From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier

2026-01-22 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

The position paper "From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier" advocates for a critical shift in AI for Mathematics (AI4Math) systems from predefined problem-solvers to research agents. While LLM-driven theorem provers have achieved remarkable success in formal proof generation for well-defined problems using Interactive Theorem Proving (ITP) languages like Lean, they remain limited in tackling frontier research mathematics, such as discovering new theorems or resolving open conjectures. The authors systematically review the field, identifying core limitations across datasets, relational structure, mathematical exploration, tool ecosystems, and human-AI collaboration. They outline a strategic roadmap for the future, highlighting AI contributions to Erdős problems, which show rapid progress in literature review and formalization, but a smaller share of genuinely novel AI-primary solutions.

Key takeaway

For Research Scientists and Directors of AI/ML aiming to push mathematical AI capabilities, recognize that current LLM provers, while adept at competition-level problems, are not true research agents. Your teams should prioritize developing systems that can autonomously generate conjectures, faithfully formalize open-ended problems, and seamlessly integrate diverse external mathematical tools. Emphasize building robust human-AI collaboration frameworks and explainability into your systems to facilitate genuine mathematical discovery.

Key insights

AI4Math must transition from solving predefined problems to acting as research agents for mathematical discovery.

Principles

• Formal mathematical data scarcity limits LLM scaling.
• AI systems require modeling deep mathematical relationships.
• Human-AI collaboration is crucial for mathematical discovery.

Method

Neural theorem proving involves autoformalization (translating natural language to formal statements) and proof generation (producing verifiable symbolic sequences), typically optimized via supervised fine-tuning and reinforcement learning.

In practice

• Leverage ITPs like Lean for rigorous proof verification.
• Employ synthetic data generation to overcome data scarcity.
• Integrate external tools such as SMT solvers and CAS.

Topics

• AI for Mathematics
• Large Language Models
• Formal Verification
• Interactive Theorem Proving
• Automated Theorem Proving
• Mathematical Discovery
• Human-AI Collaboration

Code references

• ericjiang18/Awesome-Formal-Mathematics
• teorth/erdosproblems
• google-deepmind/formal-conjectures

Best for: AI Scientist, Research Scientist, Director of AI/ML

Related on AIssential

• ‘It is incredible’: How AI is transforming mathematics — Large language models are demonstrating surprising "original thought" in mathematics, solving complex problems with novel strategies.
• Advancing Mathematics Research with AI-Driven Formal Proof Search — LLM-driven formal proof search, verified by Lean, can solve open mathematical problems and advance research.
• Inference: When Will We Solve Intelligence — Solving intelligence requires a hybrid approach combining formal verification with large language models for robust reasoning.
• What it Means to Be a Mathematician When AI Does the Math — AI is shifting mathematics from human intuition and struggle to automated discovery and formal verification, raising existential questions.
• The AI Revolution in Math Has Arrived — AI is rapidly transforming mathematical research by accelerating discovery and proof generation, despite inherent model errors.
• OpenAI just SOLVED MATH.... — AI's breakthrough stemmed from its ability to connect disparate mathematical fields, revealing novel solutions.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.