Theorem proving serves as a major testbed for evaluating complex reasoning abilities in large language models (LLMs). However, traditional automated theorem proving (ATP) approaches rely heavily on formal proof systems that poorly align with LLMs' strength derived from informal, natural language knowledge acquired during pre-training. In this work, we propose DeepTheorem, a comprehensive informal theorem-proving framework exploiting natural language to enhance LLM mathematical reasoning. DeepTheorem includes a large-scale benchmark dataset consisting of 121K high-quality IMO-level informal theorems and proofs spanning diverse mathematical domains, rigorously annotated for correctness, difficulty, and topic categories, accompanied by systematically constructed verifiable theorem variants. We devise a novel reinforcement learning strategy (RL-Zero) explicitly tailored to informal theorem proving, leveraging the verified theorem variants to incentivize robust mathematical inference. Additionally, we propose comprehensive outcome and process evaluation metrics examining proof correctness and the quality of reasoning steps. Extensive experimental analyses demonstrate DeepTheorem significantly improves LLM theorem-proving performance compared to existing datasets and supervised fine-tuning protocols, achieving state-of-the-art accuracy and reasoning quality. Our findings highlight DeepTheorem's potential to fundamentally advance automated informal theorem proving and mathematical exploration.
The emergence of large language model (LLM)-based agents has significantly advanced the development of autonomous machine learning (ML) engineering. However, most existing approaches rely heavily on manual prompt engineering, failing to adapt and optimize based on diverse experimental experiences. Focusing on this, for the first time, we explore the paradigm of learning-based agentic ML, where an LLM agent learns through interactive experimentation on ML tasks using online reinforcement learning (RL). To realize this, we propose a novel agentic ML training framework with three key components: (1) exploration-enriched fine-tuning, which enables LLM agents to generate diverse actions for enhanced RL exploration; (2) step-wise RL, which enables training on a single action step, accelerating experience collection and improving training efficiency; (3) an agentic ML-specific reward module, which unifies varied ML feedback signals into consistent rewards for RL optimization. Leveraging this framework, we train ML-Agent, driven by a 7B-sized Qwen-2.5 LLM for autonomous ML. Remarkably, despite being trained on merely 9 ML tasks, our 7B-sized ML-Agent outperforms the 671B-sized DeepSeek-R1 agent. Furthermore, it achieves continuous performance improvements and demonstrates exceptional cross-task generalization capabilities.
Enhancing the mathematical reasoning capabilities of LLMs has garnered significant attention in both the mathematical and computer science communities. Recent works have made substantial progress in both Natural Language (NL) reasoning and Formal Language (FL) reasoning by leveraging the potential of pure Reinforcement Learning (RL) methods on base models. However, RL approaches struggle to impart new capabilities not presented in the base model, highlighting the need to integrate more knowledge like FL into NL math reasoning effectively. Yet, this integration is challenging due to inherent disparities in problem structure and reasoning format between NL and FL. To address these challenges, we introduce **NL-FL HybridReasoning**, an end-to-end framework designed to incorporate the FL expert into NL math problem-solving. To bridge the NL and FL input format gap, we propose the *NL-FL Problem Alignment* method, which reformulates the Question-Answering (QA) problems in NL as existence theorems in FL. Subsequently, the *Mixed Problem Input* technique we provide enables the FL reasoner to handle both QA and existence problems concurrently. Lastly, we mitigate the NL and FL output format gap in reasoning through an LLM-based *Answer Extraction* mechanism. Comprehensive experiments demonstrate that the **HybridReasoning** framework achieves **89.80%** and **84.34%** accuracy rates on the MATH-500 and the AMC benchmarks, surpassing the NL baseline by 4.60% and 4.82%, respectively. Notably, some problems resolved by our framework remain unsolved by the NL baseline model even under a larger number of trials.
Recent decoding methods improve the factuality of large language models~(LLMs) by refining how the next token is selected during generation. These methods typically operate at the token level, leveraging internal representations to suppress superficial patterns. Nevertheless, LLMs remain prone to hallucinations, especially over longer contexts. In this paper, we propose Active Layer-Contrastive Decoding (ActLCD), a novel decoding strategy that actively decides when to apply contrasting layers during generation. By casting decoding as a sequential decision-making problem, ActLCD employs a reinforcement learning policy guided by a reward-aware classifier to optimize factuality beyond the token level. Our experiments demonstrate that ActLCD surpasses state-of-the-art methods across five benchmarks, showcasing its effectiveness in mitigating hallucinations in diverse generation scenarios.
