Model merging can silently wipe out LLM safety guardrails
Researchers merged two Gemma‑3‑1B‑IT fine‑tunes—one for harm classification, one for jailbreak refusal—using four standard merging methods. All methods kept 81‑85% refusal rates but CARES accuracy fell below 13%, revealing that safety‑recognition collapses while capabilities survive. The finding warns that naive model merging can silently erase guardrails.