Scaling test-time compute is crucial for enhancing the reasoning capabilities of large language models (LLMs). Existing approaches typically employ reinforcement learning (RL) to maximize a verifiable reward obtained at the end of reasoning traces. However, such methods optimize only the final performance under a large and fixed token budget, which hinders efficiency in both training and deployment. In this work, we present a novel framework, AnytimeReasoner, to optimize anytime reasoning performance, which aims to improve token efficiency and the flexibility of reasoning under varying token budget constraints. To achieve this, we truncate the complete thinking process to fit within sampled token budgets from a prior distribution, compelling the model to summarize the optimal answer for each truncated thinking for verification. This introduces verifiable dense rewards into the reasoning process, facilitating more effective credit assignment in RL optimization. We then optimize the thinking and summary policies in a decoupled manner to maximize the cumulative reward. Additionally, we introduce a novel variance reduction technique, Budget Relative Policy Optimization (BRPO), to enhance the robustness and efficiency of the learning process when reinforcing the thinking policy. Empirical results in mathematical reasoning tasks demonstrate that our method consistently outperforms GRPO across all thinking budgets under various prior distributions, enhancing both training and token efficiency.
The hallucination of code generation models hinders their applicability to systems requiring higher safety standards. One critical bottleneck in addressing code hallucination is the difficulty of identifying the functional correctness of generated code, due to its unnatural form. We address this core bottleneck by automatically generating unit tests using dynamic code analysis tools, leveraging the \emph{executable nature} of code. Accordingly, we propose a \emph{selective code generator} that abstains from uncertain generations -- based on the functional correctness evaluated by generated unit tests -- to theoretically control the correctness among non-abstained answers, \ie the false discovery rate. Finally, we propose to use generated unit tests in evaluation as well as in learning for precise code evaluation, calling this paradigm \emph{FuzzEval}. We demonstrate the efficacy of our method along with the controllability of code hallucination and reasonable selection efficiency.
Geometry problem solving is a key area of mathematical reasoning, which is widely involved in many important fields such as education, mathematical ability assessment of artificial intelligence, and multimodal ability assessment. In recent years, the rapid development of deep learning technology, especially the rise of multimodal large language models, has triggered a widespread research boom. This paper provides a survey of the applications of deep learning in geometry problem solving, including (i) a comprehensive summary of the relevant tasks in geometry problem solving; (ii) a thorough review of related deep learning methods; (iii) a detailed analysis of evaluation metrics and methods; and (iv) a critical discussion of the current challenges and future directions that can be explored. Our goal is to provide a comprehensive and practical reference of deep learning for geometry problem solving to promote further developments in this field. We create a continuously updated list of papers on GitHub: https://github.com/majianz/dl4gps.
Recent advances in reinforcement learning (RL) have led to substantial improvements in the mathematical reasoning abilities of large language models (LLMs), as measured by standard benchmarks. However, these gains often persist even when models are trained with flawed signals, such as random or inverted rewards, raising a fundamental question: do such improvements reflect true reasoning, or are they merely artifacts of overfitting to benchmark-specific patterns? To address this question, we take an evaluation-centric perspective and identify two critical shortcomings in existing protocols. First, \emph{benchmark contamination} arises from the public availability of test problems, increasing the risk of data leakage. Second, \emph{evaluation fragility} stems from the reliance on single-instance assessments, which are highly sensitive to stochastic outputs and fail to capture reasoning consistency. To overcome these limitations, we introduce {VAR-MATH}, a symbolic evaluation framework designed to probe genuine reasoning ability. By converting fixed numerical problems into symbolic templates and requiring models to solve multiple instantiations of each, VAR-MATH enforces consistent reasoning across structurally equivalent variants, thereby mitigating contamination and improving evaluation robustness. We apply VAR-MATH to transform two popular benchmarks, AMC23 and AIME24, into their symbolic counterparts, VAR-AMC23 and VAR-AIME24. Experimental results reveal substantial performance drops for RL-trained models on the variabilized versions, especially for smaller models, with average declines of 48.0\% on AMC23 and 58.3\% on AIME24. These findings suggest that many existing RL methods rely on superficial heuristics and fail to generalize beyond specific numerical forms. Overall, VAR-MATH offers a principled, contamination-resistant evaluation paradigm for mathematical reasoning.
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.
Chain-of-thought traces have been shown to improve performance of large language models in a plethora of reasoning tasks, yet there is no consensus on the mechanism through which this performance boost is achieved. To shed more light on this, we introduce Causal CoT Graphs (CCGs), which are directed acyclic graphs automatically extracted from reasoning traces that model fine-grained causal dependencies in the language model output. A collection of $1671$ mathematical reasoning problems from MATH500, GSM8K and AIME, and their associated CCGs are compiled into our dataset -- \textbf{KisMATH}. Our detailed empirical analysis with 15 open-weight LLMs shows that (i) reasoning nodes in the CCG are mediators for the final answer, a condition necessary for reasoning; and (ii) LLMs emphasise reasoning paths given by the CCG, indicating that models internally realise structures akin to our graphs. KisMATH enables controlled, graph-aligned interventions and opens up avenues for further investigation into the role of chain-of-thought in LLM reasoning.
