Goodfire finds language model neurons are sparse and describable, publishing 1,500 neuron descriptions for Qwen3.6
The interpretability lab's technical report argues MLP neurons are interpretable units of computation: ablating the weakest 92% of activations has only modest effects on capabilities, and short descriptions let a judge model predict when a neuron fires.
Key facts
- Report: "Neurons as an Interpretable Basis", 10 November 2026
- Authors: Curt Tigges, Matthew Kowal, Thomas McGrath, Owen Lewis, Nick Cammarata
- Model studied: Qwen3.6 35B-A3B
- Descriptions released for 1,500 neurons
- Ablating the weakest 92% of activations has only modest effects