EMNLP paper: fine-tuning on benign data quietly erodes LLM safety alignment

hanjie_chen · x · 2026-10-24

A paper accepted to EMNLP 2026 Main Conference (GR-SAP) examines a counterintuitive risk: fine-tuning an aligned LLM degrades its safety alignment even on fully benign data.
The standard fix—mixing original safety alignment data back in—rarely works because such data is seldom released. The authors analyze the problem theoretically and empirically, and propose an efficient, data- and model-agnostic method that uses generative replay, having the model synthesize its own safety data to preserve alignment.

More from Safety