engram-steering

The article explores Dynamic Abliteration, a non-destructive method to suppress refusal behavior in open-weight LLMs like Qwen3-4B at runtime without permanently altering model weights, using multi-layer engram steering via PyTorch forward hooks.