MuMev: test-time training for profitable discovery
Luc Chartier
Session summary
The presenter describes building an experimental environment where an agent controls a wallet and interacts with historical Ethereum data to search for profitable opportunities. Using archived public mempool data and a compressed chain snapshot, the agent writes Python programs that inspect on-chain state, observe public order flow, and attempt to construct bundles that increase the wallet balance. The programs can decompile bytecode, run tests, and even try to exploit contracts within deterministic snapshots, competing against simple ordering strategies.
He frames this as a reinforcement-learning-style problem and compares several search approaches, including simple looping, tree search methods, GRPO, and test-time training, which he views as better aligned with finding rare, exceptional outcomes. He also reflects on the difficulty of wrapping RL around a complex tool harness, citing long traces and hard-to-debug failures, and notes alternatives like structured search with tree-based methods. Ultimately, the project consumed substantial storage, ended after a solid-state drive failure, and did not generate profits, but informed his views on search, MEV-style discovery, and model training.
Key takeaways
- The presenter built a historical Ethereum environment where an agent-controlled wallet writes Python scripts to act on archived public mempool data and chain snapshots, aiming to increase its balance.
- He enabled the agent’s programs to inspect blockchain state, analyze mempool data, decompile bytecode, run tests, and attempt contract exploits or bundles in deterministic snapshots.
- The presenter concludes that while this large data experiment was difficult to engineer and ultimately unprofitable, it clarified how reinforcement learning and test-time training might better target rare, high-value outcomes in similar search problems.