Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CoSineVerifier: Tool-Augmented Answer Verification for Computation-Oriented Scientific Questions

Ruixiang Feng, Zhenwei An, Yuntao Wen, Ran Le, Yiming Jia, Chen Yang, Zongchao Chen, Lisi Chen, Shen Gao, Shuo Shang, Yang Song, Tao Zhang

TL;DR

CoSineVerifier introduces a tool-augmented verifier that uses external executors to perform precise calculations and symbolic simplifications for computation-heavy scientific questions. It employs a two-stage training pipeline—cold-start supervised fine-tuning followed by reinforcement learning with tool-use encouragement—together with long-tail and difficulty-oriented data augmentation. The approach achieves state-of-the-art verification performance on VerifyBench and Sci-Bench, and serves as a fast, reliable reward model that improves RLVR results on AIME 2024 and 2025, highlighting the value of tool-based verification for enhanced reasoning. Notably, a compact 4B model with external tools can surpass larger models in both accuracy and efficiency, underscoring practical benefits for deployment in real-world verification and RL workflows.

Abstract

Answer verification methods are widely employed in language model training pipelines spanning data curation, evaluation, and reinforcement learning with verifiable rewards (RLVR). While prior work focus on developing unified verifiers applicable across multiple reasoning scenarios, significant challenges remain in computation-oriented scientific domains, such as algebraic equivalence checking and physical constant substitution. In this paper, we introduce \model, a tool-augmented verifier that leverages external executors to perform precise computations and symbolic simplifications. \model enables robust verification that goes beyond simple semantic matching. We propose a novel two-stage pipeline, which begin with cold-start fine-tuning and followed by multi-turn reinforcement learning with tool integration. Extensive experiments conducted on STEM subjects, general QA, and long-form reasoning tasks demonstrates strong generalization of \model. The results shows that the \model achieves state-of-the-art performance on VerifyBench-Hard and SCI-Bench. And we also employ our \model in RLVR as a reward model, the results show that it consistently outperforms both rubric-based and model-based verifiers on AIME'24 and AIME'25, demonstrating strong potential to enhance reasoning capabilities of LLM. Our model is released at \hyperlink{https://huggingface.co/Nanbeige/CoSineVerifier-Tool-4B}{https://huggingface.co/Nanbeige/CoSineVerifier-Tool-4B}.

CoSineVerifier: Tool-Augmented Answer Verification for Computation-Oriented Scientific Questions

TL;DR

CoSineVerifier introduces a tool-augmented verifier that uses external executors to perform precise calculations and symbolic simplifications for computation-heavy scientific questions. It employs a two-stage training pipeline—cold-start supervised fine-tuning followed by reinforcement learning with tool-use encouragement—together with long-tail and difficulty-oriented data augmentation. The approach achieves state-of-the-art verification performance on VerifyBench and Sci-Bench, and serves as a fast, reliable reward model that improves RLVR results on AIME 2024 and 2025, highlighting the value of tool-based verification for enhanced reasoning. Notably, a compact 4B model with external tools can surpass larger models in both accuracy and efficiency, underscoring practical benefits for deployment in real-world verification and RL workflows.

Abstract

Answer verification methods are widely employed in language model training pipelines spanning data curation, evaluation, and reinforcement learning with verifiable rewards (RLVR). While prior work focus on developing unified verifiers applicable across multiple reasoning scenarios, significant challenges remain in computation-oriented scientific domains, such as algebraic equivalence checking and physical constant substitution. In this paper, we introduce \model, a tool-augmented verifier that leverages external executors to perform precise computations and symbolic simplifications. \model enables robust verification that goes beyond simple semantic matching. We propose a novel two-stage pipeline, which begin with cold-start fine-tuning and followed by multi-turn reinforcement learning with tool integration. Extensive experiments conducted on STEM subjects, general QA, and long-form reasoning tasks demonstrates strong generalization of \model. The results shows that the \model achieves state-of-the-art performance on VerifyBench-Hard and SCI-Bench. And we also employ our \model in RLVR as a reward model, the results show that it consistently outperforms both rubric-based and model-based verifiers on AIME'24 and AIME'25, demonstrating strong potential to enhance reasoning capabilities of LLM. Our model is released at \hyperlink{https://huggingface.co/Nanbeige/CoSineVerifier-Tool-4B}{https://huggingface.co/Nanbeige/CoSineVerifier-Tool-4B}.

Paper Structure

This paper contains 42 sections, 15 equations, 8 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Answer Verification
  4. Tool Integrated Reasoning
  5. Preliminaries
  6. Answer verification
  7. Methodology: CoSineVerifier
  8. Tool-augmented Verification
  9. Data Curation
  10. Data Collection
  11. Data Annotation
  12. Data augmentation
  13. Long-tail data augmentation
  14. Difficulty-oriented data augmentation
  15. Post Training
  16. ...and 27 more sections

Figures (8)

  • Figure 1: Comparison between CoSineVerifier and existing methods, which demonstrates our CoSineVerifier can enable accurate judgment on calculation-intensive scientific scenarios
  • Figure 2: Error distribution on existing verification methods. More Details can be found in the Appendix \ref{['appendix:error_analysis']}.
  • Figure 3: (a) Workflow of tool-augmented verification. (b) The overall training framework of CoSineVerifier.
  • Figure 4: Performance Analysis. (a) Analysis on accuracy across different domains. (b) Inference efficiency analysis of CoSineVerifier-4B-Tool.
  • Figure 5: (a) Ablation study of CoSineVerifier-4B-Tool. (b) Average verification accuracy with tools before RL and after RL
  • ...and 3 more figures