Research map / Large language models / Agents, code and math
ProofOptimizer: Training Language Models to Simplify Proofs without Human Demonstrations
Alex Gu, Bartosz Piotrowski, Fabian Gloeckle, Kaiyu Yang, Aram H. Markosyan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RL training after expert iteration (ablation) | ProofOptimizer · miniF2F | red@1 (%): 49.0% -> 63.6% | Improved |
| add RL training after expert iteration (ablation) | ProofOptimizer · miniF2F | red@32 (%): 72.3% -> 70.9% | Hurt |
| add RL training after expert iteration (ablation) | ProofOptimizer · PutnamBench | red@32 (%): 26.3% -> 21.1% | Hurt |
| add test-time RL (ablation) | ProofOptimizer-RL (It 2) · PutnamBench | red@1 (%): 14.9% -> 23.8% | Improved |
| more expert iteration rounds (ablation) | ProofOptimizer-ExpIt · miniF2F | red@32 (%): 67.0% -> 72.3% | Improved |
| fine-tune 7B model for proof simplification vs Gemini-2.5-Pro | Qwen2.5-7B-Instruct · miniF2F | red@32 (%): 57.2% -> 72.3% | Improved |
| add proof state annotations to Gemini-2.5-Pro prompt (ablation) | Gemini-2.5-Pro · miniF2F | red@1 (%): 24.3% -> 26.4% | No clear effect |
| repair failed simplifications with execution feedback instead of resampling (ablation) | ProofOptimizer-ExpIt · miniF2F, PutnamBench | proof length: repaired length 644 -> 576 after linter (still longer than originals often) | Hurt |
| iterative proof shortening with more iterations/samples (ablation) | ProofOptimizer-ExpIt · miniF2F | avg proof length (tokens): 334 -> 75 | Improved |
| SFT on simplified proofs instead of original proofs (ablation) | Qwen2.5-7B-Instruct Lean base · miniF2F | pass rate: about +2% | Improved |
| optimize heartbeats instead of proof length (ablation) | ProofOptimizer-ExpIt · miniF2F | proofs with >10% speedup: 114/194 -> 138/194 (proofs somewhat longer) | Mixed |
| shorten proofs by length | ProofOptimizer-ExpIt · PutnamBench | execution time: 50/75 proofs >10% speedup | Mixed |
Closest papers by meaning
- ProofAug: Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis (ICML 2025 · similarity 0.91)
- Lean-STaR: Learning to Interleave Thinking and Proving (ICLR 2025 · similarity 0.91)
- DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search (ICLR 2025 · similarity 0.91)
- ImProver: Agent-Based Automated Proof Optimization (ICLR 2025 · similarity 0.91)
- Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models (NeurIPS 2025 · similarity 0.90)
- LEGO-Prover: Neural Theorem Proving with Growing Libraries (ICLR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