The reasoning capabilities of large language models (LLMs) have been a longstanding focus of research. Recent works have further enhanced these capabilities using reinforcement learning (RL), with many new methods claiming significant improvements with minimal or no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance reasoning performance. However, these breakthroughs are mostly reported on the Qwen2.5 model family and evaluated on well-known benchmarks such as MATH-500, AMC, and AIME, while failing to achieve similar gains on other models like Llama, which warrants further investigation. Our analysis shows that although Qwen2.5 achieves strong mathematical reasoning performance, its pretraining on large-scale web corpora makes it vulnerable to data contamination in popular benchmarks. As a result, results derived from these benchmarks may be unreliable. To address this, we introduce a generator that produces fully synthetic arithmetic problems of arbitrary length and difficulty, yielding a clean dataset we call RandomCalculation. Using these leakage-free datasets, we show that only accurate reward signals consistently improve performance, while noisy or incorrect signals do not. We advocate for evaluating RL methods on uncontaminated benchmarks and across diverse model families to ensure trustworthy conclusions.
Current long chain-of-thought (long-CoT) models excel at mathematical reasoning but rely on slow and error-prone natural language traces. Tool-augmented agents address arithmetic via code execution, but often falter on complex logical tasks. We introduce a fine-tuning framework, DualDistill, that distills complementary reasoning strategies from multiple teachers into a unified student model. Using this approach, we train Agentic-R1, which dynamically selects the optimal strategy for each query, invoking tools for arithmetic and algorithmic problems, and using text-based reasoning for abstract ones. Our method improves accuracy across a range of tasks, including both computation-intensive and standard benchmarks, demonstrating the effectiveness of multi-strategy distillation in achieving robust and efficient reasoning. Our project is available at https://github.com/StigLidu/DualDistill
Enhancing the mathematical reasoning of Large Language Models (LLMs) is a pivotal challenge in advancing AI capabilities. While Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are the dominant training paradigms, a systematic methodology for combining them to maximize both accuracy and efficiency remains largely unexplored. This paper introduces a practical and effective training recipe that strategically integrates extended SFT with RL from online inference (GRPO). We posit that these methods play complementary, not competing, roles: a prolonged SFT phase first pushes the model's accuracy to its limits, after which a GRPO phase dramatically improves token efficiency while preserving this peak performance. Our experiments reveal that extending SFT for as many as 10 epochs is crucial for performance breakthroughs, and that the primary role of GRPO in this framework is to optimize solution length. The efficacy of our recipe is rigorously validated through top-tier performance on challenging benchmarks, including a high rank among over 2,200 teams in the strictly leak-free AI Mathematical Olympiad (AIMO). This work provides the community with a battle-tested blueprint for developing state-of-the-art mathematical reasoners that are both exceptionally accurate and practically efficient. To ensure full reproducibility and empower future research, we will open-source our entire framework, including all code, model checkpoints, and training configurations at https://github.com/analokmaus/kaggle-aimo2-fast-math-r1.
This paper deals with improving querying large language models (LLMs). It is well-known that without relevant contextual information, LLMs can provide poor quality responses or tend to hallucinate. Several initiatives have proposed integrating LLMs with external tools to provide them with up-to-date data to improve accuracy. In this paper, we propose a framework to integrate external tools to enhance the capabilities of LLMs in answering queries in educational settings. Precisely, we develop a framework that allows accessing external APIs to request additional relevant information. Integrated tools can also provide computational capabilities such as calculators or calendars. The proposed framework has been evaluated using datasets from the Multi-Modal Language Understanding (MMLU) collection. The data consists of questions on mathematical and scientific reasoning. Results compared to state-of-the-art language models show that the proposed approach significantly improves performance. Our Athena framework achieves 83% accuracy in mathematical reasoning and 88% in scientific reasoning, substantially outperforming all tested models including GPT-4o, LLaMA-Large, Mistral-Large, Phi-Large, and GPT-3.5, with the best baseline model (LLaMA-Large) achieving only 67% and 79% respectively. These promising results open the way to creating complex computing ecosystems around LLMs to make their use more natural to support various tasks and activities.
