Large Language Models (LLMs) often struggle with mathematical reasoning tasks requiring precise, verifiable computation. While Reinforcement Learning (RL) from outcome-based rewards enhances text-based reasoning, understanding how agents autonomously learn to leverage external tools like code execution remains crucial. We investigate RL from outcome-based rewards for Tool-Integrated Reasoning, ZeroTIR, training base LLMs to spontaneously generate and execute Python code for mathematical problems without supervised tool-use examples. Our central contribution is we demonstrate that as RL training progresses, key metrics scale predictably. Specifically, we observe strong positive correlations where increased training steps lead to increases in the spontaneous code execution frequency, the average response length, and, critically, the final task accuracy. This suggests a quantifiable relationship between computational effort invested in training and the emergence of effective, tool-augmented reasoning strategies. We implement a robust framework featuring a decoupled code execution environment and validate our findings across standard RL algorithms and frameworks. Experiments show ZeroTIR significantly surpasses non-tool ZeroRL baselines on challenging math benchmarks. Our findings provide a foundational understanding of how autonomous tool use is acquired and scales within Agent RL, offering a reproducible benchmark for future studies. Code is released at \href{https://github.com/yyht/openrlhf_async_pipline}{https://github.com/yyht/openrlhf\_async\_pipline}.
As Test-Time Scaling emerges as an active research focus in the large language model community, advanced post-training methods increasingly emphasize extending chain-of-thought (CoT) generation length, thereby enhancing reasoning capabilities to approach Deepseek R1-like reasoning models. However, recent studies reveal that reasoning models (even Qwen3) consistently exhibit excessive thought redundancy in CoT generation. This overthinking problem stems from conventional outcome-reward reinforcement learning's systematic neglect in regulating intermediate reasoning steps. This paper proposes Serial-Group Decaying-Reward Policy Optimization (namely S-GRPO), a novel reinforcement learning method that empowers models with the capability to determine the sufficiency of reasoning steps, subsequently triggering early exit of CoT generation. Specifically, unlike GRPO, which samples multiple possible completions (parallel group) in parallel, we select multiple temporal positions in the generation of one CoT to allow the model to exit thinking and instead generate answers (serial group), respectively. For the correct answers in a serial group, we assign rewards that decay according to positions, with lower rewards towards the later ones, thereby reinforcing the model's behavior to generate higher-quality answers at earlier phases with earlier exits of thinking. Empirical evaluations demonstrate compatibility with state-of-the-art reasoning models, including Qwen3 and Deepseek-distill models, achieving 35.4% ~ 61.1\% sequence length reduction with 0.72% ~ 6.08% accuracy improvements across GSM8K, AIME 2024, AMC 2023, MATH-500, and GPQA Diamond benchmarks.
Data driven discovery of partial differential equations (PDEs) is a promising approach for uncovering the underlying laws governing complex systems. However, purely data driven techniques face the dilemma of balancing search space with optimization efficiency. This study introduces a knowledge guided approach that incorporates existing PDEs documented in a mathematical handbook to facilitate the discovery process. These PDEs are encoded as sentence like structures composed of operators and basic terms, and used to train a generative model, called EqGPT, which enables the generation of free form PDEs. A loop of generation evaluation optimization is constructed to autonomously identify the most suitable PDE. Experimental results demonstrate that this framework can recover a variety of PDE forms with high accuracy and computational efficiency, particularly in cases involving complex temporal derivatives or intricate spatial terms, which are often beyond the reach of conventional methods. The approach also exhibits generalizability to irregular spatial domains and higher dimensional settings. Notably, it succeeds in discovering a previously unreported PDE governing strongly nonlinear surface gravity waves propagating toward breaking, based on real world experimental data, highlighting its applicability to practical scenarios and its potential to support scientific discovery.
