Safe: Enhancing Mathematical Reasoning in Large Language Models via Retrospective Step-aware Formal Verification

Chain-of-Thought (CoT) prompting has become the de facto method to elicit reasoning capabilities from large language models (LLMs). However, to mitigate hallucinations in CoT that are notoriously difficult to detect, current methods such as process reward models (PRMs) or self-consistency operate as opaque boxes and do not provide checkable evidence for their judgments, possibly limiting their effectiveness. To address this issue, we draw inspiration from the idea that "the gold standard for supporting a mathematical claim is to provide a proof". We propose a retrospective, step-aware formal verification framework $Safe$. Rather than assigning arbitrary scores, we strive to articulate mathematical claims in formal mathematical language Lean 4 at each reasoning step and provide formal proofs to identify hallucinations. We evaluate our framework $Safe$ across multiple language models and various mathematical datasets, demonstrating a significant performance improvement while offering interpretable and verifiable evidence. We also propose $FormalStep$ as a benchmark for step correctness theorem proving with $30,809$ formal statements. To the best of our knowledge, our work represents the first endeavor to utilize formal mathematical language Lean 4 for verifying natural language content generated by LLMs, aligning with the reason why formal mathematical languages were created in the first place: to provide a robust foundation for hallucination-prone human-written proofs.

Pangu DeepDiver: Adaptive Search Intensity Scaling via Open-Web Reinforcement Learning

Information seeking demands iterative evidence gathering and reflective reasoning, yet large language models (LLMs) still struggle with it in open-web question answering. Existing methods rely on static prompting rules or training with Wikipedia-based corpora and retrieval environments, limiting adaptability to the real-world web environment where ambiguity, conflicting evidence, and noise are prevalent. These constrained training settings hinder LLMs from learning to dynamically decide when and where to search, and how to adjust search depth and frequency based on informational demands. We define this missing capacity as Search Intensity Scaling (SIS)--the emergent skill to intensify search efforts under ambiguous or conflicting conditions, rather than settling on overconfident, under-verification answers. To study SIS, we introduce WebPuzzle, the first dataset designed to foster information-seeking behavior in open-world internet environments. WebPuzzle consists of 24K training instances and 275 test questions spanning both wiki-based and open-web queries. Building on this dataset, we propose DeepDiver, a Reinforcement Learning (RL) framework that promotes SIS by encouraging adaptive search policies through exploration under a real-world open-web environment. Experimental results show that Pangu-7B-Reasoner empowered by DeepDiver achieve performance on real-web tasks comparable to the 671B-parameter DeepSeek-R1. We detail DeepDiver's training curriculum from cold-start supervised fine-tuning to a carefully designed RL phase, and present that its capability of SIS generalizes from closed-form QA to open-ended tasks such as long-form writing. Our contributions advance adaptive information seeking in LLMs and provide a valuable benchmark and dataset for future research.

Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning

Although large language models demonstrate strong performance across various domains, they still struggle with numerous bad cases in mathematical reasoning. Previous approaches to learning from errors synthesize training data by solely extrapolating from isolated bad cases, thereby failing to generalize the extensive patterns inherent within these cases. This paper presents Self-Error-Instruct (SEI), a framework that addresses these model weaknesses and synthesizes more generalized targeted training data. Specifically, we explore a target model on two mathematical datasets, GSM8K and MATH, to pinpoint bad cases. Then, we generate error keyphrases for these cases based on the instructor model's (GPT-4o) analysis and identify error types by clustering these keyphrases. Next, we sample a few bad cases during each generation for each identified error type and input them into the instructor model, which synthesizes additional training data using a self-instruct approach. This new data is refined through a one-shot learning process to ensure that only the most effective examples are kept. Finally, we use these curated data to fine-tune the target model, iteratively repeating the process to enhance performance. We apply our framework to various models and observe improvements in their reasoning abilities across both in-domain and out-of-domain mathematics datasets. These results demonstrate the effectiveness of self-error instruction in improving LLMs' mathematical reasoning through error generalization.

Stepwise Reasoning Checkpoint Analysis: A Test Time Scaling Method to Enhance LLMs' Reasoning

Mathematical reasoning through Chain-of-Thought (CoT) has emerged as a powerful capability of Large Language Models (LLMs), which can be further enhanced through Test-Time Scaling (TTS) methods like Beam Search and DVTS. However, these methods, despite improving accuracy by allocating more computational resources during inference, often suffer from path homogenization and inefficient use of intermediate results. To address these limitations, we propose Stepwise Reasoning Checkpoint Analysis (SRCA), a framework that introduces checkpoints between reasoning steps. It incorporates two key strategies: (1) Answer-Clustered Search, which groups reasoning paths by their intermediate checkpoint answers to maintain diversity while ensuring quality, and (2) Checkpoint Candidate Augmentation, which leverages all intermediate answers for final decision-making. Our approach effectively reduces path homogenization and creates a fault-tolerant mechanism by utilizing high-quality intermediate results. Experimental results show that SRCA improves reasoning accuracy compared to existing TTS methods across various mathematical datasets.

The Real Barrier to LLM Agent Usability is Agentic ROI

Large Language Model (LLM) agents represent a promising shift in human-AI interaction, moving beyond passive prompt-response systems to autonomous agents capable of reasoning, planning, and goal-directed action. Despite the widespread application in specialized, high-effort tasks like coding and scientific research, we highlight a critical usability gap in high-demand, mass-market applications. This position paper argues that the limited real-world adoption of LLM agents stems not only from gaps in model capabilities, but also from a fundamental tradeoff between the value an agent can provide and the costs incurred during real-world use. Hence, we call for a shift from solely optimizing model performance to a broader, utility-driven perspective: evaluating agents through the lens of the overall agentic return on investment (Agent ROI). By identifying key factors that determine Agentic ROI--information quality, agent time, and cost--we posit a zigzag development trajectory in optimizing agentic ROI: first scaling up to improve the information quality, then scaling down to minimize the time and cost. We outline the roadmap across different development stages to bridge the current usability gaps, aiming to make LLM agents truly scalable, accessible, and effective in real-world contexts.

