mini-Lean-agent: bash-only coding agent solves 12/12 Putnam 2025 at $1.28 per problem
wellecks · x · 2026-10-21
A new arXiv paper (Ortega et al., arXiv:2610.17331) introduces mini-Lean-agent, a deliberately minimal agent for Lean theorem proving whose only model-visible tool is bash — the model must inspect/edit the repo, search local libraries, invoke Lean, and manage intermediate work itself, with a separate trusted evaluator verifying the final artifact. The design isolates whether benchmark gains come from better base models, specialized harnesses, or specific tools.
Key results on the Putnam 2025 benchmark:
- With Gemini 3.8 Flash as base: 7/12 solved within a $2-per-problem limit, vs. 4/12 for AxProverBase.
- With Astra as base: 12/12 solved at an average cost of $1.28 per problem.
The takeaway: a generic coding agent with a minimal harness already performs strongly on Lean benchmarks, and the authors propose it as a standard baseline for evaluating theorem-proving infrastructure. A benchmark and the open-source Leanimum Agent were released alongside.
Related event: Minimal Bash-Only Lean Agent Aces Putnam 2025(2 posts)→
More from coding & agent
- PyTorch accelerator integration group details progress and priorities at PyTorchCon — PyTorch · 2026-10-21
- wshobson/agents hits 41k stars as a multi-harness plugin marketplace for Claude Code, Codex, Cursor and more — wshobson · 2026-10-21
- Neo4j Desktop 2.3.0 embeds terminal and AI coding-agent skills into graph workflows — JeremyCMorgan · 2026-10-21
- ChatGPT assistant dot sent a business email without showing a draft, breaking user trust — denby10562 · 2026-10-21
- Characterizing a cheap SG90 servo with Claude's plan to build a MuJoCo model — _Stocko_ · 2026-10-21
- Decision model leads, LLM reviews: SRE diagnosis pass rate jumps to 89.5% — tianyin_xu · 2026-10-21