Canonical Schema Views for Activation Steering Under Tool-Schema Churn: A Negative Result

Abstract

Tool-calling agents increasingly rely on external APIs, yet these APIs undergo frequent schema changes---parameter renamings, restructurings, and deprecations---that can degrade agent performance. Activation Steering via Abstention (ASA) offers a training-free approach to improve function-calling accuracy through representation engineering, but its reliance on schema-specific steering vectors raises concerns about robustness to such churn. We investigate whether schema canonicalization---mapping diverse schemas to a unified lexical form before ASA processing---can provide churn-invariant steering. Our experiments reveal a decisive negative result: DelexGate-ASA degrades clean-schema AST accuracy by 15.2 percentage points (from 81.4% to 61.6%) while providing no robustness benefit under schema perturbations. This failure stems from the removal of semantic information in parameter names, which LLMs rely upon for correct argument-value mapping. We further demonstrate that ASA assets trained on clean schemas transfer effectively to perturbed schemas without recalibration, suggesting the steering vectors themselves are inherently robust to lexical churn.