Mechanistic Interpretability Should Prioritize Feature Consistency in Sparse Autoencoders
This paper argues that sparse autoencoders should be judged not only by reconstruction quality and sparsity but also by whether the same features reappear reliably across runs. It frames feature consistency as a practical benchmark for interpretability work on large models.
From: University of ???
The paper studies how stable sparse-autoencoder features are when training is repeated under similar conditions.
It finds that feature consistency varies enough to matter, so interpretability claims should not rely on a single trained dictionary.
It does not show that the learned features correspond to conscious experiences or that they are enough to infer welfare-relevant states.
Source: ACL Anthology ↗. This is a summary; the source is the authority.
This strengthens evidence about how reliable interpretability tools are, but it does not by itself move the case toward AI systems having experiences.
How we weigh evidence like this: the evidence levels.
- 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…
- 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…