLLMs) due to the dynamic temporal patterns, ambiguous semantics, and lack of temporal priors. In this work, we introduce TimeMaster, a reinforcement learning (RL)-based method that enables time-series MLLMs to perform structured, interpretable reasoning directly over visualized time-series inputs and task prompts. TimeMaster adopts a three-part structured output format, reasoning, classification, and domain-specific extension, and is optimized via a composite reward function that aligns format adherence, prediction accuracy, and open-ended insight quality. The model is trained using a two-stage pipeline: we first apply supervised fine-tuning (SFT) to establish a good initialization, followed by Group Relative Policy Optimization (GRPO) at the token level to enable stable and targeted reward-driven improvement in time-series reasoning. We evaluate TimeMaster on the TimerBed benchmark across six real-world classification tasks based on Qwen2.5-VL-3B-Instruct. TimeMaster achieves state-of-the-art performance, outperforming both classical time-series models and few-shot GPT-4o by over 14.6% and 7.3% performance gain, respectively. Notably, TimeMaster goes beyond time-series classification: it also exhibits expert-like reasoning behavior, generates context-aware explanations, and delivers domain-aligned insights. Our results highlight that reward-driven RL can be a scalable and promising path toward integrating temporal understanding into time-series MLLMs.
Recommender systems have been widely used in various large-scale user-oriented platforms for many years. However, compared to the rapid developments in the AI community, recommendation systems have not achieved a breakthrough in recent years. For instance, they still rely on a multi-stage cascaded architecture rather than an end-to-end approach, leading to computational fragmentation and optimization inconsistencies, and hindering the effective application of key breakthrough technologies from the AI community in recommendation scenarios. To address these issues, we propose OneRec, which reshapes the recommendation system through an end-to-end generative approach and achieves promising results. Firstly, we have enhanced the computational FLOPs of the current recommendation model by 10 $\times$ and have identified the scaling laws for recommendations within certain boundaries. Secondly, reinforcement learning techniques, previously difficult to apply for optimizing recommendations, show significant potential in this framework. Lastly, through infrastructure optimizations, we have achieved 23.7% and 28.8% Model FLOPs Utilization (MFU) on flagship GPUs during training and inference, respectively, aligning closely with the LLM community. This architecture significantly reduces communication and storage overhead, resulting in operating expense that is only 10.6% of traditional recommendation pipelines. Deployed in Kuaishou/Kuaishou Lite APP, it handles 25% of total queries per second, enhancing overall App Stay Time by 0.54% and 1.24%, respectively. Additionally, we have observed significant increases in metrics such as 7-day Lifetime, which is a crucial indicator of recommendation experience. We also provide practical lessons and insights derived from developing, optimizing, and maintaining a production-scale recommendation system with significant real-world impact.
The development of large language models (LLMs) has entered in a experience-driven era, flagged by the emergence of environment feedback-driven learning via reinforcement learning and tool-using agents. This encourages the emergenece of model context protocol (MCP), which defines the standard on how should a LLM interact with external services, such as \api and data. However, as MCP becomes the de facto standard for LLM agent systems, it also introduces new safety risks. In particular, MCP introduces third-party services, which are not controlled by the LLM developers, into the agent systems. These third-party MCP services provider are potentially malicious and have the economic incentives to exploit vulnerabilities and sabotage user-agent interactions. In this position paper, we advocate the research community in LLM safety to pay close attention to the new safety risks issues introduced by MCP, and develop new techniques to build safe MCP-powered agent systems. To establish our position, we argue with three key parts. (1) We first construct \framework, a controlled framework to examine safety issues in MCP-powered agent systems. (2) We then conduct a series of pilot experiments to demonstrate the safety risks in MCP-powered agent systems is a real threat and its defense is not trivial. (3) Finally, we give our outlook by showing a roadmap to build safe MCP-powered agent systems. In particular, we would call for researchers to persue the following research directions: red teaming, MCP safe LLM development, MCP safety evaluation, MCP safety data accumulation, MCP service safeguard, and MCP safe ecosystem construction. We hope this position paper can raise the awareness of the research community in MCP safety and encourage more researchers to join this important research direction. Our code is available at https://github.com/littlelittlenine/SafeMCP.git.