Achieving both accuracy and diverse reasoning remains challenging for Large Language Models (LLMs) in complex domains like mathematics. A key bottleneck is evaluating intermediate reasoning steps to guide generation without costly human annotations. To address this, we first introduce a novel Process Reward Model (PRM) trained automatically using Monte Carlo Tree Search coupled with a similarity-based data augmentation technique, effectively capturing step-level reasoning quality. Leveraging this PRM, we then adapt Generative Flow Networks (GFlowNets) to operate at the reasoning step level. Unlike traditional reinforcement learning focused on maximizing a single reward, GFlowNets naturally sample diverse, high-quality solutions proportional to their rewards, as measured by our PRM. Empirical evaluation shows strong improvements in both accuracy and solution diversity on challenging mathematical benchmarks (e.g., +2.59% absolute accuracy on MATH Level 5 for Llama3.2-3B), with effective generalization to unseen datasets (+9.4% absolute on SAT MATH). Our work demonstrates the potential of PRM-guided, step-level GFlowNets for developing more robust and versatile mathematical reasoning in LLMs.
Prevalent reinforcement learning~(RL) methods for fine-tuning LLM reasoners, such as GRPO or Leave-one-out PPO, abandon the learned value function in favor of empirically estimated returns. This hinders test-time compute scaling that relies on using the value-function for verification. In this work, we propose RL$^V$ that augments any ``value-free'' RL method by jointly training the LLM as both a reasoner and a generative verifier using RL-generated data, adding verification capabilities without significant overhead. Empirically, RL$^V$ boosts MATH accuracy by over 20\% with parallel sampling and enables $8-32\times$ efficient test-time compute scaling compared to the base RL method. RL$^V$ also exhibits strong generalization capabilities for both easy-to-hard and out-of-domain tasks. Furthermore, RL$^V$ achieves $1.2-1.6\times$ higher performance when jointly scaling parallel and sequential test-time compute with a long reasoning R1 model.
Large reasoning language models such as OpenAI-o1 and Deepseek-R1 have recently attracted widespread attention due to their impressive task-solving abilities. However, the enormous model size and the generation of lengthy thought chains introduce significant reasoning costs and response latency. Existing methods for efficient reasoning mainly focus on reducing the number of model parameters or shortening the chain-of-thought length. In this paper, we introduce Speculative Chain-of-Thought (SCoT), which reduces reasoning latency from another perspective by accelerated average reasoning speed through large and small model collaboration. SCoT conducts thought-level drafting using a lightweight draft model. Then it selects the best CoT draft and corrects the error cases with the target model. The proposed thinking behavior alignment improves the efficiency of drafting and the draft selection strategy maintains the prediction accuracy for complex problems. Experimental results on GSM8K, MATH, GaoKao, CollegeMath and Olympiad datasets show that SCoT reduces reasoning latency by 48\%$\sim$66\% for Deepseek-R1-Distill-Qwen-32B while achieving near-target-model-level performance. Our code is available at https://github.com/Jikai0Wang/Speculative_CoT.
Recent advancements in Large Vision-Language Models (LVLMs) have significantly enhanced their ability to integrate visual and linguistic information, achieving near-human proficiency in tasks like object recognition, captioning, and visual question answering. However, current benchmarks typically focus on knowledge-centric evaluations that assess domain-specific expertise, often neglecting the core ability to reason about fundamental mathematical elements and visual concepts. We identify a gap in evaluating elementary-level math problems, which rely on explicit visual dependencies-requiring models to discern, integrate, and reason across multiple images while incorporating commonsense knowledge, all of which are crucial for advancing toward broader AGI capabilities. To address this gap, we introduce VCBENCH, a comprehensive benchmark for multimodal mathematical reasoning with explicit visual dependencies. VCBENCH includes 1,720 problems across six cognitive domains, featuring 6,697 images (averaging 3.9 per question) to ensure multi-image reasoning. We evaluate 26 state-of-the-art LVLMs on VCBENCH, revealing substantial performance disparities, with even the top models unable to exceed 50% accuracy. Our findings highlight the ongoing challenges in visual-mathematical integration and suggest avenues for future LVLM advancements. The project can be found at https://alibaba-damo-academy.github.io/VCBench/.
