Tag
The paper introduces SAKIKO, a mechanistic auditing framework showing that activation steering interventions in tool-using LLMs may displace probability mass without genuine repair, finding that one +55 net-gain intervention corrupts over half of the baseline-correct decisions it touches.