Directional Ablation, and How Astra Fixed Qwen3.5-0.8B's Reasoning Loops
How an LLM edited 0.016% of Qwen3.5-0.8B’s weights to break its reasoning loops: directional ablation from the refusal-direction work, then pushed further with a strength α and 64 selected neurons per layer.