Clinical decision-making is a dynamic, interactive, and cyclic process where doctors have to repeatedly decide on which clinical action to perform and consider newly uncovered information for diagnosis and treatment. Large Language Models (LLMs) have the potential to support clinicians in this process, however, most applications of LLMs in clinical decision support suffer from one of two limitations: Either they assume the unrealistic scenario of immediate availability of all patient information and do not model the interactive and iterative investigation process, or they restrict themselves to the limited "out-of-the-box" capabilities of large pre-trained models without performing task-specific training. In contrast to this, we propose to model clinical decision-making for diagnosis with a hypothesis-driven uncertainty-aware language agent, LA-CDM, that converges towards a diagnosis via repeatedly requesting and interpreting relevant tests. Using a hybrid training paradigm combining supervised and reinforcement learning, we train LA-CDM with three objectives targeting critical aspects of clinical decision-making: accurate hypothesis generation, hypothesis uncertainty estimation, and efficient decision-making. We evaluate our methodology on MIMIC-CDM, a real-world dataset covering four abdominal diseases containing various clinical tests and show the benefit of explicitly training clinical decision-making for increasing diagnostic performance and efficiency.
Current Reinforcement Learning (RL) methodologies for Large Language Models (LLMs) often rely on simplistic, outcome-based reward signals (e.g., final answer correctness), which limits the depth of learning from each interaction. This paper introduces Socratic Reinforcement Learning (Socratic-RL), a novel, process-oriented framework designed to address this limitation. Socratic-RL operates on the principle that deeper understanding is achieved by reflecting on the causal reasons for errors and successes within the reasoning process itself. The framework employs a decoupled "Teacher-Student" architecture, where a "Teacher AI" analyzes interaction histories, extracts causal insights, and formulates them into structured "viewpoints." These viewpoints, acting as distilled guidance, are then used by a "Student AI" to enhance its subsequent reasoning. A key innovation is the iterative self-improvement of the Teacher AI, enabling its reflective capabilities to evolve through a meta-learning loop. To manage the accumulation of knowledge, a distillation mechanism compresses learned viewpoints into the Student's parameters. By focusing on process rather than just outcome, Socratic-RL presents a pathway toward enhanced sample efficiency, superior interpretability, and a more scalable architecture for self-improving AI systems. This paper details the foundational concepts, formal mechanisms, synergies, challenges, and a concrete research roadmap for this proposed framework.
Recent advances in Large Language Models (LLMs) have showcased impressive reasoning abilities in structured tasks like mathematics and programming, largely driven by Reinforcement Learning with Verifiable Rewards (RLVR), which uses outcome-based signals that are scalable, effective, and robust against reward hacking. However, applying similar techniques to open-ended long-form reasoning tasks remains challenging due to the absence of generic, verifiable reward signals. To address this, we propose Direct Reasoning Optimization (DRO), a reinforcement learning framework for fine-tuning LLMs on open-ended, particularly long-form, reasoning tasks, guided by a new reward signal: the Reasoning Reflection Reward (R3). At its core, R3 selectively identifies and emphasizes key tokens in the reference outcome that reflect the influence of the model's preceding chain-of-thought reasoning, thereby capturing the consistency between reasoning and reference outcome at a fine-grained level. Crucially, R3 is computed internally using the same model being optimized, enabling a fully self-contained training setup. Additionally, we introduce a dynamic data filtering strategy based on R3 for open-ended reasoning tasks, reducing cost while improving downstream performance. We evaluate DRO on two diverse datasets -- ParaRev, a long-form paragraph revision task, and FinQA, a math-oriented QA benchmark -- and show that it consistently outperforms strong baselines while remaining broadly applicable across both open-ended and structured domains.
Recent advancements in large language models (LLMs) have witnessed a surge in the development of advanced reasoning paradigms, which are now being integrated into multimodal large language models (MLLMs). However, existing approaches often fall short: methods solely employing reinforcement learning (RL) can struggle with sample inefficiency and activating entirely absent reasoning capabilities, while conventional pipelines that initiate with a cold-start supervised fine-tuning (SFT) phase before RL may restrict the model's exploratory capacity and face suboptimal convergence. In this work, we introduce \textbf{Metis-RISE} (\textbf{R}L \textbf{I}ncentivizes and \textbf{S}FT \textbf{E}nhances) for multimodal reasoning model learning. Unlike conventional approaches, Metis-RISE distinctively omits an initial SFT stage, beginning instead with an RL phase (e.g., using a Group Relative Policy Optimization variant) to incentivize and activate the model's latent reasoning capacity. Subsequently, the targeted SFT stage addresses two key challenges identified during RL: (1) \textit{inefficient trajectory sampling} for tasks where the model possesses but inconsistently applies correct reasoning, which we tackle using self-distilled reasoning trajectories from the RL model itself; and (2) \textit{fundamental capability absence}, which we address by injecting expert-augmented knowledge for prompts where the model entirely fails. This strategic application of RL for incentivization followed by SFT for enhancement forms the core of Metis-RISE, leading to two versions of our MLLMs (7B and 72B parameters). Evaluations on the OpenCompass Multimodal Reasoning Leaderboard demonstrate that both models achieve state-of-the-art performance among similar-sized models, with the 72B version ranking fourth overall.
