Identifying Intervenable and Interpretable Features via Orthogonality Regularization

2026-02-04

AI Summary

This work addresses the challenge of entangled, superimposed, and poorly interpretable feature representations in language models by proposing an orthogonality-regularized fine-tuning approach for sparse autoencoders. Specifically, the decoder matrix is decomposed into an approximately orthogonal set of feature bases, grounded in the principle of independent causal mechanisms. The orthogonality constraint enhances feature uniqueness and modularity, while feature identifiability and interpretability are further improved through embedding distance metrics and causal intervention analysis. This enables more precise and isolated manipulation of individual features. Experimental results demonstrate that the proposed method yields substantially clearer and more controllable internal representations, with negligible degradation in downstream task performance.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

With recent progress on fine-tuning language models around a fixed sparse autoencoder, we disentangle the decoder matrix into almost orthogonal features. This reduces interference and superposition between the features, while keeping performance on the target dataset essentially unchanged. Our orthogonality penalty leads to identifiable features, ensuring the uniqueness of the decomposition. Further, we find that the distance between embedded feature explanations increases with stricter orthogonality penalty, a desirable property for interpretability. Invoking the $\textit{Independent Causal Mechanisms}$ principle, we argue that orthogonality promotes modular representations amenable to causal intervention. We empirically show that these increasingly orthogonalized features allow for isolated interventions. Our code is available under $\texttt{https://github.com/mrtzmllr/sae-icm}$.

Problem

• intervenable features
• interpretable features
• orthogonality
• causal intervention
• feature disentanglement

Innovation

• orthogonality regularization
• interpretable features
• causal intervention
• sparse autoencoder
• modular representation