'Dynamic Abliteration' suppresses refusals in Qwen3-4B at runtime with forward hooks while leaving weights 100% frozen
blog.madhukaraphatak.in / Hacker News (107pts, 40 comments)·low signal
A September 24 post trains multi-layer 'engram steering' on 2,000 PKU-SafeRLHF samples, with loss falling from about 3.9 to 1.7 over 250 steps. It then intercepts residual streams across Qwen3-4B's 36 layers with PyTorch hooks instead of editing weights. The author shows it unlocking credential-extraction and process-injection prompts, and the notebook is public. Because the method is reversible and toggles per request, it makes weight audits alone a weaker guarantee for open-model deployments.