This paper presents our winning submission to the AI Mathematical Olympiad - Progress Prize 2 (AIMO-2) competition. Our recipe for building state-of-the-art mathematical reasoning models relies on three key pillars. First, we create a large-scale dataset comprising 540K unique high-quality math problems, including olympiad-level problems, and their 3.2M long-reasoning solutions. Second, we develop a novel method to integrate code execution with long reasoning models through iterative training, generation, and quality filtering, resulting in 1.7M high-quality Tool-Integrated Reasoning solutions. Third, we create a pipeline to train models to select the most promising solution from many candidates. We show that such generative solution selection (GenSelect) can significantly improve upon majority voting baseline. Combining these ideas, we train a series of models that achieve state-of-the-art results on mathematical reasoning benchmarks. To facilitate further research, we release our code, models, and the complete OpenMathReasoning dataset under a commercially permissive license.
Reasoning in large language models (LLMs) tends to produce substantially longer token generation sequences than simpler language modeling tasks. This extended generation length reflects the multi-step, compositional nature of reasoning and is often correlated with higher solution accuracy. From an efficiency perspective, longer token generation exacerbates the inherently sequential and memory-bound decoding phase of LLMs. However, not all parts of this expensive reasoning process are equally difficult to generate. We leverage this observation by offloading only the most challenging parts of the reasoning process to a larger, more capable model, while performing most of the generation with a smaller, more efficient model; furthermore, we teach the smaller model to identify these difficult segments and independently trigger offloading when needed. To enable this behavior, we annotate difficult segments across 18k reasoning traces from the OpenR1-Math-220k chain-of-thought (CoT) dataset. We then apply supervised fine-tuning (SFT) and reinforcement learning fine-tuning (RLFT) to a 1.5B-parameter reasoning model, training it to learn to offload the most challenging parts of its own reasoning process to a larger model. This approach improves AIME24 reasoning accuracy by 24% and 28.3% while offloading 1.35% and 5% of the generated tokens respectively. We open-source our SplitReason model, data, code and logs.
Recent advances in large reasoning language models (LRLMs) rely on test-time scaling, which extends long chain-of-thought (CoT) generation to solve complex tasks. However, overthinking in long CoT not only slows down the efficiency of problem solving, but also risks accuracy loss due to the extremely detailed or redundant reasoning steps. We propose a simple yet effective method that allows LLMs to self-truncate CoT sequences by early exit during generation. Instead of relying on fixed heuristics, the proposed method monitors model behavior at potential reasoning transition points (e.g.,"Wait" tokens) and dynamically terminates the next reasoning chain's generation when the model exhibits high confidence in a trial answer. Our method requires no additional training and can be seamlessly integrated into existing o1-like reasoning LLMs. Experiments on 10 reasoning benchmarks (e.g., GSM8K, MATH-500, AMC, GPQA, AIME and LiveCodeBench) show that the proposed method is consistently effective on 11 cutting-edge reasoning LLMs of varying series and sizes, reducing the length of CoT sequences by an average of 19.1% to 80.1% while improving accuracy by 0.3% to 5.0%.
Recent reasoning models through test-time scaling have demonstrated that long chain-of-thoughts can unlock substantial performance boosts in hard reasoning tasks such as math and code. However, the benefit of such long thoughts for system-2 reasoning is relatively less explored in other domains such as perceptual tasks where shallower, system-1 reasoning seems sufficient. In this paper, we introduce LongPerceptualThoughts, a new synthetic dataset with 30K long-thought traces for perceptual tasks. The key challenges in synthesizing elaborate reasoning thoughts for perceptual tasks are that off-the-shelf models are not yet equipped with such thinking behavior and that it is not straightforward to build a reliable process verifier for perceptual tasks. Thus, we propose a novel three-stage data synthesis framework that first synthesizes verifiable multiple-choice questions from dense image descriptions, then extracts simple CoTs from VLMs for those verifiable problems, and finally expands those simple thoughts to elaborate long thoughts via frontier reasoning models. In controlled experiments with a strong instruction-tuned 7B model, we demonstrate notable improvements over existing visual reasoning data-generation methods. Our model, trained on the generated dataset, achieves an average +3.4 points improvement over 5 vision-centric benchmarks, including +11.8 points on V$^*$ Bench. Notably, despite being tuned for vision tasks, it also improves performance on the text reasoning benchmark, MMLU-Pro, by +2 points.