In this work, we present the first study to explore inference-time scaling on table reasoning tasks. We develop and evaluate two post-training strategies to enable inference-time scaling: distillation from frontier model reasoning traces and reinforcement learning with verifiable rewards (RLVR). For distillation, we introduce a large-scale dataset of reasoning traces generated by DeepSeek-R1, which we use to fine-tune LLMs into the Table-R1-SFT model. For RLVR, we propose task-specific verifiable reward functions and apply the GRPO algorithm to obtain the Table-R1-Zero model. We evaluate our Table-R1-series models across diverse table reasoning tasks, including short-form QA, fact verification, and free-form QA. Notably, the Table-R1-Zero model matches or exceeds the performance of GPT-4.1 and DeepSeek-R1, while using only a 7B-parameter LLM. It also demonstrates strong generalization to out-of-domain datasets. Extensive ablation and qualitative analyses reveal the benefits of instruction tuning, model architecture choices, and cross-task generalization, as well as emergence of essential table reasoning skills during RL training.
Unlocking deep, interpretable biological reasoning from complex genomic data is a major AI challenge hindering scientific discovery. Current DNA foundation models, despite strong sequence representation, struggle with multi-step reasoning and lack inherent transparent, biologically intuitive explanations. We introduce BioReason, a pioneering architecture that, for the first time, deeply integrates a DNA foundation model with a Large Language Model (LLM). This novel connection enables the LLM to directly process and reason with genomic information as a fundamental input, fostering a new form of multimodal biological understanding. BioReason's sophisticated multi-step reasoning is developed through supervised fine-tuning and targeted reinforcement learning, guiding the system to generate logical, biologically coherent deductions. On biological reasoning benchmarks including KEGG-based disease pathway prediction - where accuracy improves from 88% to 97% - and variant effect prediction, BioReason demonstrates an average 15% performance gain over strong single-modality baselines. BioReason reasons over unseen biological entities and articulates decision-making through interpretable, step-by-step biological traces, offering a transformative approach for AI in biology that enables deeper mechanistic insights and accelerates testable hypothesis generation from genomic data. Data, code, and checkpoints are publicly available at https://github.com/bowang-lab/BioReason
The reasoning capabilities of large language models (LLMs) have advanced rapidly, particularly following the release of DeepSeek R1, which has inspired a surge of research into data quality and reinforcement learning (RL) algorithms. Despite the pivotal role diversity plays in RL, its influence on LLM reasoning remains largely underexplored. To bridge this gap, this work presents a systematic investigation into the impact of diversity in RL-based training for LLM reasoning, and proposes a novel diversity-aware policy optimization method. Across evaluations on 12 LLMs, we observe a strong positive correlation between the solution diversity and Potential at k (a novel metric quantifying an LLM's reasoning potential) in high-performing models. This finding motivates our method to explicitly promote diversity during RL training. Specifically, we design a token-level diversity and reformulate it into a practical objective, then we selectively apply it to positive samples. Integrated into the R1-zero training framework, our method achieves a 3.5 percent average improvement across four mathematical reasoning benchmarks, while generating more diverse and robust solutions.
Large Language Models (LLMs) generate functionally correct solutions but often fall short in code efficiency, a critical bottleneck for real-world deployment. In this paper, we introduce a novel test-time iterative optimization framework to address this, employing a closed-loop system where LLMs iteratively refine code based on empirical performance feedback from an execution sandbox. We explore three training strategies: Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Group Relative Policy Optimization~(GRPO). Experiments on our Venus dataset and the APPS benchmark show that SFT and DPO rapidly saturate in efficiency gains. In contrast, GRPO, using reinforcement learning (RL) with execution feedback, continuously optimizes code performance, significantly boosting both pass@1 (from 47% to 62%) and the likelihood of outperforming human submissions in efficiency (from 31% to 45%). Our work demonstrates effective test-time code efficiency improvement and critically reveals the power of RL in teaching LLMs to truly self-improve code efficiency.
