Research map / Large language models / Agents, code and math
Compile to Compress: Boosting Formal Theorem Provers by Compiler Outputs
Guchan Li, Rui Tian, Hongning Wang
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 |
|---|---|---|---|
| fine-tune on refinement data with expert iteration plus random tree search | Goedel-Prover-V2-32B · PutnamBench (budget 64) | problems solved: 32 -> 63 | Improved |
| fine-tune on refinement data with expert iteration plus random tree search | Kimina-Distill-8B · PutnamBench (budget 64) | problems solved: 10 -> 17 | Improved |
| replace refinement training with direct-solution SFT on more data (ablation) | Kimina-Distill-8B · MiniF2F-test | accuracy [%]: 77.46 -> 75.41 | Hurt |
| replace refinement training with direct-solution SFT on more data (ablation) | Goedel-Prover-V2-32B · MathOlympiadBench | accuracy [%]: 14.72 -> 12.50 | Hurt |
| add value-guided tree search over random tree search (ablation) | Goedel-Prover-V2-32B · MiniF2F-test | accuracy [%]: 84.43 -> 86.89 | Improved |
| add value-guided tree search over random tree search (ablation) | Kimina-Distill-8B · ProofNet | accuracy [%]: 15.63 -> 15.36 | Hurt |
| add value-guided search on hard problems (ablation) | Goedel-Prover-V2-32B · MiniF2F-test hard | accuracy [%]: 18.42 -> 26.32 | Improved |
| use DFS refinement instead of BFS sampling (ablation) | Kimina-Distill-8B · MiniF2F-test hard | accuracy [%]: 10.91 -> 20.00 | Improved |
| use refinement instead of direct sampling on easy problems | provers · MiniF2F-test | accuracy [%] | No clear effect |
| scale test budget from 64 to 256 | Goedel-Expert (value) · PutnamBench | problems solved: 63 -> 110 | Improved |
| use Markovian refinement with Goedel-Expert | Goedel-Prover-V2-32B · PutnamBench total output tokens | tokens: 924,911 K -> 476,133 K | Improved |
Closest papers by meaning
- MPS-Prover: Advancing Stepwise Theorem Proving by Multi-Perspective Search and Data Curation (NeurIPS 2025 · similarity 0.90)
- APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning (NeurIPS 2025 · similarity 0.88)
- DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search (ICLR 2025 · similarity 0.88)
- FVEL: Interactive Formal Verification Environment with Large Language Models via Theorem Proving (NeurIPS 2024 · similarity 0.88)
- ProofAug: Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis (ICML 2025 · similarity 0.87)
- ImProver: Agent-Based Automated Proof Optimization (ICLR 2025 · similarity 0.87)
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.
