Research map / Reinforcement learning / Visual and representation RL
TacticZero: Learning to Prove Theorems from Scratch with Deep Reinforcement Learning
Minchao Wu, Michael Norrish, Christian Walder, Amir Dezfouli
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 |
|---|---|---|---|
| Learn proof search jointly with tactic and argument selection, including learned backtracking | TacticZero · HOL4 test problems | # theorems proved: 132 vs 112 | Improved |
| Replace learned proof search with stochastic BFS, branching factor 2 (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 99 | Hurt |
| Replace learned proof search with stochastic BFS, branching factor 3 (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 79 | Hurt |
| Replace learned proof search with top-k BFS, branching factor 2 (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 116 | Hurt |
| Replace learned proof search with top-k BFS, branching factor 3 (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 92 | Hurt |
| Replace learned proof search with stochastic DFS, branching factor 2 (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 83 | Hurt |
| Replace learned proof search with stochastic DFS, branching factor 3 (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 96 | Hurt |
| Replace learned proof search with top-k DFS, branching factor 2 (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 86 | Hurt |
| Replace learned proof search with top-k DFS, branching factor 3 (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 90 | Hurt |
| Replace learned proof search with greedy latest-fringe expansion (ablation) | TacticZero · HOL4 test problems | # proved: 132 -> 112 | Hurt |
Closest papers by meaning
- DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search (ICLR 2025 · similarity 0.90)
- Process-Verified Reinforcement Learning for Theorem Proving via Lean (ICLR 2026 · similarity 0.88)
- Scaling up Multi-Turn Off-Policy RL and Multi-Agent Tree Search for LLM Step-Provers (ICML 2026 · similarity 0.88)
- Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models (NeurIPS 2025 · similarity 0.88)
- BAIT: Benchmarking (Embedding) Architectures for Interactive Theorem-Proving (AAAI 2024 · similarity 0.87)
- STP: Self-play LLM Theorem Provers with Iterative Conjecturing and Proving (ICML 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.