Process reward models (PRMs) provide more nuanced supervision compared to outcome reward models (ORMs) for optimizing policy models, positioning them as a promising approach to enhancing the capabilities of LLMs in complex reasoning tasks. Recent efforts have advanced PRMs from step-level to token-level granularity by integrating reward modeling into the training of generative models, with reward scores derived from token generation probabilities. However, the conflict between generative language modeling and reward modeling may introduce instability and lead to inaccurate credit assignments. To address this challenge, we revisit token-level reward assignment by decoupling reward modeling from language generation and derive a token-level reward model through the optimization of a discriminative policy, termed the Q-function Reward Model (Q-RM). We theoretically demonstrate that Q-RM explicitly learns token-level Q-functions from preference data without relying on fine-grained annotations. In our experiments, Q-RM consistently outperforms all baseline methods across various benchmarks. For example, when integrated into PPO/REINFORCE algorithms, Q-RM enhances the average Pass@1 score by 5.85/4.70 points on mathematical reasoning tasks compared to the ORM baseline, and by 4.56/5.73 points compared to the token-level PRM counterpart. Moreover, reinforcement learning with Q-RM significantly enhances training efficiency, achieving convergence 12 times faster than ORM on GSM8K and 11 times faster than step-level PRM on MATH. Code and data are available at https://github.com/homzer/Q-RM.
Rewards serve as proxies for human preferences and play a crucial role in Reinforcement Learning from Human Feedback (RLHF). However, if these rewards are inherently imperfect, exhibiting various biases, they can adversely affect the alignment of large language models (LLMs). In this paper, we collectively define the various biases present in rewards as the problem of reward unfairness. We propose a bias-agnostic method to address the issue of reward fairness from a resource allocation perspective, without specifically designing for each type of bias, yet effectively mitigating them. Specifically, we model preference learning as a resource allocation problem, treating rewards as resources to be allocated while considering the trade-off between utility and fairness in their distribution. We propose two methods, Fairness Regularization and Fairness Coefficient, to achieve fairness in rewards. We apply our methods in both verification and reinforcement learning scenarios to obtain a fairness reward model and a policy model, respectively. Experiments conducted in these scenarios demonstrate that our approach aligns LLMs with human preferences in a more fair manner.
The rapid development of multimodal AI and Large Language Models (LLMs) has greatly enhanced real-time interaction, decision-making, and collaborative tasks. However, in wireless multi-agent scenarios, limited bandwidth poses significant challenges to exchanging semantically rich multimodal information efficiently. Traditional semantic communication methods, though effective, struggle with redundancy and loss of crucial details. To overcome these challenges, we propose a Retrieval-Augmented Multimodal Semantic Communication (RAMSemCom) framework. RAMSemCom incorporates iterative, retrieval-driven semantic refinement tailored for distributed multi-agent environments, enabling efficient exchange of critical multimodal elements through local caching and selective transmission. Our approach dynamically optimizes retrieval using deep reinforcement learning (DRL) to balance semantic fidelity with bandwidth constraints. A comprehensive case study on multi-agent autonomous driving demonstrates that our DRL-based retrieval strategy significantly improves task completion efficiency and reduces communication overhead compared to baseline methods.
In next-generation wireless networks, supporting real-time applications such as augmented reality, autonomous driving, and immersive Metaverse services demands stringent constraints on bandwidth, latency, and reliability. Existing semantic communication (SemCom) approaches typically rely on static models, overlooking dynamic conditions and contextual cues vital for efficient transmission. To address these challenges, we propose CaSemCom, a context-aware SemCom framework that leverages a Large Language Model (LLM)-based gating mechanism and a Mixture of Experts (MoE) architecture to adaptively select and encode only high-impact semantic features across multiple data modalities. Our multimodal, multi-user case study demonstrates that CaSemCom significantly improves reconstructed image fidelity while reducing bandwidth usage, outperforming single-agent deep reinforcement learning (DRL) methods and traditional baselines in convergence speed, semantic accuracy, and retransmission overhead.