The Da Vinci Code, a game of logical deduction and imperfect information, presents unique challenges for artificial intelligence, demanding nuanced reasoning beyond simple pattern recognition. This paper investigates the efficacy of various AI paradigms in mastering this game. We develop and evaluate three distinct agent architectures: a Transformer-based baseline model with limited historical context, several Large Language Model (LLM) agents (including Gemini, DeepSeek, and GPT variants) guided by structured prompts, and an agent based on Proximal Policy Optimization (PPO) employing a Transformer encoder for comprehensive game history processing. Performance is benchmarked against the baseline, with the PPO-based agent demonstrating superior win rates ($58.5\% \pm 1.0\%$), significantly outperforming the LLM counterparts. Our analysis highlights the strengths of deep reinforcement learning in policy refinement for complex deductive tasks, particularly in learning implicit strategies from self-play. We also examine the capabilities and inherent limitations of current LLMs in maintaining strict logical consistency and strategic depth over extended gameplay, despite sophisticated prompting. This study contributes to the broader understanding of AI in recreational games involving hidden information and multi-step logical reasoning, offering insights into effective agent design and the comparative advantages of different AI approaches.
With the rapid development of large language models (LLMs), they have significantly improved efficiency across a wide range of domains. However, recent studies have revealed that LLMs often exhibit gender bias, leading to serious social implications. Detecting, classifying, and mitigating gender bias in LLMs has therefore become a critical research focus. In the NLPCC 2025 Shared Task 7: Chinese Corpus for Gender Bias Detection, Classification and Mitigation Challenge, we investigate how to enhance the capabilities of LLMs in gender bias detection, classification, and mitigation. We adopt reinforcement learning, chain-of-thoughts (CoT) reasoning, and supervised fine-tuning to handle different Subtasks. Specifically, for Subtasks 1 and 2, we leverage the internal reasoning capabilities of LLMs to guide multi-step thinking in a staged manner, which simplifies complex biased queries and improves response accuracy. For Subtask 3, we employ a reinforcement learning-based approach, annotating a preference dataset using GPT-4. We then apply Direct Preference Optimization (DPO) to mitigate gender bias by introducing a loss function that explicitly favors less biased completions over biased ones. Our approach ranked first across all three subtasks of the NLPCC 2025 Shared Task 7.
Intelligent Transportation Systems (ITSs) have emerged as a promising solution towards ameliorating urban traffic congestion, with Traffic Signal Control (TSC) identified as a critical component. Although Multi-Agent Reinforcement Learning (MARL) algorithms have shown potential in optimizing TSC through real-time decision-making, their scalability and effectiveness often suffer from large-scale and complex environments. Typically, these limitations primarily stem from a fundamental mismatch between the exponential growth of the state space driven by the environmental heterogeneities and the limited modeling capacity of current solutions. To address these issues, this paper introduces a novel MARL framework that integrates Dynamic Graph Neural Networks (DGNNs) and Topological Data Analysis (TDA), aiming to enhance the expressiveness of environmental representations and improve agent coordination. Furthermore, inspired by the Mixture of Experts (MoE) architecture in Large Language Models (LLMs), a topology-assisted spatial pattern disentangling (TSD)-enhanced MoE is proposed, which leverages topological signatures to decouple graph features for specialized processing, thus improving the model's ability to characterize dynamic and heterogeneous local observations. The TSD module is also integrated into the policy and value networks of the Multi-agent Proximal Policy Optimization (MAPPO) algorithm, further improving decision-making efficiency and robustness. Extensive experiments conducted on real-world traffic scenarios, together with comprehensive theoretical analysis, validate the superior performance of the proposed framework, highlighting the model's scalability and effectiveness in addressing the complexities of large-scale TSC tasks.