Multimodal reasoning, which integrates language and visual cues into problem solving and decision making, is a fundamental aspect of human intelligence and a crucial step toward artificial general intelligence. However, the evaluation of multimodal reasoning capabilities in Multimodal Large Language Models (MLLMs) remains inadequate. Most existing reasoning benchmarks are constrained by limited data size, narrow domain coverage, and unstructured knowledge distribution. To close these gaps, we introduce MDK12-Bench, a multi-disciplinary benchmark assessing the reasoning capabilities of MLLMs via real-world K-12 examinations. Spanning six disciplines (math, physics, chemistry, biology, geography, and information science), our benchmark comprises 140K reasoning instances across diverse difficulty levels from primary school to 12th grade. It features 6,827 instance-level knowledge point annotations based on a well-organized knowledge structure, detailed answer explanations, difficulty labels and cross-year partitions, providing a robust platform for comprehensive evaluation. Additionally, we present a novel dynamic evaluation framework to mitigate data contamination issues by bootstrapping question forms, question types, and image styles during evaluation. Extensive experiment on MDK12-Bench reveals the significant limitation of current MLLMs in multimodal reasoning. The findings on our benchmark provide insights into the development of the next-generation models. Our data and codes are available at https://github.com/LanceZPF/MDK12.
Despite the recent success of large language models (LLMs) in reasoning such as DeepSeek, we for the first time identify a key dilemma in reasoning robustness and generalization: significant performance degradation on novel or incomplete data, suggesting a reliance on memorized patterns rather than systematic reasoning. Our closer examination reveals four key unique limitations underlying this issue:(1) Positional bias--models favor earlier queries in multi-query inputs but answering the wrong one in the latter (e.g., GPT-4o's accuracy drops from 75.8 percent to 72.8 percent); (2) Instruction sensitivity--performance declines by 5.0 to 7.5 percent in the Qwen2.5 Series and by 5.0 percent in DeepSeek-V3 with auxiliary guidance; (3) Numerical fragility--value substitution sharply reduces accuracy (e.g., GPT-4o drops from 97.5 percent to 82.5 percent, GPT-o1-mini drops from 97.5 percent to 92.5 percent); and (4) Memory dependence--models resort to guesswork when missing critical data. These findings further highlight the reliance on heuristic recall over rigorous logical inference, demonstrating challenges in reasoning robustness. To comprehensively investigate these robustness challenges, this paper introduces a novel benchmark, termed as Math-RoB, that exploits hallucinations triggered by missing information to expose reasoning gaps. This is achieved by an instruction-based approach to generate diverse datasets that closely resemble training distributions, facilitating a holistic robustness assessment and advancing the development of more robust reasoning frameworks. Bad character(s) in field Abstract.
Step-by-step verifiers -- also known as process reward models (PRMs) -- are a key ingredient for test-time scaling. PRMs require step-level supervision, making them expensive to train. This work aims to build data-efficient PRMs as verbalized step-wise reward models that verify every step in the solution by generating a verification chain-of-thought (CoT). We propose ThinkPRM, a long CoT verifier fine-tuned on orders of magnitude fewer process labels than those required by discriminative PRMs. Our approach capitalizes on the inherent reasoning abilities of long CoT models, and outperforms LLM-as-a-Judge and discriminative verifiers -- using only 1% of the process labels in PRM800K -- across several challenging benchmarks. Specifically, ThinkPRM beats the baselines on ProcessBench, MATH-500, and AIME '24 under best-of-N selection and reward-guided search. In an out-of-domain evaluation on a subset of GPQA-Diamond and LiveCodeBench, our PRM surpasses discriminative verifiers trained on the full PRM800K by 8% and 4.5%, respectively. Lastly, under the same token budget, ThinkPRM scales up verification compute more effectively compared to LLM-as-a-Judge, outperforming it by 7.2% on a subset of ProcessBench. Our work highlights the value of generative, long CoT PRMs that can scale test-time compute for verification while requiring minimal supervision for training. Our code, data, and models will be released at https://github.com/mukhal/thinkprm.