Reinforcement learning from human feedback (RLHF) is an essential technique for ensuring that large language models (LLMs) are aligned with human values and preferences during the post-training phase. As an effective RLHF approach, group relative policy optimization (GRPO) has demonstrated success in many LLM-based applications. However, efficient GRPO-based RLHF training remains a challenge. Recent studies reveal that a higher reward variance of the initial policy model leads to faster RLHF training. Inspired by this finding, we propose a practical reward adjustment model to accelerate RLHF training by provably increasing the reward variance and preserving the relative preferences and reward expectation. Our reward adjustment method inherently poses a nonconvex optimization problem, which is NP-hard to solve in general. To overcome the computational challenges, we design a novel $O(n \log n)$ algorithm to find a global solution of the nonconvex reward adjustment model by explicitly characterizing the extreme points of the feasible set. As an important application, we naturally integrate this reward adjustment model into the GRPO algorithm, leading to a more efficient GRPO with reward variance increase (GRPOVI) algorithm for RLHF training. As an interesting byproduct, we provide an indirect explanation for the empirical effectiveness of GRPO with rule-based reward for RLHF training, as demonstrated in DeepSeek-R1. Experiment results demonstrate that the GRPOVI algorithm can significantly improve the RLHF training efficiency compared to the original GRPO algorithm.
Multimodal Large Language Models (MLLMs) have demonstrated significant visual understanding capabilities, yet their fine-grained visual perception in complex real-world scenarios, such as densely crowded public areas, remains limited. Inspired by the recent success of reinforcement learning (RL) in both LLMs and MLLMs, in this paper, we explore how RL can enhance visual perception ability of MLLMs. Then we develop a novel RL-based framework, Deep Inspection and Perception with RL (DIP-R1) designed to enhance the visual perception capabilities of MLLMs, by comprehending complex scenes and looking through visual instances closely. DIP-R1 guides MLLMs through detailed inspection of visual scene via three simply designed rule-based reward modelings. First, we adopt a standard reasoning reward encouraging the model to include three step-by-step processes: 1) reasoning for understanding visual scenes, 2) observing for looking through interested but ambiguous regions, and 3) decision-making for predicting answer. Second, a variance-guided looking reward is designed to examine uncertain regions for the second observing process. It explicitly enables the model to inspect ambiguous areas, improving its ability to mitigate perceptual uncertainties. Third, we model a weighted precision-recall accuracy reward enhancing accurate decision-making. We explore its effectiveness across diverse fine-grained object detection data consisting of challenging real-world environments, such as densely crowded scenes. Built upon existing MLLMs, DIP-R1 achieves consistent and significant improvement across various in-domain and out-of-domain scenarios. It also outperforms various existing baseline models and supervised fine-tuning methods. Our findings highlight the substantial potential of integrating RL into MLLMs for enhancing capabilities in complex real-world perception tasks.
Recent advancements in large language models (LLMs) have demonstrated substantial progress in reasoning capabilities, such as DeepSeek-R1, which leverages rule-based reinforcement learning to enhance logical reasoning significantly. However, extending these achievements to multimodal large language models (MLLMs) presents critical challenges, which are frequently more pronounced for Multimodal Small Language Models (MSLMs) given their typically weaker foundational reasoning abilities: (1) the scarcity of high-quality multimodal reasoning datasets, (2) the degradation of reasoning capabilities due to the integration of visual processing, and (3) the risk that direct application of reinforcement learning may produce complex yet incorrect reasoning processes. To address these challenges, we design a novel framework Infi-MMR to systematically unlock the reasoning potential of MSLMs through a curriculum of three carefully structured phases and propose our multimodal reasoning model Infi-MMR-3B. The first phase, Foundational Reasoning Activation, leverages high-quality textual reasoning datasets to activate and strengthen the model's logical reasoning capabilities. The second phase, Cross-Modal Reasoning Adaptation, utilizes caption-augmented multimodal data to facilitate the progressive transfer of reasoning skills to multimodal contexts. The third phase, Multimodal Reasoning Enhancement, employs curated, caption-free multimodal data to mitigate linguistic biases and promote robust cross-modal reasoning. Infi-MMR-3B achieves both state-of-the-art multimodal math reasoning ability (43.68% on MathVerse testmini, 27.04% on MathVision test, and 21.33% on OlympiadBench) and general reasoning ability (67.2% on MathVista testmini).
