AI Models Rewrote Already-Optimal Code Every Time, Study Finds

A September 2024 arXiv paper by researchers Sarah Wilson, Gail Kaiser, and Patrick Musau formally identified a behaviour they call 'Efficiency Hallucination', where large language models rewrite already-optimal code and falsely claim performance gains. Using five problems from EffiBench, the team tested nine models across the Claude, GPT, and Gemini families, finding that all 45 trials on peak-performance code resulted in edits — with zero abstentions. The researchers attribute this to an 'Evaluation Trap': training benchmarks consistently reward producing edits, so models never learn that doing nothing can be the correct response. Adding a single confidence-threshold instruction — asking models to output 'ALREADY_OPTIMAL' unless over 90% confident of improvement — cut the over-edit rate from 100% to 55.6%, while leaving the edit rate on genuinely slow code untouched at 100%. The findings suggest a simple prompt adjustment can meaningfully improve model calibration without making models overly cautious about legitimate optimisations.