New Paper Steers LLM Tool-Call Rates

This is a compelling inference-time control knob for agent cost and reliability, though it is closer to behavioral steering than a complete tool-use policy.

• –Difference-of-means activation vectors are extracted from the model’s own tool-use propensity signal, requiring no fine-tuning, labels, gradients, or prompt changes.
• –A shared direction generalizes across six unseen tools while primarily changing whether the model calls something—not which tool it selects.
• –Induced calls concentrate on obscure questions the model is unlikely to know, producing a useful cost–accuracy Pareto frontier.
• –The approach transfers across dense, MoE, and multimodal models, but strong positive steering can create malformed calls and does not improve tool-execution quality.
• –The authors publicly release implementation code at [GitHub](https://github.com/YuqiChen4188/Steering-Tool-Use-Propensity).