This paper investigates Reinforcement Learning (RL) on data without explicit labels for reasoning tasks in Large Language Models (LLMs). The core challenge of the problem is reward estimation during inference while not having access to ground-truth information. While this setting appears elusive, we find that common practices in Test-Time Scaling (TTS), such as majority voting, yield surprisingly effective rewards suitable for driving RL training. In this work, we introduce Test-Time Reinforcement Learning (TTRL), a novel method for training LLMs using RL on unlabeled data. TTRL enables self-evolution of LLMs by utilizing the priors in the pre-trained models. Our experiments demonstrate that TTRL consistently improves performance across a variety of tasks and models. Notably, TTRL boosts the pass@1 performance of Qwen-2.5-Math-7B by approximately 159% on the AIME 2024 with only unlabeled test data. Furthermore, although TTRL is only supervised by the Maj@N metric, TTRL has demonstrated performance to consistently surpass the upper limit of the initial model, and approach the performance of models trained directly on test data with ground-truth labels. Our experimental findings validate the general effectiveness of TTRL across various tasks, and highlight TTRL's potential for broader tasks and domains. GitHub: https://github.com/PRIME-RL/TTRL
Effective reasoning remains a core challenge for large language models (LLMs) in the financial domain, where tasks often require domain-specific knowledge, precise numerical calculations, and strict adherence to compliance rules. We propose DianJin-R1, a reasoning-enhanced framework designed to address these challenges through reasoning-augmented supervision and reinforcement learning. Central to our approach is DianJin-R1-Data, a high-quality dataset constructed from CFLUE, FinQA, and a proprietary compliance corpus (Chinese Compliance Check, CCC), combining diverse financial reasoning scenarios with verified annotations. Our models, DianJin-R1-7B and DianJin-R1-32B, are fine-tuned from Qwen2.5-7B-Instruct and Qwen2.5-32B-Instruct using a structured format that generates both reasoning steps and final answers. To further refine reasoning quality, we apply Group Relative Policy Optimization (GRPO), a reinforcement learning method that incorporates dual reward signals: one encouraging structured outputs and another rewarding answer correctness. We evaluate our models on five benchmarks: three financial datasets (CFLUE, FinQA, and CCC) and two general reasoning benchmarks (MATH-500 and GPQA-Diamond). Experimental results show that DianJin-R1 models consistently outperform their non-reasoning counterparts, especially on complex financial tasks. Moreover, on the real-world CCC dataset, our single-call reasoning models match or even surpass the performance of multi-agent systems that require significantly more computational cost. These findings demonstrate the effectiveness of DianJin-R1 in enhancing financial reasoning through structured supervision and reward-aligned learning, offering a scalable and practical solution for real-world applications.
Existing methods to enhance the reasoning capability of large language models predominantly rely on supervised fine-tuning (SFT) followed by reinforcement learning (RL) on reasoning-specific data. These approaches critically depend on external supervisions--such as labeled reasoning traces, verified golden answers, or pre-trained reward models. In this work, we propose Entropy Minimized Policy Optimization (\ours), which makes an early attempt at fully unsupervised LLM reasoning incentivization. By continuously minimizing the predictive entropy of LLMs on unlabeled questions in a latent semantic space, \ours achieves competitive performance compared to supervised counterparts on both mathematical and free-form natural reasoning tasks. Specifically, without any supervised signals, \ours boosts the accuracy of Qwen2.5-Math-7B Base from 30.7\% to 48.1\% on mathematical benchmarks and improves the accuracy of Qwen2.5-7B Base from 32.1\% to 50.1\% on MMLU-Pro. Primary experiments and analysis are also provided to interpret the effectiveness of \ours. Code is available at https://github.com/QingyangZhang/EMPO.
