Fugu-MT: arxivの論文翻訳

DAPS: Dependency-Aware Premise Selection for LLM Theorem Proving

Authors Zixuan Chen, Wenyuan Jiang, Junling Wang, Mrinmaya Sachan, Yinya Huang
Affiliations ETH Zurich
Categories Method / Premise Selection / Dependency-aware neural premise ranking, Application / Automated Theorem Proving / LLM-based theorem proving in Lean 4, Evaluation / Model Adaptation / Adaptation procedure for theorem proving datasets
License CC BY 4.0

Abstract Overview

DAPS is a premise selector for Lean 4 theorem proving that uses dependency structure to identify potentially useful lemmas. It combines a structural neighborhood encoder, a group-level contrastive objective, and adaptation from Mathlib4 to research-project Blueprints. The authors construct a typed Mathlib4 dependency graph, mine 53 Blueprint projects, and evaluate retrieval on four benchmarks spanning different distributions. DAPS outperforms the tested selectors on retrieval and improves downstream proof results, although the extent of improvement depends on the prover and evaluation setting.

Novelty

The distinctive contribution is the combination of typed dependency neighborhoods and group-level premise supervision with Mathlib-to-Blueprint adaptation. The work also systematically mines Lean Blueprint projects to provide dependency data containing informal and research-level statements beyond Mathlib4.

Results

DAPS achieves mean Recall@32 of 89.25 on Mathlib4-Heldout, 11.41 percentage points above LeanHammer, and leads the tested baselines on all four retrieval benchmarks. On miniF2F, its premises improve proof success over using no premises in all six tested settings; on IMO-ProofBench, they produce the highest autograded scores among the compared selectors for all four tested language models.

Key Points

  1. The dependency resources include a Mathlib4 graph with more than 275,000 nodes and 6.5 million edges, plus 3,805 nodes from 53 Blueprint projects.
  2. Ablation of the group-level objective in favor of single-positive contrastive training substantially reduces retrieval recall.
  3. Informal theorem proving uses a formalizer to convert the query into Lean before retrieval; generated informal proofs are evaluated by an autograder rather than Lean verification.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.