Agent-Guided Process Verification for Mathematical Self-Correction
Summary
- Researchers propose Agent-Guided Process Verification (AGPV) for mathematical self-correction in large language models.
- AGPV decomposes solutions into key claims, verifies them individually using a mathematical toolbox, and stops at the first error.
- A summarization agent diagnoses failures to guide a correction agent, and AGPV also verifies solution completeness.
- Evaluated across five datasets with four models from DeepSeek, Qwen, and GPT, AGPV achieves 63.8% average accuracy, outperforming baselines.
Why it matters
AGPV addresses multi-step reasoning errors in large language models by providing granular process verification and structured self-correction, improving mathematical problem-solving accuracy.