This survey paper outlines the key developments in the field of Large Language Models (LLMs), such as enhancing their reasoning skills, adaptability to various tasks, increased computational efficiency, and ability to make ethical decisions. The techniques that have been most effective in bridging the gap between human and machine communications include the Chain-of-Thought prompting, Instruction Tuning, and Reinforcement Learning from Human Feedback. The improvements in multimodal learning and few-shot or zero-shot techniques have further empowered LLMs to handle complex jobs with minor input. They also manage to do more with less by applying scaling and optimization tricks for computing power conservation. This survey also offers a broader perspective on recent advancements in LLMs going beyond isolated aspects such as model architecture or ethical concerns. It categorizes emerging methods that enhance LLM reasoning, efficiency, and ethical alignment. It also identifies underexplored areas such as interpretability, cross-modal integration and sustainability. With recent progress, challenges like huge computational costs, biases, and ethical risks remain constant. Addressing these requires bias mitigation, transparent decision-making, and clear ethical guidelines. Future research will focus on enhancing models ability to handle multiple input, thereby making them more intelligent, safe, and reliable.
Medical Question-Answering (QA) encompasses a broad spectrum of tasks, including multiple choice questions (MCQ), open-ended text generation, and complex computational reasoning. Despite this variety, a unified framework for delivering high-quality medical QA has yet to emerge. Although recent progress in reasoning-augmented large language models (LLMs) has shown promise, their ability to achieve comprehensive medical understanding is still largely unexplored. In this paper, we present Med-U1, a unified framework for robust reasoning across medical QA tasks with diverse output formats, ranging from MCQs to complex generation and computation tasks. Med-U1 employs pure large-scale reinforcement learning with mixed rule-based binary reward functions, incorporating a length penalty to manage output verbosity. With multi-objective reward optimization, Med-U1 directs LLMs to produce concise and verifiable reasoning chains. Empirical results reveal that Med-U1 significantly improves performance across multiple challenging Med-QA benchmarks, surpassing even larger specialized and proprietary models. Furthermore, Med-U1 demonstrates robust generalization to out-of-distribution (OOD) tasks. Extensive analysis presents insights into training strategies, reasoning chain length control, and reward design for medical LLMs. The code will be released.
Self-reflection -- the ability of a large language model (LLM) to revisit, evaluate, and revise its own reasoning -- has recently emerged as a powerful behavior enabled by reinforcement learning with verifiable rewards (RLVR). While self-reflection correlates with improved reasoning accuracy, its origin and underlying mechanisms remain poorly understood. In this work, {\it we first show that self-reflection is not exclusive to RLVR fine-tuned models: it already emerges, albeit rarely, in pretrained models}. To probe this latent ability, we introduce Reflection-Inducing Probing, a method that injects reflection-triggering reasoning traces from fine-tuned models into pretrained models. This intervention raises self-reflection frequency of Qwen2.5 from 0.6\% to 18.6\%, revealing a hidden capacity for reflection. Moreover, our analysis of internal representations shows that both pretrained and fine-tuned models maintain hidden states that distinctly separate self-reflective from non-reflective contexts. Leveraging this observation, {\it we then construct a self-reflection vector, a direction in activation space associated with self-reflective reasoning}. By manipulating this vector, we enable bidirectional control over the self-reflective behavior for both pretrained and fine-tuned models. Experiments across multiple reasoning benchmarks show that enhancing these vectors improves reasoning performance by up to 12\%, while suppressing them reduces computational cost, providing a flexible mechanism to navigate the trade-off between reasoning quality and efficiency without requiring additional training. Our findings further our understanding of self-reflection and support a growing body of work showing that understanding model internals can enable precise behavioral control.
Reinforcement learning (RL) with tree search has demonstrated superior performance in traditional reasoning tasks. Compared to conventional independent chain sampling strategies with outcome supervision, tree search enables better exploration of the reasoning space and provides dense, on-policy process rewards during RL training but remains under-explored in On-Policy LLM RL. We propose TreeRL, a reinforcement learning framework that directly incorporates on-policy tree search for RL training. Our approach includes intermediate supervision and eliminates the need for a separate reward model training. Existing approaches typically train a separate process reward model, which can suffer from distribution mismatch and reward hacking. We also introduce a cost-effective tree search approach that achieves higher search efficiency under the same generation token budget by strategically branching from high-uncertainty intermediate steps rather than using random branching. Experiments on challenging math and code reasoning benchmarks demonstrate that TreeRL achieves superior performance compared to traditional ChainRL, highlighting the potential of tree search for LLM. TreeRL is open-sourced at https://github.com/THUDM/TreeRL.
