ft-fine-tuning-catastrophic-multihop-failure
IN premise — summaries/2026/08/24/zhong-2023-mquake-sR-references.md
Created 2026-08-25T02:59:08+00:00
Fine-tuning (FT) on layer 21/31 yields 0–2.8% multi-hop accuracy compared to approximately 40% for the unedited base model.
Summary
Fine-tuning a model on just two specific layers (21 and 31) essentially destroys its ability to chain together multi-step reasoning, dropping performance from roughly 40 percent correct to nearly zero. This matters because it shows that a routine and widely used adaptation technique can silently and catastrophically break a model's core reasoning capability, meaning any pipeline that fine-tunes must actively verify that multi-hop performance survives the change.