How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study

Large Reasoning Models (LRMs) have achieved remarkable success on reasoning-intensive tasks such as mathematics and programming. However, their enhanced reasoning capabilities do not necessarily translate to improved safety performance-and in some cases, may even degrade it. This raises an important research question: how can we enhance the safety of LRMs? In this paper, we present a comprehensive empirical study on how to enhance the safety of LRMs through Supervised Fine-Tuning (SFT). Our investigation begins with an unexpected observation: directly distilling safe responses from DeepSeek-R1 fails to significantly enhance safety. We analyze this phenomenon and identify three key failure patterns that contribute to it. We then demonstrate that explicitly addressing these issues during the data distillation process can lead to substantial safety improvements. Next, we explore whether a long and complex reasoning process is necessary for achieving safety. Interestingly, we find that simply using short or template-based reasoning process can attain comparable safety performance-and are significantly easier for models to learn than more intricate reasoning chains. These findings prompt a deeper reflection on the role of reasoning in ensuring safety. Finally, we find that mixing math reasoning data during safety fine-tuning is helpful to balance safety and over-refusal. Overall, we hope our empirical study could provide a more holistic picture on enhancing the safety of LRMs. The code and data used in our experiments are released in https://github.com/thu-coai/LRM-Safety-Study.

Instruction-Tuning Data Synthesis from Scratch via Web Reconstruction

The improvement of LLMs' instruction-following capabilities depends critically on the availability of high-quality instruction-response pairs. While existing automatic data synthetic methods alleviate the burden of manual curation, they often rely heavily on either the quality of seed data or strong assumptions about the structure and content of web documents. To tackle these challenges, we propose Web Reconstruction (WebR), a fully automated framework for synthesizing high-quality instruction-tuning (IT) data directly from raw web documents with minimal assumptions. Leveraging the inherent diversity of raw web content, we conceptualize web reconstruction as an instruction-tuning data synthesis task via a novel dual-perspective paradigm--Web as Instruction and Web as Response--where each web document is designated as either an instruction or a response to trigger the reconstruction process. Comprehensive experiments show that datasets generated by WebR outperform state-of-the-art baselines by up to 16.65% across four instruction-following benchmarks. Notably, WebR demonstrates superior compatibility, data efficiency, and scalability, enabling enhanced domain adaptation with minimal effort. The data and code are publicly available at https://github.com/YJiangcm/WebR.

ToolACE-R: Tool Learning with Adaptive Self-Refinement

Tool learning, which allows Large Language Models (LLMs) to leverage external tools for solving complex user tasks, has emerged as a promising avenue for extending model capabilities. However, current approaches primarily focus on data synthesis for fine-tuning LLMs to invoke tools effectively, largely ignoring how to fully stimulate the potential of the model. In this paper, we propose ToolACE-R, a novel method that introduces adaptive self-refinement for tool invocations. Our approach features a model-aware iterative training procedure that progressively incorporates more training samples based on the model's evolving capabilities. Additionally, it allows LLMs to iteratively refine their tool calls, optimizing performance without requiring external feedback. To further enhance computational efficiency, we integrate an adaptive mechanism when scaling the inference time, enabling the model to autonomously determine when to stop the refinement process. We conduct extensive experiments across several benchmark datasets, showing that ToolACE-R achieves competitive performance compared to advanced API-based models, even without any refinement. Furthermore, its performance can be further improved efficiently through adaptive self-refinement. Our results demonstrate the effectiveness of the proposed method, which is compatible with base models of various sizes, offering a promising direction for more efficient tool learning.

DAST: Difficulty-Aware Self-Training on Large Language Models

Present Large Language Models (LLM) self-training methods always under-sample on challenging queries, leading to inadequate learning on difficult problems which limits LLMs' ability. Therefore, this work proposes a difficulty-aware self-training (DAST) framework that focuses on improving both the quantity and quality of self-generated responses on challenging queries during self-training. DAST is specified in three components: 1) sampling-based difficulty level estimation, 2) difficulty-aware data augmentation, and 3) the self-training algorithm using SFT and DPO respectively. Experiments on mathematical tasks demonstrate the effectiveness and generalization of DAST, highlighting the critical role of difficulty-aware strategies in advancing LLM self-training.

Learning to Align Multi-Faceted Evaluation: A Unified and Robust Framework

Large Language Models (LLMs) are being used more and more extensively for automated evaluation in various scenarios. Previous studies have attempted to fine-tune open-source LLMs to replicate the evaluation explanations and judgments of powerful proprietary models, such as GPT-4. However, these methods are largely limited to text-based analyses under predefined general criteria, resulting in reduced adaptability for unseen instructions and demonstrating instability in evaluating adherence to quantitative and structural constraints. To address these limitations, we propose a novel evaluation framework, ARJudge, that adaptively formulates evaluation criteria and synthesizes both text-based and code-driven analyses to evaluate LLM responses. ARJudge consists of two components: a fine-tuned Analyzer that generates multi-faceted evaluation analyses and a tuning-free Refiner that combines and refines all analyses to make the final judgment. We construct a Composite Analysis Corpus that integrates tasks for evaluation criteria generation alongside text-based and code-driven analysis generation to train the Analyzer. Our results demonstrate that ARJudge outperforms existing fine-tuned evaluators in effectiveness and robustness. Furthermore, it demonstrates the importance of multi-faceted evaluation and code-driven analyses in enhancing evaluation capabilities.