Research map / Large language models / Retrieval and knowledge
DAPS: Dependency-Aware Premise Selection for LLM Theorem Proving
Zixuan Chen, Wenyuan Jiang, Junling Wang, Mrinmaya Sachan, Yinya Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DAPS, a dependency-aware premise selector for LLM-based Lean 4 theorem proving that aims to surface newly contributed lemmas buried under popular foundational ones.
Problem
Premise selection for LLM theorem provers in Lean 4: retrieval skews toward frequently used foundational lemmas, so the right premises, especially newly contributed ones, are buried among hundreds of thousands of candidates.
Approach
Moves beyond semantic-similarity neural selectors, which have frequency-skewed retrieval, isolated pairwise ranking, and a third limitation cut off in the abstract; the dependency-aware approach is inferred from the title only and the details are not stated in the available text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace single-positive InfoNCE with the group-level contrastive objective (ablation) | DAPS · Mathlib4-Heldout | Recall@32: 56.00 -> 89.25 | Improved |
| add group-level contrastive objective (ablation) | DAPS · DAPS-Collection | Recall@32: 15.25 -> 23.53 | Improved |
| use DAPS with structural neighborhood encoding instead of LeanHammer | premise selector · Mathlib4-Heldout | Recall@32: 77.84 -> 89.25 | Improved |
| DAPS under competition-style query shift | premise selector · DAPS-Collection | Recall@32: 3.19 -> 23.53 | Improved |
| add Blueprint continual training after Mathlib pretraining | DAPS · DAPS-Blueprint | Recall@32: 40.42 -> 50.08 vs LeanHammer | Improved |
| use DAPS premises for Goedel-Prover-V2-8B | Goedel-Prover-V2-8B · miniF2F-test | pass@4: 49.90 -> 54.03 | Improved |
| use LeanHammer premises for Goedel-Prover-V2-8B | Goedel-Prover-V2-8B · miniF2F-test | pass@4: 49.90 -> 48.93 | Hurt |
| use DAPS premises for DeepSeek-Prover-V2-7B | DeepSeek-Prover-V2-7B · miniF2F-test | pass@1: 34.07 -> 38.11 | Improved |
| use DAPS premises instead of LeanExplore for Goedel-Prover-V2-8B | Goedel-Prover-V2-8B · miniF2F-test | pass@1: 40.37 -> 39.96 | Mixed |
| supply DAPS top-16 premises to general LLMs for informal proving | Qwen3-235b, DeepSeek-V3.1, kimi-k2, Gemini-3.6-Flash · IMO-ProofBench Pass@8 | grade: +0.15 to +3.18 over Qwen3-Embedding-8B | Improved |
| swap formalizer: Gemini-3-Pro to Goedel-Formalizer-V2-32B to manual formalization (ablation) | informal proving pipeline · IMO-ProofBench, Gemini-3.6-Flash, Pass@4 | grade: 46.55 -> 47.50 -> 48.27 | Improved |
| add DAPS structural features, rare-premise bucket | DAPS vs LeanHammer · Mathlib4-Heldout hit@32 rarest bucket | hit@32: 39.4 -> 50.1 | Improved |
Closest papers by meaning
- Magnushammer: A Transformer-Based Approach to Premise Selection (ICLR 2024 · similarity 0.90)
- BAIT: Benchmarking (Embedding) Architectures for Interactive Theorem-Proving (AAAI 2024 · similarity 0.89)
- LEGO-Prover: Neural Theorem Proving with Growing Libraries (ICLR 2024 · similarity 0.89)
- Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMs (ICML 2025 · similarity 0.89)
- APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning (NeurIPS 2025 · similarity 0.89)
- MA-LoT: Model-Collaboration Lean-based Long Chain-of-Thought Reasoning enhances Formal Theorem Proving (ICML 2025 · similarity 0.89)
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.