Large language models (LLMs) have shown promising potential in persuasion, but existing works on training LLM persuaders are still preliminary. Notably, while humans are skilled in modeling their opponent's thoughts and opinions proactively and dynamically, current LLMs struggle with such Theory of Mind (ToM) reasoning, resulting in limited diversity and opponent awareness. To address this limitation, we introduce Theory of Mind Augmented Persuader (ToMAP), a novel approach for building more flexible persuader agents by incorporating two theory of mind modules that enhance the persuader's awareness and analysis of the opponent's mental state. Specifically, we begin by prompting the persuader to consider possible objections to the target central claim, and then use a text encoder paired with a trained MLP classifier to predict the opponent's current stance on these counterclaims. Our carefully designed reinforcement learning schema enables the persuader learns how to analyze opponent-related information and utilize it to generate more effective arguments. Experiments show that the ToMAP persuader, while containing only 3B parameters, outperforms much larger baselines, like GPT-4o, with a relative gain of 39.4% across multiple persuadee models and diverse corpora. Notably, ToMAP exhibits complex reasoning chains and reduced repetition during training, which leads to more diverse and effective arguments. The opponent-aware feature of ToMAP also makes it suitable for long conversations and enables it to employ more logical and opponent-aware strategies. These results underscore our method's effectiveness and highlight its potential for developing more persuasive language agents. Code is available at: https://github.com/ulab-uiuc/ToMAP.
Large language models (LLMs)-empowered web agents enables automating complex, real-time web navigation tasks in enterprise environments. However, existing web agents relying on supervised fine-tuning (SFT) often struggle with generalization and robustness due to insufficient reasoning capabilities when handling the inherently dynamic nature of web interactions. In this study, we introduce WorkForceAgent-R1, an LLM-based web agent trained using a rule-based R1-style reinforcement learning framework designed explicitly to enhance single-step reasoning and planning for business-oriented web navigation tasks. We employ a structured reward function that evaluates both adherence to output formats and correctness of actions, enabling WorkForceAgent-R1 to implicitly learn robust intermediate reasoning without explicit annotations or extensive expert demonstrations. Extensive experiments on the WorkArena benchmark demonstrate that WorkForceAgent-R1 substantially outperforms SFT baselines by 10.26-16.59%, achieving competitive performance relative to proprietary LLM-based agents (gpt-4o) in workplace-oriented web navigation tasks.
Systematic reviews in medicine play a critical role in evidence-based decision-making by aggregating findings from multiple studies. A central bottleneck in automating this process is extracting numeric evidence and determining study-level conclusions for specific outcomes and comparisons. Prior work has framed this problem as a textual inference task by retrieving relevant content fragments and inferring conclusions from them. However, such approaches often rely on shallow textual cues and fail to capture the underlying numeric reasoning behind expert assessments. In this work, we conceptualise the problem as one of quantitative reasoning. Rather than inferring conclusions from surface text, we extract structured numerical evidence (e.g., event counts or standard deviations) and apply domain knowledge informed logic to derive outcome-specific conclusions. We develop a numeric reasoning system composed of a numeric data extraction model and an effect estimate component, enabling more accurate and interpretable inference aligned with the domain expert principles. We train the numeric data extraction model using different strategies, including supervised fine-tuning (SFT) and reinforcement learning (RL) with a new value reward model. When evaluated on the CochraneForest benchmark, our best-performing approach -- using RL to train a small-scale number extraction model -- yields up to a 21% absolute improvement in F1 score over retrieval-based systems and outperforms general-purpose LLMs of over 400B parameters by up to 9%. Our results demonstrate the promise of reasoning-driven approaches for automating systematic evidence synthesis.
Computer-Aided Design (CAD) plays a central role in engineering and manufacturing, making it possible to create precise and editable 3D models. Using a variety of sensor or user-provided data as inputs for CAD reconstruction can democratize access to design applications. However, existing methods typically focus on a single input modality, such as point clouds, images, or text, which limits their generalizability and robustness. Leveraging recent advances in vision-language models (VLM), we propose a multi-modal CAD reconstruction model that simultaneously processes all three input modalities. Inspired by large language model (LLM) training paradigms, we adopt a two-stage pipeline: supervised fine-tuning (SFT) on large-scale procedurally generated data, followed by reinforcement learning (RL) fine-tuning using online feedback, obtained programatically. Furthermore, we are the first to explore RL fine-tuning of LLMs for CAD tasks demonstrating that online RL algorithms such as Group Relative Preference Optimization (GRPO) outperform offline alternatives. In the DeepCAD benchmark, our SFT model outperforms existing single-modal approaches in all three input modalities simultaneously. More importantly, after RL fine-tuning, cadrille sets new state-of-the-art on three challenging datasets, including a real-world one.
