A single linear direction in the residual stream dials an agent's tool-call rate from ~0% to over 90% without retraining or prompt changes

arXiv 2608.25198·medium signalXBlueskyLinkedInCopy link

The authors extract a steering direction from the model's existing tool-use preference signal and apply it at inference, producing monotonic control over how often the agent reaches for a tool while keeping invocations valid. Open-domain QA accuracy with live tool execution nearly doubled, 0.29 to 0.56, by tuning the rate rather than the prompt. It generalizes to unseen tools and works across dense, mixture-of-experts, and multimodal architectures, which makes it a cost-versus-accuracy knob you can set per deployment instead of relitigating in the system prompt.