Reproduce “Learning Interestingness in Automated Mathematical Theory Formation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
George Tsoukalas, Rahul Saha, Amitayush Thakur, Sabrina Reguyal, Swarat Chaudhuri
A new reinforcement learning environment, FERMAT, models open-ended mathematical theory formation, employing an LLM-driven evolutionary algorithm called EvoAbstract to autonomously learn "interestingness" measures. This system effectively rediscovered fundamental mathematical concepts in elementary number theory and finite fields, demonstrating superior performance over static heuristics and one-shot LLM approaches.
Key insights
- Learning explicit, programmatic interestingness measures is an effective strategy for navigating the vast search space of open-ended mathematical discovery.
- Combining Large Language Models with evolutionary program synthesis enables the generation of sophisticated and adaptable interestingness heuristics.
- An abstraction learning component can promote the modularity and interpretability of learned programs by identifying and reusing common subroutines, potentially accelerating early optimization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.