Mathematical reasoning tasks have become prominent benchmarks for assessing the reasoning capabilities of LLMs, especially with reinforcement learning (RL) methods such as GRPO showing significant performance gains. However, accuracy metrics alone do not support fine-grained assessment of capabilities and fail to reveal which problem-solving skills have been internalized. To better understand these capabilities, we propose to decompose problem solving into fundamental capabilities: Plan (mapping questions to sequences of steps), Execute (correctly performing solution steps), and Verify (identifying the correctness of a solution). Empirically, we find that GRPO mainly enhances the execution skill-improving execution robustness on problems the model already knows how to solve-a phenomenon we call temperature distillation. More importantly, we show that RL-trained models struggle with fundamentally new problems, hitting a 'coverage wall' due to insufficient planning skills. To explore RL's impact more deeply, we construct a minimal, synthetic solution-tree navigation task as an analogy for mathematical problem-solving. This controlled setup replicates our empirical findings, confirming RL primarily boosts execution robustness. Importantly, in this setting, we identify conditions under which RL can potentially overcome the coverage wall through improved exploration and generalization to new solution paths. Our findings provide insights into the role of RL in enhancing LLM reasoning, expose key limitations, and suggest a path toward overcoming these barriers. Code is available at https://github.com/cfpark00/RL-Wall.
Recent studies on post-training large language models (LLMs) for reasoning through reinforcement learning (RL) typically focus on tasks that can be accurately verified and rewarded, such as solving math problems. In contrast, our research investigates the impact of reward noise, a more practical consideration for real-world scenarios involving the post-training of LLMs using reward models. We found that LLMs demonstrate strong robustness to substantial reward noise. For example, manually flipping 40% of the reward function's outputs in math tasks still allows a Qwen-2.5-7B model to achieve rapid convergence, improving its performance on math tasks from 5% to 72%, compared to the 75% accuracy achieved by a model trained with noiseless rewards. Surprisingly, by only rewarding the appearance of key reasoning phrases (namely reasoning pattern reward, RPR), such as ``first, I need to''-without verifying the correctness of answers, the model achieved peak downstream performance (over 70% accuracy for Qwen-2.5-7B) comparable to models trained with strict correctness verification and accurate rewards. Recognizing the importance of the reasoning process over the final results, we combined RPR with noisy reward models. RPR helped calibrate the noisy reward models, mitigating potential false negatives and enhancing the LLM's performance on open-ended tasks. These findings suggest the importance of improving models' foundational abilities during the pre-training phase while providing insights for advancing post-training techniques. Our code and scripts are available at https://github.com/trestad/Noisy-Rewards-in-Learning-to-Reason.
Code generation with large language models (LLMs), often termed vibe coding, is increasingly adopted in production but fails to ensure code quality, particularly in security (e.g., SQL injection vulnerabilities) and maintainability (e.g., missing type annotations). Existing methods, such as supervised fine-tuning and rule-based post-processing, rely on labor-intensive annotations or brittle heuristics, limiting their scalability and effectiveness. We propose REAL, a reinforcement learning framework that incentivizes LLMs to generate production-quality code using program analysis-guided feedback. Specifically, REAL integrates two automated signals: (1) program analysis detecting security or maintainability defects and (2) unit tests ensuring functional correctness. Unlike prior work, our framework is prompt-agnostic and reference-free, enabling scalable supervision without manual intervention. Experiments across multiple datasets and model scales demonstrate that REAL outperforms state-of-the-art methods in simultaneous assessments of functionality and code quality. Our work bridges the gap between rapid prototyping and production-ready code, enabling LLMs to deliver both speed and quality.
This paper aims to overcome a major obstacle in scaling RL for reasoning with LLMs, namely the collapse of policy entropy. Such phenomenon is consistently observed across vast RL runs without entropy intervention, where the policy entropy dropped sharply at the early training stage, this diminished exploratory ability is always accompanied with the saturation of policy performance. In practice, we establish a transformation equation R=-a*e^H+b between entropy H and downstream performance R. This empirical law strongly indicates that, the policy performance is traded from policy entropy, thus bottlenecked by its exhaustion, and the ceiling is fully predictable H=0, R=-a+b. Our finding necessitates entropy management for continuous exploration toward scaling compute for RL. To this end, we investigate entropy dynamics both theoretically and empirically. Our derivation highlights that, the change in policy entropy is driven by the covariance between action probability and the change in logits, which is proportional to its advantage when using Policy Gradient-like algorithms. Empirical study shows that, the values of covariance term and entropy differences matched exactly, supporting the theoretical conclusion. Moreover, the covariance term stays mostly positive throughout training, further explaining why policy entropy would decrease monotonically. Through understanding the mechanism behind entropy dynamics, we motivate to control entropy by restricting the update of high-covariance tokens. Specifically, we propose two simple yet effective techniques, namely Clip-Cov and KL-Cov, which clip and apply KL penalty to tokens with high covariances respectively. Experiments show that these methods encourage exploration, thus helping policy escape entropy collapse and achieve better downstream performance.