Database connectors are critical components enabling applications to interact with underlying database management systems (DBMS), yet their security vulnerabilities often remain overlooked. Unlike traditional software defects, connector vulnerabilities exhibit subtle behavioral patterns and are inherently challenging to detect. Besides, nonstandardized implementation of connectors leaves potential risks (a.k.a. unsafe implementations) but is more elusive. As a result, traditional fuzzing methods are incapable of finding such vulnerabilities. Even for LLM-enable test case generation, due to a lack of domain knowledge, they are also incapable of generating test cases that invoke all interface and internal logic of connectors. In this paper, we propose reinforcement learning (RL)-guided LLM test-case generation for database connector testing. Specifically, to equip the LLM with sufficient and appropriate domain knowledge, a parameterized prompt template is composed which can be utilized to generate numerous prompts. Test cases are generated via LLM with a prompt, and are dynamically evaluated through differential testing across multiple connectors. The testing is iteratively conducted, with each round RL is adopted to select optimal prompt based on prior-round behavioral feedback, so as to maximize control flow coverage. We implement aforementioned methodology in a practical tool and evaluate it on two widely used JDBC connectors: MySQL Connector/J and OceanBase Connector/J. In total, we reported 16 bugs, among them 10 are officially confirmed and the rest are acknowledged as unsafe implementations.
The recent advent of reasoning models like OpenAI's o1 was met with excited speculation by the AI community about the mechanisms underlying these capabilities in closed models, followed by a rush of replication efforts, particularly from the open source community. These speculations were largely settled by the demonstration from DeepSeek-R1 that chains-of-thought and reinforcement learning (RL) can effectively replicate reasoning on top of base LLMs. However, it remains valuable to explore alternative methods for theoretically eliciting reasoning that could help elucidate the underlying mechanisms, as well as providing additional methods that may offer complementary benefits. Here, we build on the long-standing literature in cognitive psychology and cognitive architectures, which postulates that reasoning arises from the orchestrated, sequential execution of a set of modular, predetermined cognitive operations. Crucially, we implement this key idea within a modern agentic tool-calling framework. In particular, we endow an LLM with a small set of "cognitive tools" encapsulating specific reasoning operations, each executed by the LLM itself. Surprisingly, this simple strategy results in considerable gains in performance on standard mathematical reasoning benchmarks compared to base LLMs, for both closed and open-weight models. For instance, providing our "cognitive tools" to GPT-4.1 increases its pass@1 performance on AIME2024 from 26.7% to 43.3%, bringing it very close to the performance of o1-preview. In addition to its practical implications, this demonstration contributes to the debate regarding the role of post-training methods in eliciting reasoning in LLMs versus the role of inherent capabilities acquired during pre-training, and whether post-training merely uncovers these latent abilities.
Traditional information retrieval (IR) methods excel at textual and semantic matching but struggle in reasoning-intensive retrieval tasks that require multi-hop inference or complex semantic understanding between queries and documents. One promising solution is to explicitly rewrite or augment queries using large language models (LLMs) to elicit reasoning-relevant content prior to retrieval. However, the widespread use of large-scale language models like GPT-4 or LLaMA3-70B remains impractical due to their high inference cost and limited deployability in real-world systems. In this work, we introduce TongSearch QR (Previously Known as "TongSearch Reasoner"), a family of small-scale language models for query reasoning and rewriting in reasoning-intensive retrieval. With a novel semi-rule-based reward function, we employ reinforcement learning approaches enabling smaller language models, e,g, Qwen2.5-7B-Instruct and Qwen2.5-1.5B-Instruct, to achieve query reasoning performance rivaling large-scale language models without their prohibitive inference costs. Experiment results on BRIGHT benchmark show that with BM25 as retrievers, both TongSearch QR-7B and TongSearch QR-1.5B models significantly outperform existing baselines, including prompt-based query reasoners and some latest dense retrievers trained for reasoning-intensive retrieval tasks, offering superior adaptability for real-world deployment.
Although modern vulnerability detection tools enable developers to efficiently identify numerous security flaws, indiscriminate remediation efforts often lead to superfluous development expenses. This is particularly true given that a substantial portion of detected vulnerabilities either possess low exploitability or would incur negligible impact in practical operational environments. Consequently, vulnerability severity assessment has emerged as a critical component in optimizing software development efficiency. Existing vulnerability assessment methods typically rely on manually crafted descriptions associated with source code artifacts. However, due to variability in description quality and subjectivity in intention interpretation, the performance of these methods is seriously limited. To address this issue, this paper introduces VulStamp, a novel intention-guided framework, to facilitate description-free vulnerability assessment. Specifically, VulStamp adopts static analysis together with Large Language Model (LLM) to extract the intention information of vulnerable code. Based on the intention information, VulStamp uses a prompt-tuned model for vulnerability assessment. Furthermore, to mitigate the problem of imbalanced data associated with vulnerability types, VulStamp integrates a Reinforcement Learning (RL)-based prompt-tuning method to train the assessment model.
