RMA: an Agentic System for Research-Level Mathematical Problems

Zelin Zhao, Bo Yuan, Jaemoo Choi, Yongxin Chen

arXiv:2605.22875 · 2026-05-25 공개 · arXiv · PDF

mathematical-reasoning structured-reasoning iterative-refinement multi-agent-system agent-framework proof-verification literature-search research-math-agents

Abstract

We present $\textbf{Research Math Agents (RMA)}$, an agentic framework for automated reasoning on research-level mathematical problems. Unlike prior studies centered on competition mathematics or formal theorem proving, RMA targets research-level mathematical problems that require long-horizon reasoning, literature grounding, and iterative proof refinement. RMA decomposes research-level proof solving into specialized modules for problem analysis, literature search and understanding, fair comparison, knowledge-bank construction, and proof verification, all coordinated by initializer, proposer, and verifier agents through a shared structured memory. Within this unified framework, these agents operate in a multi-role, multi-round workflow, collaboratively generating, refining, and verifying candidate proofs through iterative feedback. We evaluate RMA on the First Proof benchmark, which consists of ten research-level problems contributed by expert mathematicians across diverse domains. Through comprehensive expert evaluation, RMA outperforms strong baselines on the First Proof benchmark, including GPT-5.2R and Aletheia, solving eight out of ten research problems and producing more logically sound and readable proofs. Our comprehensive ablation studies further show that performance gains arise from the interaction of structured reasoning modules, iterative refinement, and verifier-based feedback, rather than any single component. Our solutions and implementations will be made publicly available upon acceptance.

한국어 요약

📋 한 줄 요약

**[Research Math Agent / Long-Horizon Proof]** RMA가 problem analysis·literature search·knowledge-bank·proof verification 모듈을 initializer/proposer/verifier 에이전트로 협조시켜 First Proof 벤치마크에서 10문제 중 8개 해결, GPT-5.2R·Aletheia 능가.

🎯 핵심 기여도

💡 핵심 아이디어

Research-level 수학 증명은 단발 LLM 호출이나 formal proof search로는 불충분하며, 문제 분석·문헌 활용·증명 비교·지식 베이스 구축·검증을 모듈로 분해하고 multi-role·multi-round 협업을 shared memory 위에서 반복해야 인간 수학자 수준에 접근하는 증명이 가능하다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Research-level 수학에 적합한 첫 본격 multi-agent 프레임워크, 모듈 분해와 에이전트 협업이 long-horizon proof에 효과적임을 정량 입증, 향후 수학 자동화의 design template 제공. **한계**: First Proof 10문제로 평가 규모가 작음, 전문가 평가의 주관성 잔존, frontier 모델 의존도와 비용 부담, 자동화된 verifier가 모든 분야에 통할지 추가 검증 필요.

🚀 실용적 활용