AI

Fix, Not Just Resample: Bounded Compiler-Guided Repair in Lean 4

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 21/10/2026, 04:03:25 UTC.

Briefings

A new method replaces discarding failed proof attempts with targeted repair using a pretrained model, improving solve rates by up to 32% on the LeanDojo benchmark. The approach samples multiple tactics per step and repairs at most R times, trading runtime for proof reachability. Results were validated on a second benchmark with zero regressions.

Key points

• Automated theorem provers discard failed steps, but a new method repairs them using a pretrained model instead.
• The method achieves up to a 32% relative gain in solve rates on the LeanDojo benchmark.

Context

This research addresses a limitation in automated theorem proving where failed steps are discarded rather than repaired. The method samples multiple tactics per step and repairs at most R times, trading runtime for proof reachability.

Limitations

Results are limited to the LeanDojo benchmark subset of 1,990 theorems. The study uses preprint status and does not include additional background or causal claims beyond the reported findings.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 21/10/2026, 04:33:13 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeicivytax3jpl554glbnpbygiilsm4lg665dk2hzcolew4ln5ify3y