Reinforcement learning (RL) has become a key technique for enhancing LLMs' reasoning abilities, yet its data inefficiency remains a major bottleneck. To address this critical yet challenging issue, we present a novel gradient-alignment-based method, named LearnAlign, which intelligently selects the learnable and representative training reasoning data for RL post-training. To overcome the well-known issue of response-length bias in gradient norms, we introduce the data learnability based on the success rate, which can indicate the learning potential of each data point. Experiments across three mathematical reasoning benchmarks demonstrate that our method significantly reduces training data requirements while achieving minor performance degradation or even improving performance compared to full-data training. For example, it reduces data requirements by up to 1,000 data points with better performance (77.53%) than that on the full dataset on GSM8K benchmark (77.04%). Furthermore, we show its effectiveness in the staged RL setting. This work provides valuable insights into data-efficient RL post-training and establishes a foundation for future research in optimizing reasoning data selection.To facilitate future work, we will release code.
Recent advances in reinforcement learning (RL) with verifiable outcome rewards have significantly improved the reasoning capabilities of large language models (LLMs), especially when combined with multi-turn tool interactions. However, existing methods lack both meaningful verification signals from realistic environments and explicit optimization for verification, leading to unreliable self-verification. To address these limitations, we propose ReVeal, a multi-turn reinforcement learning framework that interleaves code generation with explicit self-verification and tool-based evaluation. ReVeal enables LLMs to autonomously generate test cases, invoke external tools for precise feedback, and improves performance via a customized RL algorithm with dense, per-turn rewards. As a result, ReVeal fosters the co-evolution of a model's generation and verification capabilities through RL training, expanding the reasoning boundaries of the base model, demonstrated by significant gains in Pass@k on LiveCodeBench. It also enables test-time scaling into deeper inference regimes, with code consistently evolving as the number of turns increases during inference, ultimately surpassing DeepSeek-R1-Zero-Qwen-32B. These findings highlight the promise of ReVeal as a scalable and effective paradigm for building more robust and autonomous AI agents.
Reinforcement Learning from Verifiable Rewards (RLVR) has been widely adopted as the de facto method for enhancing the reasoning capabilities of large language models and has demonstrated notable success in verifiable domains like math and competitive programming tasks. However, the efficacy of RLVR diminishes significantly when applied to agentic environments. These settings, characterized by multi-step, complex problem solving, lead to high failure rates even for frontier LLMs, as the reward landscape is too sparse for effective model training via conventional RLVR. In this work, we introduce Agent-RLVR, a framework that makes RLVR effective in challenging agentic settings, with an initial focus on software engineering tasks. Inspired by human pedagogy, Agent-RLVR introduces agent guidance, a mechanism that actively steers the agent towards successful trajectories by leveraging diverse informational cues. These cues, ranging from high-level strategic plans to dynamic feedback on the agent's errors and environmental interactions, emulate a teacher's guidance, enabling the agent to navigate difficult solution spaces and promotes active self-improvement via additional environment exploration. In the Agent-RLVR training loop, agents first attempt to solve tasks to produce initial trajectories, which are then validated by unit tests and supplemented with agent guidance. Agents then reattempt with guidance, and the agent policy is updated with RLVR based on the rewards of these guided trajectories. Agent-RLVR elevates the pass@1 performance of Qwen-2.5-72B-Instruct from 9.4% to 22.4% on SWE-Bench Verified. We find that our guidance-augmented RLVR data is additionally useful for test-time reward model training, shown by further boosting pass@1 to 27.8%. Agent-RLVR lays the groundwork for training agents with RLVR in complex, real-world environments where conventional RL methods struggle.
Large language models (LLMs) are powerful but static; they lack mechanisms to adapt their weights in response to new tasks, knowledge, or examples. We introduce Self-Adapting LLMs (SEAL), a framework that enables LLMs to self-adapt by generating their own finetuning data and update directives. Given a new input, the model produces a self-edit-a generation that may restructure the information in different ways, specify optimization hyperparameters, or invoke tools for data augmentation and gradient-based updates. Through supervised finetuning (SFT), these self-edits result in persistent weight updates, enabling lasting adaptation. To train the model to produce effective self-edits, we use a reinforcement learning loop with the downstream performance of the updated model as the reward signal. Unlike prior approaches that rely on separate adaptation modules or auxiliary networks, SEAL directly uses the model's own generation to control its adaptation process. Experiments on knowledge incorporation and few-shot generalization show that SEAL is a promising step toward language models capable of self-directed adaptation. Our website and code is available at https://jyopari.github.io/posts/seal.
