Self-evolving Lean proof agents reach 45.1% on miniF2F with coevolving benchmarks
omarsar0 · x · 2026-07-22
A paper on self-modifying Lean proof agents argues that agents should co-evolve with their benchmarks instead of optimizing against a fixed test set.
- The system uses a small trusted runtime around a fully mutable workspace covering workflow, prompts, and tools.
- Between generations, the current champion updates the task distribution with a mastery-throttled curriculum, adding harder proof obligations only after the current level is mastered.
- A single-anchor recalibration reruns the champion on the updated benchmark so scores stay comparable as difficulty rises.
- All success is grounded in a Lean verifier: a result only counts if it produces verified proofs under a trusted snapshot.
- Over 15 generations, the best coevolving agent reaches 45.1% held-out solve rate on miniF2F, versus 12.7% for the seed and 32.0% for the best fixed-benchmark agent.
The paper argues that verifier-grounded self-evolution can improve formal proof workflows while avoiding reward hacking.
More from AGI Musings
- Matt Perault says AI law should fit existing legal principles, not rewrite 1L — MattPerault · 2026-07-22
- Paper frames AI alignment as a moving sociotechnical target — weballergy · 2026-07-22
- AI coding tools are making fast social-science foresight experiments much cheaper — weballergy · 2026-07-22
- Researchers warn static alignment could cause value lock-in and societal stagnation — weballergy · 2026-07-22
- A population model suggests AI alignment can slow social progress under strong lock-in — weballergy · 2026-07-22
- Paper argues AI alignment breaks when human values keep evolving — weballergy · 2026-07-22