Research map / Reinforcement learning / Visual and representation RL
Reinforcement Learning of Theorem Proving
Cezary Kaliszyk, Josef Urban, Henryk Michalewski, Miroslav Olšák
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 jointly learned policy and value guidance with UCT | rlCoP · Miz40 test | problems proved: 1143 -> 1624 | Improved |
| Add jointly learned policy and value guidance with UCT | rlCoP · M2k | problems proved: 876 -> 1235 | Improved |
| Use learned policy only, retaining heuristic value instead of jointly learning value (ablation) | rlCoP · M2k | problems proved: 1235 -> 1087 | Hurt |
| Use learned value only and remove learned policy (ablation) | rlCoP · M2k | problems proved: 1235 -> 1144 | Hurt |
| Add learned policy guidance while keeping heuristic state values (ablation) | rlCoP with UCT · M2k | problems proved: 770 -> 1087 | Improved |
| Add learned value guidance without learned policy (ablation) | rlCoP with UCT · M2k | problems proved: 770 -> 856 | Improved |
| Replace iterative deepening with unguided random search (ablation) | bare prover · M2k | problems proved: 876 -> 434 | Hurt |
| Add UCT with simple goal-counting evaluation, but no learned policy or value (ablation) | bare prover · M2k | problems proved: 434 -> 770 | Improved |
| Add UCT with simple goal-counting evaluation, but no learned policy or value (ablation) | bare prover · Miz40 test | problems proved: 431 -> 804 | Improved |
| Use the jointly learned policy and value from iteration 5 rather than the 10-iteration result (ablation) | rlCoP · Miz40 test | problems proved: 1624 -> 1621 | No clear effect |
Closest papers by meaning
- Subgoal-based Demonstration Learning for Formal Theorem Proving (ICML 2024 · similarity 0.90)
- Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models (NeurIPS 2025 · similarity 0.89)
- DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search (ICLR 2025 · similarity 0.89)
- STP: Self-play LLM Theorem Provers with Iterative Conjecturing and Proving (ICML 2025 · similarity 0.88)
- Learning Formal Mathematics From Intrinsic Motivation (NeurIPS 2024 · similarity 0.88)
- Process-Verified Reinforcement Learning for Theorem Proving via Lean (ICLR 2026 · similarity 0.88)
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.