We introduce Magistral, Mistral's first reasoning model and our own scalable reinforcement learning (RL) pipeline. Instead of relying on existing implementations and RL traces distilled from prior models, we follow a ground up approach, relying solely on our own models and infrastructure. Notably, we demonstrate a stack that enabled us to explore the limits of pure RL training of LLMs, present a simple method to force the reasoning language of the model, and show that RL on text data alone maintains most of the initial checkpoint's capabilities. We find that RL on text maintains or improves multimodal understanding, instruction following and function calling. We present Magistral Medium, trained for reasoning on top of Mistral Medium 3 with RL alone, and we open-source Magistral Small (Apache 2.0) which further includes cold-start data from Magistral Medium.
To advance time series forecasting (TSF), various methods have been proposed to improve prediction accuracy, evolving from statistical techniques to data-driven deep learning architectures. Despite their effectiveness, most existing methods still adhere to a fast thinking paradigm-relying on extracting historical patterns and mapping them to future values as their core modeling philosophy, lacking an explicit thinking process that incorporates intermediate time series reasoning. Meanwhile, emerging slow-thinking LLMs (e.g., OpenAI-o1) have shown remarkable multi-step reasoning capabilities, offering an alternative way to overcome these issues. However, prompt engineering alone presents several limitations - including high computational cost, privacy risks, and limited capacity for in-depth domain-specific time series reasoning. To address these limitations, a more promising approach is to train LLMs to develop slow thinking capabilities and acquire strong time series reasoning skills. For this purpose, we propose Time-R1, a two-stage reinforcement fine-tuning framework designed to enhance multi-step reasoning ability of LLMs for time series forecasting. Specifically, the first stage conducts supervised fine-tuning for warmup adaptation, while the second stage employs reinforcement learning to improve the model's generalization ability. Particularly, we design a fine-grained multi-objective reward specifically for time series forecasting, and then introduce GRIP (group-based relative importance for policy optimization), which leverages non-uniform sampling to further encourage and optimize the model's exploration of effective reasoning paths. Experiments demonstrate that Time-R1 significantly improves forecast performance across diverse datasets.
Large language models (LLMs) have demonstrated significant advancements in reasoning capabilities, performing well on various challenging benchmarks. Techniques like Chain-of-Thought prompting have been introduced to further improve reasoning. However, these approaches frequently generate longer outputs, which in turn increase computational latency. Although some methods use reinforcement learning to shorten reasoning, they often apply uniform penalties without considering the problem's complexity, leading to suboptimal outcomes. In this study, we seek to enhance the efficiency of LLM reasoning by promoting conciseness for simpler problems while preserving sufficient reasoning for more complex ones for accuracy, thus improving the model's overall performance. Specifically, we manage the model's reasoning efficiency by dividing the reward function and including a novel penalty for output length. Our approach has yielded impressive outcomes in benchmark evaluations across three datasets: GSM8K, MATH500, and AIME2024. For the comparatively simpler datasets GSM8K and MATH500, our method has effectively shortened output lengths while preserving or enhancing accuracy. On the more demanding AIME2024 dataset, our approach has resulted in improved accuracy.
Large Language Models (LLMs) have demonstrated impressive capabilities in complex reasoning tasks, yet they still struggle to reliably verify the correctness of their own outputs. Existing solutions to this verification challenge often depend on separate verifier models or require multi-stage self-correction training pipelines, which limit scalability. In this paper, we propose Policy as Generative Verifier (PAG), a simple and effective framework that empowers LLMs to self-correct by alternating between policy and verifier roles within a unified multi-turn reinforcement learning (RL) paradigm. Distinct from prior approaches that always generate a second attempt regardless of model confidence, PAG introduces a selective revision mechanism: the model revises its answer only when its own generative verification step detects an error. This verify-then-revise workflow not only alleviates model collapse but also jointly enhances both reasoning and verification abilities. Extensive experiments across diverse reasoning benchmarks highlight PAG's dual advancements: as a policy, it enhances direct generation and self-correction accuracy; as a verifier, its self-verification outperforms self-consistency.
