What problem does it solve? Choosing a steering coefficient that is too small drowns the causal effect in noise, while one that is too large collapses fluency and produces false conclusions about specificity. This Skill guides the systematic tuning of intervention strength (alpha, beta, dose, magnitude) so mechanistic interpretability experiments yield valid, evidence-backed results. ## Core Features & Use Cases - Coarse-to-Fine Sweep Strategy: Start with a wide geometric range (e.g. [1, 2, 4, 8, 16, 32]) and progressively narrow toward the optimum instead of trusting a single value copied from a paper. - Layer-Dependent Guidance: Accounts for the fact that the best coefficient depends on the intervention site, with mid layers typically carrying high-level control signals and late layers breaking into repetition or format spam. - Dual-Metric Scoring: Requires scoring every sweep point on both a target metric and a fluency/general-ability metric, keeping only Pareto-optimal candidates. - Use Case: When a steering vector appears to have no effect, or a random direction matches your steering vector, use this Skill to re-sweep the coefficient range before concluding the experiment failed. ## Quick Start Ask the agent to sweep steering coefficients for your intervention across a wide geometric range and score each point on both the target behavior and a fluency metric.
