Refusal as a LoRA: abliteration becomes a runtime layer for self-hosted LLMs
A Reddit post shows how for Qwen3.8-Flash-Next the refusal direction, largely rank-1, compresses into a small LoRA applied at runtime on the existing quantization. llama.cpp activates it per request, separating refusal behavior from model weights. Th...