Large reasoning models (LRMs) have demonstrated impressive capabilities in domains like mathematics and program synthesis. Despite their strong performance, LRMs often exhibit overthinking -- excessive and redundant reasoning steps that introduce inefficiencies during inference. This phenomenon raises an important question for LRM self-evaluation: How can a model autonomously assess the correctness of its own reasoning trajectory without external labels? To address this, we propose Chain-of-Reasoning Embedding (CoRE), a series of hidden states in latent space to enable label-free self-evaluation on intermediate reasoning steps of LRMs, so as to enhance metacognition abilities for improved reasoning efficiency. By analyzing the geometric properties of the CoRE trajectories, we reveal that redundant reasoning usually presents cyclical fluctuations, which correspond to repetitive and unconscious reflection/exploration. Leveraging this insight, we further introduce a training-free, label-free self-evaluation framework, CoRE-Eval, to detect such patterns and dynamically determine whether to terminate reasoning early. Extensive experiments on mathematical reasoning benchmarks (GSM8K, MATH-500, and AIME) and across model sizes from 7B to 32B demonstrate that CoRE-Eval reduces chain-of-thought length by 13.7% to 33.2% while improving answer accuracy by around 10%, achieving 70.0% accuracy on the challenging AIME benchmark with the 32B model.
Translating natural language mathematical statements into formal, executable code is a fundamental challenge in automated theorem proving. While prior work has focused on generation and compilation success, little attention has been paid to the critic phase-the evaluation of whether generated formalizations truly capture the semantic intent of the original problem. In this paper, we introduce CriticLean, a novel critic-guided reinforcement learning framework that elevates the role of the critic from a passive validator to an active learning component. Specifically, first, we propose the CriticLeanGPT, trained via supervised fine-tuning and reinforcement learning, to rigorously assess the semantic fidelity of Lean 4 formalizations. Then, we introduce CriticLeanBench, a benchmark designed to measure models' ability to distinguish semantically correct from incorrect formalizations, and demonstrate that our trained CriticLeanGPT models can significantly outperform strong open- and closed-source baselines. Building on the CriticLean framework, we construct FineLeanCorpus, a dataset comprising over 285K problems that exhibits rich domain diversity, broad difficulty coverage, and high correctness based on human evaluation. Overall, our findings highlight that optimizing the critic phase is essential for producing reliable formalizations, and we hope our CriticLean will provide valuable insights for future advances in formal mathematical reasoning.
We introduce Skywork-R1V3, an advanced, open-source vision-language model (VLM) that pioneers a new approach to visual reasoning. Its key innovation lies in effectively transferring reasoning skills from text-only Large Language Models (LLMs) to visual tasks. The strong performance of Skywork-R1V3 primarily stems from our elaborate post-training RL framework, which effectively activates and enhances the model's reasoning ability, without the need for additional continue pre-training. Through this framework, we further uncover the fundamental role of the connector module in achieving robust cross-modal alignment for multimodal reasoning models. In addition, we introduce a unique indicator of reasoning capability, the entropy of critical reasoning tokens, which has proven highly effective for checkpoint selection during RL training. Skywork-R1V3 achieves state-of-the-art results on MMMU, significantly improving from 64.3% to 76.0%. This performance matches entry-level human capabilities. Remarkably, our RL-powered post-training approach enables even the 38B parameter model to rival top closed-source VLMs. The implementation successfully transfers mathematical reasoning to other subject-related reasoning tasks. We also include an analysis of curriculum learning and reinforcement finetuning strategies, along with a broader discussion on multimodal reasoning. Skywork-R1V3 represents a significant leap in multimodal reasoning, showcasing RL as a powerful engine for advancing open-source VLM capabilities.
The progress of Large Language Models (LLMs) like ChatGPT raises the question of how they can be integrated into education. One hope is that they can support mathematics learning, including word-problem solving. Since LLMs can handle textual input with ease, they appear well-suited for solving mathematical word problems. Yet their real competence, whether they can make sense of the real-world context, and the implications for classrooms remain unclear. We conducted a scoping review from a mathematics-education perspective, including three parts: a technical overview, a systematic review of word problems used in research, and a state-of-the-art empirical evaluation of LLMs on mathematical word problems. First, in the technical overview, we contrast the conceptualization of word problems and their solution processes between LLMs and students. In computer-science research this is typically labeled mathematical reasoning, a term that does not align with usage in mathematics education. Second, our literature review of 213 studies shows that the most popular word-problem corpora are dominated by s-problems, which do not require a consideration of realities of their real-world context. Finally, our evaluation of GPT-3.5-turbo, GPT-4o-mini, GPT-4.1, and o3 on 287 word problems shows that most recent LLMs solve these s-problems with near-perfect accuracy, including a perfect score on 20 problems from PISA. LLMs still showed weaknesses in tackling problems where the real-world context is problematic or non-sensical. In sum, we argue based on all three aspects that LLMs have mastered a superficial solution process but do not make sense of word problems, which potentially limits their value as instructional tools in mathematics classrooms.
