Institutional Repository of Vilnius University
Analysis of artificial neural networks for solving reasoning tasks /
Abstract
dc:descriptionThis thesis examines the use of artificial neural networks for automated theorem proving in the Lean 4 formal proof assistant. The central problem is semantic hallucination: models often produce proof steps that are syntactically correct but logically incorrect. A formal error probability framework is presented to quantify this problem at the step level. Phase I evaluates three models under identical conditions on 1,000 theorems from the LeanDojo Benchmark 4. InternLM2.5StepProver achieves 45.7% Pass@1, ReProver 44.6%, and DeepSeekProverv1.5RL 40.3%. The framework shows that 75–83% of syntactically valid tactics are logically incorrect across all models. Phase II proposes two modifications of InternLM2.5StepProver. Proposition A introduces a featurebased value head that prunes unpromising proof branches, achieving 51.9% Pass@1 (+6.2pp). Proposition B trains an errorconditioned refiner using Direct Preference Optimization on real Lean 4 compiler error pairs, achieving 45.9% Pass@1 with a 16.3% correction rate.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Sadi, Belkacem,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308157690&prefLang=en_US