Mechanistic Interpretability for Neural Networks
This is a broad survey of mechanistic interpretability, covering the goal of reverse-engineering internal algorithms in neural networks. It discusses sparse autoencoders, transcoders, and the problem of superposition in modern models.
From: arXiv
The article synthesizes existing methods and explains how interpretability tools try to decompose tangled activations into simpler features.
It presents sparse autoencoders and related methods as central approaches for making internal representations more legible.
It does not establish that these methods can tell whether a model is conscious, only that they can help analyze internal structure.
Source: arXiv ↗. This is a summary; the source is the authority.
This mainly sharpens the technical toolkit for studying models, while leaving the consciousness question open.
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…
- Constructing Interpretable Features from Compositional Neuron Representations in Language ModelsThis paper proposes an alternative to standard sparse autoencoders by decomposing MLP activations with semi-nonnegative matrix factorization…
- 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…