Constructing Interpretable Features from Compositional Neuron Representations in Language Models
This paper proposes an alternative to standard sparse autoencoders by decomposing MLP activations with semi-nonnegative matrix factorization. It argues that the resulting features are more directly tied to the model’s computations and can be easier to interpret causally.
From: ACL Anthology
The authors compare their factorization approach with sparse autoencoders and a supervised baseline across several language models.
They report stronger causal steering performance and better alignment with human-interpretable concepts than the comparators.
It does not show that these features are introspective reports from the model itself, nor that causal interpretability implies subjective experience.
Source: ACL Anthology ↗. This is a summary; the source is the authority.
This improves evidence that internal features can be extracted and manipulated, but it still falls short of evidence for experience or welfare.
How we weigh evidence like this: the evidence levels.
- Mechanistic Interpretability Should Prioritize Feature Consistency in Sparse AutoencodersThis paper argues that sparse autoencoders should be judged not only by reconstruction quality and sparsity but also by whether the same fea…
- Mechanistic Interpretability for Neural NetworksThis is a broad survey of mechanistic interpretability, covering the goal of reverse-engineering internal algorithms in neural networks. It …
- Interpretability Transfer from Language to Vision via Sparse Autoencoder Transfer AlignmentThis work transfers labeled sparse-autoencoder features from language to vision inside a multimodal model. It uses those features to make vi…
- Multilingual Sparse Autoencoders and Principled Layer Selection for Cross-Lingual Steering in Large Language ModelsThis paper studies sparse autoencoders for feature-level interpretability and steering in multilingual language models. It also proposes a m…