Market Incentivization for Theorem Proving with LLM Agents

Agentic and multi-agent systems have been increasingly used for solving frontier-level math problems. We present preliminary experimental results with a market-based multi-agent system for proving mathematical theorems, in which agent coordination is mediated by market signals rather than a hand-engineered harness. Agent communities learn via evolutionary signals from market profits and losses. We report initial findings on whether evolutionary pressure produces more profitable agent populations and discuss remaining challenges towards an autonomous multi-agent mathematical system that learns from markets.