Reinforcement learning has been shown to improve the performance of large language models. However, traditional approaches like RLHF or RLAIF treat the problem as single-step. As focus shifts toward more complex reasoning and agentic tasks, language models must take multiple steps of text generation, reasoning and environment interaction before generating a solution. We propose a synthetic data generation and RL methodology targeting multi-step optimization scenarios. This approach, called Step-Wise Reinforcement Learning (SWiRL), iteratively generates multi-step reasoning and tool use data, and then learns from that data. It employs a simple step-wise decomposition that breaks each multi-step trajectory into multiple sub-trajectories corresponding to each action by the original model. It then applies synthetic data filtering and RL optimization on these sub-trajectories. We evaluated SWiRL on a number of multi-step tool use, question answering, and mathematical reasoning tasks. Our experiments show that SWiRL outperforms baseline approaches by 21.5%, 12.3%, 14.8%, 11.1%, and 15.3% in relative accuracy on GSM8K, HotPotQA, CofCA, MuSiQue, and BeerQA, respectively. Excitingly, the approach exhibits generalization across tasks: for example, training only on HotPotQA (text question-answering) improves zero-shot performance on GSM8K (a math dataset) by a relative 16.9%.
Recent advancements in Large Language Models (LLMs) have shown that it is promising to utilize Process Reward Models (PRMs) as verifiers to enhance the performance of LLMs. However, current PRMs face three key challenges: (1) limited process supervision and generalization capabilities, (2) dependence on scalar value prediction without leveraging the generative abilities of LLMs, and (3) inability to scale the test-time compute of PRMs. In this work, we introduce GenPRM, a generative process reward model that performs explicit Chain-of-Thought (CoT) reasoning with code verification before providing judgment for each reasoning step. To obtain high-quality process supervision labels and rationale data, we propose Relative Progress Estimation (RPE) and a rationale synthesis framework that incorporates code verification. Experimental results on ProcessBench and several mathematical reasoning tasks show that GenPRM significantly outperforms prior PRMs with only 23K training data from MATH dataset. Through test-time scaling, a 1.5B GenPRM outperforms GPT-4o, and a 7B GenPRM surpasses Qwen2.5-Math-PRM-72B on ProcessBench. Additionally, GenPRM demonstrates strong abilities to serve as a critic model for policy model refinement. This work establishes a new paradigm for process supervision that bridges the gap between PRMs and critic models in LLMs. Our code, model, and data will be available in https://ryanliu112.github.io/GenPRM.
Recently, foundation models such as OpenAI's O1 and O3, along with DeepSeek's R1, have demonstrated strong reasoning capacities and problem-solving skills acquired through large-scale reinforcement learning (RL), with wide applications in mathematics, coding, science, intelligent agents, and virtual assistants. In this work, we introduce an off-policy reinforcement learning algorithm, EM Policy Gradient, aimed at enhancing LLM reasoning by optimizing expected return over reasoning trajectories. We frame the reasoning task as an Expectation-Maximization (EM) optimization problem, alternating between sampling diverse rationale trajectories and performing reward-guided fine-tuning. Unlike PPO and GRPO, which rely on complex importance weights and heuristic clipping, our method provides a simpler, more principled off-policy policy gradient approach, eliminating these complexities while maintaining strong performance. We evaluate the effectiveness of EM Policy Gradient on the GSM8K and MATH (HARD) datasets, where it achieves performance comparable to or slightly surpassing the state-of-the-art GRPO, while offering additional advantages in scalability, simplicity, and reasoning conciseness. Moreover, models fine-tuned with our method exhibit cognitive behaviors, such as sub-problem decomposition, self-verification, and backtracking, highlighting its potential to enhance both the interpretability and robustness of LLM reasoning.