Large language models can exhibit improved mathematical reasoning capabilities following post-training with instruction tuning, reinforcement learning, or knowledge distillation. However, it remains unclear whether these improvements are driven by major changes in transformer layers or from minor adjustments that leave the relative layer importance structures of the base model largely unchanged. We investigate this question through systematic layer-wise ablation experiments, examining base, instruction-tuned, knowledge-distilled, and reinforcement learning variants on mathematical reasoning benchmarks. Our findings show that mathematical reasoning gives rise to a specific layer importance structure, and this structure persists across all post-training paradigms. Removal of such layers causes accuracy drops of up to 80%. In contrast, non-mathematical tasks like factual recall exhibit no critical layers. This distinction suggests that mathematical reasoning requires specialized layers that emerge during pre-training, while other non-reasoning tasks do not. From an information-theoretic perspective, we also observe that these critical layers are the same layers where major representational transformation occurs.
Large language models (LLMs) have shown strong capabilities in complex reasoning, and test-time scaling techniques can enhance their performance with comparably low cost. Many of these methods have been developed and evaluated on mathematical reasoning benchmarks such as AIME. This paper investigates whether the lessons learned from these benchmarks generalize to the domain of advanced theoretical physics. We evaluate a range of common test-time scaling methods on the TPBench physics dataset and compare their effectiveness with results on AIME. To better leverage the structure of physics problems, we develop a novel, symbolic weak-verifier framework to improve parallel scaling results. Our empirical results demonstrate that this method significantly outperforms existing test-time scaling approaches on TPBench. We also evaluate our method on AIME, confirming its effectiveness in solving advanced mathematical problems. Our findings highlight the power of step-wise symbolic verification for tackling complex scientific problems.
Large language models (LLMs) have shown promising first-order logic (FOL) reasoning capabilities with applications in various areas. However, their effectiveness in complex mathematical reasoning involving multi-step FOL deductions is still under-researched. While LLMs perform competitively on established mathematical reasoning benchmarks, they struggle with multi-step FOL tasks, as demonstrated by Deepseek-Prover-V2-7B's low accuracy (4.2%) on our proposed theorem proving dataset. This issue arises from the limited exploration of diverse proof strategies and the potential for early reasoning mistakes to undermine entire proofs. To address these issues, we propose DREAM, a self-adaptive solution that enhances the Diversity and REAsonability of LLMs' generation strategies. DREAM incorporates an Axiom-Driven Strategy Diversification mechanism to promote varied strategic outcomes and a Sub-Proposition Error Feedback to help LLMs reflect on and correct their proofs. Our contributions include pioneering advancements in LLMs' mathematical reasoning through FOL theorem proving, introducing a novel inference stage solution that improves performance by 0.6% to 6.4%, and providing a curated dataset of 447 mathematical theorems in Lean 4 format for evaluation.
Reinforcement Learning (RL)-based post-training has significantly advanced the complex reasoning capabilities of language models, fostering sophisticated self-reflection processes. However, this ``slow thinking'' paradigm presents a critical challenge to reasoning efficiency: models may expend excessive computation on simple questions and shift reasoning prematurely for complex ones. Previous mechanisms typically rely on static length budgets or predefined rules, lacking the adaptability for varying question complexities and models' evolving capabilities. To this end, we propose AdapThink, an adaptive post-training framework designed to induce more efficient thinking while maintaining the performance of reasoning language models. Specifically, AdapThink incorporates two key mechanisms: 1) A group-relative reward function that leverages model confidence and response's characteristic to dynamically adjust the preference of reflection-related transition words without resorting to a fixed length preference. 2) A diversity-aware sampling mechanism that balances the training group's solution accuracy with reasoning diversity via an entropy-guided score. Experiments on several mathematical reasoning datasets with DeepSeek-distilled models demonstrate AdapThink's advantages in enabling adaptive reasoning patterns and mitigating the inefficiencies.
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 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.
Reinforcement learning (RL) is vital for optimizing large language models (LLMs). Recent Group Relative Policy Optimization (GRPO) estimates advantages using multiple on-policy outputs per prompt, leading to high computational costs and low data efficiency. To address this, we introduce Replay-Enhanced Policy Optimization (RePO), which leverages diverse replay strategies to retrieve off-policy samples from a replay buffer, allowing policy optimization based on a broader and more diverse set of samples for each prompt. Experiments on five LLMs across seven mathematical reasoning benchmarks demonstrate that RePO achieves absolute average performance gains of $18.4$ and $4.1$ points for Qwen2.5-Math-1.5B and Qwen3-1.7B, respectively, compared to GRPO. Further analysis indicates that RePO increases computational cost by $15\%$ while raising the number of effective optimization steps by $48\%$ for Qwen3-1.7B, with both on-policy and off-policy sample numbers set to $8$. The repository can be accessed at https://github.com/SihengLi99/RePO.