Back to search

Institutional Repository of Vilnius University

Analysis of artificial neural networks for solving reasoning tasks /

Abstract

dc:description

This 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 × 1

Rights

dc:rights
Statement dc:rights
  • info:eu-repo/semantics/openAccess
Language dc:language
eng

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
Vilnius University
Base URL
epublications.vu.lt/oai
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Sadi, Belkacem,. Analysis of artificial neural networks for solving reasoning tasks /. Institutional Repository of Vilnius University, 2026. https://repository.vu.lt/VU:ELABAETD308157690&prefLang=en_US