Every Ablation Is a Dose: Counterweights and the Semblance of Self-Repair
The paper argues that “self-repair” after ablation is often a fixed counterweight response already encoded in the model.
The authors model ablations as points on a signed intervention axis and say downstream repair follows an affine law.
Across Gemma, Qwen, LLaMA, and Mistral factual-verdict tests, 68 of 81 downstream directions fit that law.
On GPT-2 Small’s IOI circuit, seven reachable heads fit it, and all seven acted as counterweights.
ArXiv · AI/CL/LG's note
The authors model ablations as points on a signed intervention axis and say downstream repair follows an affine law.
Across Gemma, Qwen, LLaMA, and Mistral factual-verdict tests, 68 of 81 downstream directions fit that law.
On GPT-2 Small’s IOI circuit, seven reachable heads fit it, and all seven acted as counterweights.
ArXiv · AI/CL/LG's note
score 4