One or two MLP neurons predict whether a model is about to skip a tool call — enough to cut over-calling 80% and raise tool-required accuracy 14.2 points
PRISMS finds that three consequential tool-use failures — invalid arguments, unnecessary calls, and omitted calls — are each separable by a small, failure-specific set of MLP neurons with linear decision boundaries, then shares that neuron basis between a sparse detector and activation steering. Across six models from the Qwen3, Llama and Gemma families, over-calling and missing calls are detected at the pre-generation prompt boundary at ROC-AUC 0.90–1.00 using just 1–2 neurons for missing and 2–16 for over-calling (validity needs about 128 and is read from the generated span), matching or beating dense residual-stream baselines with 23–627x fewer features. Because detection happens before generation, intervention can be gated on predicted risk rather than applied unconditionally, cutting pooled over-calling from 0.131 to 0.026 while lifting tool-required accuracy from 0.689 to 0.831 — actionable only if you run open-weight models with activation access.
↳ Follow the thread