Emotion understanding includes basic tasks (e.g., sentiment/emotion classification) and advanced tasks (e.g., sarcasm/humor detection). Current methods rely on fixed-length CoT reasoning, failing to adapt to the varying complexity of emotions. We propose a task-adaptive reasoning framework that employs DeepSeek-R1 to generate variable-length reasoning chains for different emotion tasks. By combining fine-tuning with reinforcement learning, we design a composite reward function that balances four objectives: prediction accuracy, adaptive reasoning depth control, structural diversity in reasoning paths, and suppression of repetitive logic. This approach achieves dynamic context-sensitive inference while enabling LLMs to autonomously develop deep reasoning capabilities. Experimental results demonstrate consistent improvements in both Acc and F1 scores across four tasks: emotion, sentiment, humor, and sarcasm. Notably, peak enhancements reached 3.56% F1 (2.76% Acc) for basic tasks and 37.95% F1 (23.14% Acc) for advanced tasks. Our work bridges rigid CoT reasoning and emotional complexity through adaptive-depth analysis.
Large Language Model-based Multi-Agent Systems (MASs) have emerged as a powerful paradigm for tackling complex tasks through collaborative intelligence. Nevertheless, the question of how agents should be structurally organized for optimal cooperation remains largely unexplored. In this position paper, we aim to gently redirect the focus of the MAS research community toward this critical dimension: develop topology-aware MASs for specific tasks. Specifically, the system consists of three core components - agents, communication links, and communication patterns - that collectively shape its coordination performance and efficiency. To this end, we introduce a systematic, three-stage framework: agent selection, structure profiling, and topology synthesis. Each stage would trigger new research opportunities in areas such as language models, reinforcement learning, graph learning, and generative modeling; together, they could unleash the full potential of MASs in complicated real-world applications. Then, we discuss the potential challenges and opportunities in the evaluation of multiple systems. We hope our perspective and framework can offer critical new insights in the era of agentic AI.
Improving Multi-modal Large Language Models (MLLMs) in the post-training stage typically relies on supervised fine-tuning (SFT) or reinforcement learning (RL). However, these supervised methods require expensive and manually annotated multi-modal data--an ultimately unsustainable resource. While recent efforts have explored unsupervised post-training, their methods are complex and difficult to iterate. In this work, we are the first to investigate the use of GRPO, a stable and scalable online RL algorithm, for enabling continual self-improvement without any external supervision. We propose MM-UPT, a simple yet effective framework for unsupervised post-training of MLLMs. MM-UPT builds upon GRPO, replacing traditional reward signals with a self-rewarding mechanism based on majority voting over multiple sampled responses. Our experiments demonstrate that MM-UPT significantly improves the reasoning ability of Qwen2.5-VL-7B (e.g., 66.3 %$\rightarrow$72.9 % on MathVista, 62.9 %$\rightarrow$68.7 % on We-Math), using standard dataset without ground truth labels. MM-UPT also outperforms prior unsupervised baselines and even approaches the results of supervised GRPO. Furthermore, we show that incorporating synthetic questions, generated solely by MLLM itself, can boost performance as well, highlighting a promising approach for scalable self-improvement. Overall, MM-UPT offers a new paradigm for continual, autonomous enhancement of MLLMs in the absence of external supervision. Our code is available at https://github.com/waltonfuture/MM-UPT.