Story generation has been a prominent application of Large Language Models (LLMs). However, understanding LLMs' ability to produce high-quality stories remains limited due to challenges in automatic evaluation methods and the high cost and subjectivity of manual evaluation. Computational narratology offers valuable insights into what constitutes a good story, which has been applied in the symbolic narrative planning approach to story generation. This work aims to deepen the understanding of LLMs' story generation capabilities by using them to solve narrative planning problems. We present a benchmark for evaluating LLMs on narrative planning based on literature examples, focusing on causal soundness, character intentionality, and dramatic conflict. Our experiments show that GPT-4 tier LLMs can generate causally sound stories at small scales, but planning with character intentionality and dramatic conflict remains challenging, requiring LLMs trained with reinforcement learning for complex reasoning. The results offer insights on the scale of stories that LLMs can generate while maintaining quality from different aspects. Our findings also highlight interesting problem solving behaviors and shed lights on challenges and considerations for applying LLM narrative planning in game environments.
Reinforcement learning (RL) has shown great effectiveness for fine-tuning large language models (LLMs) using tasks that are challenging yet easily verifiable, such as math reasoning or code generation. However, extending this success to visual perception in vision-language models (VLMs) has been impeded by the scarcity of vision-centric tasks that are simultaneously challenging and unambiguously verifiable. To this end, we introduce ViCrit (Visual Caption Hallucination Critic), an RL proxy task that trains VLMs to localize a subtle, synthetic visual hallucination injected into paragraphs of human-written image captions. Starting from a 200-word captions, we inject a single, subtle visual description error-altering a few words on objects, attributes, counts, or spatial relations-and task the model to pinpoint the corrupted span given the image and the modified caption. This formulation preserves the full perceptual difficulty while providing a binary, exact-match reward that is easy to compute and unambiguous. Models trained with the ViCrit Task exhibit substantial gains across a variety of VL benchmarks. Crucially, the improvements transfer beyond natural-image training data to abstract image reasoning and visual math, showing promises of learning to perceive rather than barely memorizing seen objects. To facilitate evaluation, we further introduce ViCrit-Bench, a category-balanced diagnostic benchmark that systematically probes perception errors across diverse image domains and error types. Together, our results demonstrate that fine-grained hallucination criticism is an effective and generalizable objective for enhancing visual perception in VLMs.
Decompilers, which reconstruct human-readable source code from binary executables, are vital to many security tasks. Yet, despite recent advances, their output often suffers from syntactic and semantic errors and remains difficult to read. Recently, with the advent of large language models (LLMs), researchers began to explore the potential of LLMs to refine decompiler output. Nevertheless, our study of these approaches reveals significant limitations, such as introducing new errors and relying on unreliable accuracy validation. In this paper, we present D-LiFT, an automated decompiler backend that harnesses and further trains LLMs to improve the quality of decompiled code via reinforcement learning (RL). Unlike prior work that overlooks preserving accuracy, D-LiFT adheres to a key principle for enhancing the quality of decompiled code: \textit{preserving accuracy while improving readability}. Central to D-LiFT, we propose D-SCORE, an integrated quality assessment system to score the decompiled code from multiple aspects. In line with our principle, D-SCORE assigns low scores to any inaccurate output and only awards higher scores for readability to code that passes the accuracy check. Specifically, D-SCORE first verifies the syntactic and semantic correctness via the compiler and symbolic execution; only if a candidate is deemed accurate, it then evaluates readability using established metrics to compare the LLM output with the original decompiled code. The score will then be fed back to the LLM for fine-tuning. Our implementation, based on Ghidra and a range of LLMs, demonstrates significant improvements for the accurate decompiled code from the coreutils and util-linux projects. Compared to baseline LLMs without D-SCORE-driven fine-tuning, D-LiFT produces 55.3% more improved decompiled functions, as measured by D-SCORE.
Direct Preference Optimization (DPO) has become a cornerstone of reinforcement learning from human feedback (RLHF) due to its simplicity and efficiency. However, existing DPO-based approaches typically treat all preference pairs uniformly, ignoring critical variations in their inherent quality and learning utility, leading to suboptimal data utilization and performance. To address this challenge, we propose Omni-DPO, a dual-perspective optimization framework that jointly accounts for (1) the inherent quality of each preference pair and (2) the model's evolving performance on those pairs. By adaptively weighting samples according to both data quality and the model's learning dynamics during training, Omni-DPO enables more effective training data utilization and achieves better performance. Experimental results on various models and benchmarks demonstrate the superiority and generalization capabilities of Omni-DPO. On textual understanding tasks, Gemma-2-9b-it finetuned with Omni-DPO beats the leading LLM, Claude 3 Opus, by a significant margin of 6.7 points on the Arena-Hard benchmark. On mathematical reasoning tasks, Omni-DPO consistently outperforms the baseline methods across all benchmarks, providing strong empirical evidence for the effectiveness and robustness of our approach. Code and models will be available at https://github.com/pspdada/Omni-DPO.