From Brittle LLM Code Reasoning to MaxSAT-Based Verified Repairs
Image credit: University College London (UCL)Abstract
In this talk, we examine the limitations of Large Language Models (LLMs) in semantic code reasoning, showing that their predictions may change under semantics-preserving code transformations, which suggests brittle rather than robust understanding of program semantics. We then present a neuro-symbolic program repair approach that combines model checking with MaxSAT-based fault localisation to identify faulty code regions and guide LLMs towards more accurate, verifiable repairs.
Date
May 20, 2026 1:00 PM — 2:00 PM
Event
Software Systems Engineering (SSE) Seminars @ UCL
Location
Computer Science Department,
University College London (UCL), London, UK.
