Sungjun Lim, Heedong Kim, Andrew Lee, Kyungwoo Song
Mechanistic interpretability aims to explain a model's behavior by identifying causally responsible internal structures. Dictionary-based explainers such as sparse autoencoders and transcoders are a primary tool, but their faithfulness under out-of-distribution (OOD) shift has received little systematic attention. We show that distribution shift rotates the subspace that the model actively uses, misaligning the explainer's dictionary trained on in-distribution (ID) activations. We formalize this misalignment as the faithfulness gap, a geometric distance between the ID dictionary and the OOD-active subspace, and show that it bounds excess OOD projection loss. To reduce this gap, we propose the Geometry-Adaptive Explainer (GAE), which realigns the dictionary with the OOD-active subspace while preserving the original sparse feature activations. This requires only unlabeled OOD activations and no gradient updates. For the adapted dictionary, we also bound the excess projection loss caused by any remaining faithfulness gap. Empirically, GAE even matches or surpasses all training-based baselines in causal faithfulness across multiple models and OOD settings.
Sumin Park, Taero Kim, Minhyeong Cho, Subeen Park, Jong Hwan Lee, Kyungwoo Song
Learning representations that are robust to nuisance factors is essential for reliable deployment, but remains challenging when the representations lie on non-Euclidean manifolds. Existing debiasing methods typically assume Euclidean representations or apply linear interventions, which can distort the geometry of structured latent spaces. We propose IsoAct, a post-hoc framework that accounts for geometry in both nuisance identification and representation editing. It identifies nuisance-predictive directions using Fisher discriminant analysis (FDA) in tangent coordinates and applies corrections through manifold-native actions that preserve manifold membership. Controlled and real-world experiments on hyperspherical, hyperbolic, and $\mathrm{SE}(3)$ representations evaluate nuisance reduction, utility preservation, and manifold validity jointly. IsoAct reduces nuisance recoverability while retaining competitive utility. Its clearest joint result is on KLIFS, combining reduced kinase-group recoverability, strong structural prediction, and valid rigid poses.