Dynamic Abliteration: Non-Destructive Refusal Suppression via Engram Steering

Dynamic Abliteration suppresses refusal behaviors in models like Qwen3-4B by using PyTorch forward hooks to intercept intermediate residual streams at runtime. Unlike traditional orthogonal weight projection, this multi-layer steering approach leaves base model weights frozen to prevent performance degradation in non-refusal tasks.