LLM Bug Fixes Tend Toward Reimplementation Rather Than Patching

Analysis of GPT-5-series models on Codeforces datasets shows a bias toward generating entirely new solutions instead of applying minimal patches. The study compares LLM output similarity to human-written fixes to measure deviation from original buggy code.