Stable Directions, Different Effects: Intervention Fidelity Across Model Precision
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Activation interventions can be characterized on one numerical realization of a language model and later applied to a quantized version. Preserving standard model quality or direction-level representation similarity does not guarantee that the intervention effect is preserved. We formalize this as cross-precision intervention fidelity and define a signed interventionfidelity gap between native and transferred steering effects. Across several language models, precision changes, model sizes, steering estimators, and prospectively held-out concepts/tasks, we compare a prespecified directionalignment/simple-geometry baseline with local functional sensitivity. Intervention effects can change even when directions remain strongly aligned, whereas a standard first-order local approximation tracks the signed gap closely in the strongest held-out evaluations. The novelty is not the Taylor expansion itself, but the intervention-fidelity problem and the empirical accuracy of this local account across the tested replication axes. A correction frozen before fresh evaluation improves reconstruction of native intervention effects and a prespecified sign-based decision. Finally, a prospective parsimony test favors the compact first-order approximation over the richer predictor. Within the tested normalized-transfer regime, intervention reliability should therefore be evaluated functionally rather than inferred from ordinary quantization quality or direction similarity.
- Primary area:
- interpretability and explainable AI