Tag: arxiv

Efficiency Hallucination: Every Model Rewrote Code That Couldn’t Get Faster

A Columbia paper found LLMs edit already-optimal code 100% of the time. One prompt line halves it with zero false refusals. I reproduced it and…