Robust evaluation is critical for deploying trustworthy retrieval-augmented generation (RAG) systems. However, current LLM-based evaluation frameworks predominantly rely on directly prompting resource-intensive models with complex multi-stage prompts, underutilizing models' reasoning capabilities and introducing significant computational cost. In this paper, we present RAG-Zeval (RAG-Zero Evaluator), a novel end-to-end framework that formulates faithfulness and correctness evaluation as a rule-guided reasoning task. Our approach trains evaluators with reinforcement learning, facilitating compact models to generate comprehensive and sound assessments with detailed explanation in one-pass. We introduce a ranking-based outcome reward mechanism, using preference judgments rather than absolute scores, to address the challenge of obtaining precise pointwise reward signals. To this end, we synthesize the ranking references by generating quality-controlled responses with zero human annotation. Experiments demonstrate RAG-Zeval's superior performance, achieving the strongest correlation with human judgments and outperforming baselines that rely on LLMs with 10-100 times more parameters. Our approach also exhibits superior interpretability in response evaluation.
Diffusion models have recently demonstrated exceptional performance in image generation task. However, existing image generation methods still significantly suffer from the dilemma of image reasoning, especially in logic-centered image generation tasks. Inspired by the success of Chain of Thought (CoT) and Reinforcement Learning (RL) in LLMs, we propose SRRL, a self-reflective RL algorithm for diffusion models to achieve reasoning generation of logical images by performing reflection and iteration across generation trajectories. The intermediate samples in the denoising process carry noise, making accurate reward evaluation difficult. To address this challenge, SRRL treats the entire denoising trajectory as a CoT step with multi-round reflective denoising process and introduces condition guided forward process, which allows for reflective iteration between CoT steps. Through SRRL-based iterative diffusion training, we introduce image reasoning through CoT into generation tasks adhering to physical laws and unconventional physical phenomena for the first time. Notably, experimental results of case study exhibit that the superior performance of our SRRL algorithm even compared with GPT-4o. The project page is https://jadenpan0.github.io/srrl.github.io/.
This work presents Pangu Embedded, an efficient Large Language Model (LLM) reasoner developed on Ascend Neural Processing Units (NPUs), featuring flexible fast and slow thinking capabilities. Pangu Embedded addresses the significant computational costs and inference latency challenges prevalent in existing reasoning-optimized LLMs. We propose a two-stage training framework for its construction. In Stage 1, the model is finetuned via an iterative distillation process, incorporating inter-iteration model merging to effectively aggregate complementary knowledge. This is followed by reinforcement learning on Ascend clusters, optimized by a latency-tolerant scheduler that combines stale synchronous parallelism with prioritized data queues. The RL process is guided by a Multi-source Adaptive Reward System (MARS), which generates dynamic, task-specific reward signals using deterministic metrics and lightweight LLM evaluators for mathematics, coding, and general problem-solving tasks. Stage 2 introduces a dual-system framework, endowing Pangu Embedded with a "fast" mode for routine queries and a deeper "slow" mode for complex inference. This framework offers both manual mode switching for user control and an automatic, complexity-aware mode selection mechanism that dynamically allocates computational resources to balance latency and reasoning depth. Experimental results on benchmarks including AIME 2024, GPQA, and LiveCodeBench demonstrate that Pangu Embedded with 7B parameters, outperforms similar-size models like Qwen3-8B and GLM4-9B. It delivers rapid responses and state-of-the-art reasoning quality within a single, unified model architecture, highlighting a promising direction for developing powerful yet practically deployable LLM reasoners.
Recent advancements in large language models (LLMs) have demonstrated impressive chain-of-thought reasoning capabilities, with reinforcement learning (RL) playing a crucial role in this progress. While "aha moment" patterns--where models exhibit self-correction through reflection--are often attributed to emergent properties from RL, we first demonstrate that these patterns exist in multimodal LLMs (MLLMs) prior to RL training but may not necessarily correlate with improved reasoning performance. Building on these insights, we present a comprehensive study on enhancing multimodal reasoning through a two-stage approach: (1) supervised fine-tuning (SFT) as a cold start with structured chain-of-thought reasoning patterns, followed by (2) reinforcement learning via GRPO to further refine these capabilities. Our extensive experiments show that this combined approach consistently outperforms both SFT-only and RL-only methods across challenging multimodal reasoning benchmarks. The resulting models achieve state-of-the-art performance among open-source MLLMs at both 3B and 7B scales, with our 7B model showing substantial improvements over base models (e.g., 66.3 %$\rightarrow$73.4 % on MathVista, 62.9 %$\rightarrow$70.4 % on We-Math) and our 3B model achieving performance competitive with several 7B models. Overall, this work provides practical guidance for building advanced multimodal reasoning models. Our code is available at https://github.com/waltonfuture/RL-with-Cold-Start.