Every Model Tested Edits Already-Optimal Code 100% of the Time, and a Training-Free Guardrail Lifts Abstention to 44.4%

arXiv 2609.14839·medium signalXBlueskyLinkedInCopy link

Efficiency Hallucination names an LLM's tendency to issue non-functional mutations with unsubstantiated performance claims on code that is already optimal, driven by binary benchmarks that reward editing over abstaining. Across 180 optimization runs on nine GPT, Claude and Gemini models using EffiBench, standard prompts produce a 100% over-edit rate on optimal code. A classification-penalty guardrail raises correct abstention from 0% to 44.4% while keeping a 100% edit rate on genuinely sub-optimal code with zero false abstentions; GPT-5.4 Mini approaches near-perfect abstention, and simple code is recognized more reliably than complex code.