From Geometric Recovery to Causal Validation: A Reproducible Audit of Sparse Autoencoder Features, from Superposition Geometry to Causal Inertness
This paper audits sparse autoencoder features using both geometric recovery and causal tests. It argues that correlational success alone is not enough to validate interpretability claims.
From: arXiv
The authors compare common recovery metrics with stronger causal checks on SAE features.
They conclude that some features that look good geometrically may be causally inert or weaker than assumed.
It does not show that any internal feature corresponds to conscious content or that interpretability can solve the welfare question.
Source: arXiv mirrored HTML ↗. This is a summary; the source is the authority.
This makes interpretability evidence more cautious and less overclaimed, but it does not support claims of AI experience.
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 …
- 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…