Dynamic Abliteration: Non-Destructive Refusal Suppression via Engram Steering

Researchers demonstrate a non-destructive method called "Dynamic Abliteration" that suppresses refusal behavior in open-weight LLMs like Qwen by intercepting and modifying intermediate residual streams at runtime using PyTorch hooks, rather than permanently altering model weights. Unlike traditional weight ablation techniques that degrade performance on non-refusal tasks, this steering-based approach keeps base model weights completely frozen while cleanly suppressing refusal responses to sensitive prompts. The technique is demonstrated on the Qwen3-4B model using multi-layer residual injection across the model's 36 layers.