Language Model Circuits Are Sparse in the Neuron Basis | Transluce AI

We introduce a new approach to finding sparse circuits in the neuron basis, without relying on learned features.

Language Model Circuits Are Sparse in the Neuron Basis Aryaman Arora * , Zhengxuan Wu * , Jacob Steinhardt , Sarah Schwettmann * Equal contribution. Correspondence to: aryaman@transluce.org, zen@transluce.org. Transluce | Published: November 20, 2025 Many interpretability methods rely on learned feature bases—such as sparse autoencoders or cross-layer transcoders—based on the belief that neurons do not cleanly decompose model computation. We revisit this assumption and show that, with a better choice of neuron basis (MLP activations) and a stronger attribution method (RelP), raw neurons can pr

related reading

• Circuit Tracing: Revealing Computational Graphs in Language Modelstransformer-circuits.pub
• On the Biology of a Large Language Modeltransformer-circuits.pub
• Weight-Sparse Circuits May Be Interpretable Yet Unfaithful — LessWronglesswrong.com
• Towards Monosemanticity: Decomposing Language Models With Dictionary Learningtransformer-circuits.pub
• Transformer Circuits Threadtransformer-circuits.pub
• An Extremely Opinionated Annotated List of My Favourite Mechanistic Interpretability Papers v2 — AI Alignment Forumalignmentforum.org
• A Mathematical Framework for Transformer Circuitstransformer-circuits.pub
• Toy Models of Superpositiontransformer-circuits.pub
• Transluce on X: "Is your LM secretly an SAE? Most circuit-finding interpretability methods use learned features rather than raw activations, based on the belief that neurons do not cleanly decompose computation. In our new work, we show MLPx.com
• Zoom In: An Introduction to Circuitsdistill.pub
• Language models can explain neurons in language modelsopenaipublic.blob.core.windows.net
• Sparse Attention Post-Training for Mechanistic Interpretabilityarxiv.org