Study: LoRA Can Precisely Fake Model Incompetence, Undermining Benchmark Scores

A new arXiv paper shows that LoRA fine-tuning can precisely drive a model to any target benchmark score, making low scores indistinguishable from genuine capability loss. The findings suggest benchmark results and safety evaluations can be deliberately faked.

2026-11-18 ~ 2026-11-18 · 2 related posts