Research map / Label-efficient and robust learning / OOD and anomaly detection

Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift

NeurIPS 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Robust fine-tuning for OOD generalization · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Geometry-Adaptive Explainer (GAE), a training-free closed-form post-hoc method that realigns an ID-trained sparse autoencoder/transcoder decoder with the OOD-active subspace using unlabeled OOD activations, improving causal faithfulness under distribution shift.

Problem

Dictionary-based explainers (SAEs, transcoders) trained on in-distribution activations lose faithfulness under out-of-distribution shift because the model's active subspace rotates away from the explainer's dictionary.

Approach

Defines a faithfulness gap as the Frobenius distance between the explainer subspace and the OOD-active subspace, and closes it with a Procrustes rotation of the decoder onto the OOD second-moment eigenspace followed by a closed-form constrained ridge refit of the decoder, with the encoder frozen and no gradient updates.

Key results

• On GPT-2 Small with transcoders, GAE leads all nine nAOPC/nComp/|ΔCE| columns, e.g., HaluEval nComp 0.963 vs SAEBoost 0.921 with |ΔCE| 0.0014 vs 0.0212.
• GAE is training-free and completes adaptation in 0.5 s (GPT-2) and 2.9 s (Pythia-1.4B) versus minutes to hours for training-based baselines.
• Step 1 closes the subspace gap to zero by construction, but Step 2 is needed to recover nComp (0.96 vs 0.44 after Step 1 alone on HaluEval).

Surprising findings

• Step 1 alone (pure subspace rotation) drops nComp from 0.74 to 0.44 despite eliminating the faithfulness gap, showing alignment alone does not ensure causal faithfulness.
• Gradient-based Finetune does not reduce the subspace gap, indicating reconstruction loss does not drive subspace alignment.

Limitations

• Evaluated only on GPT-2 Small and Pythia-1.4B; larger models not tested.
• Relies on top-r SVD truncation of the ID decoder, discarding information in residual singular directions.
• Pythia-1.4B |ΔCE| on FineWeb is slightly worse than Finetune.

Experimental setup

Datasets: FineWeb Edgar HaluEval

Benchmarks: FineWeb Edgar HaluEval

Metrics: nAOPC nComp delta CE

Baselines: Fixed TERM Finetune Retrain SAEBoost FaithfulSAE

Builds on: sparse autoencoders transcoders Top-K SAE Davis-Kahan theorem

Closest papers by meaning

• Evaluating the Robustness of Interpretability Methods through Explanation Invariance and Equivariance (NeurIPS 2023 · similarity 0.87)
• Efficient and Accurate Explanation Estimation with Distribution Compression (ICLR 2025 · similarity 0.85)
• Faithful Model Explanations through Energy-Constrained Conformal Counterfactuals (AAAI 2024 · similarity 0.84)
• Compact Proofs of Model Performance via Mechanistic Interpretability (NeurIPS 2024 · similarity 0.84)
• A New Approach to Backtracking Counterfactual Explanations: A Unified Causal Framework for Efficient Model Interpretability (ICML 2025 · similarity 0.84)
• Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? (ICLR 2025 · similarity 0.84)