Neurons as an interpretable foundation
Neurons as an Interpretable Basis
Goodfire Research studies the neurons of Qwen3.6-35B-A3B and proposes that neurons themselves can serve as interpretable computational units. After the team ablated the weakest 92.2% of MoE neuron activations in each forward pass, language model loss increased by only 2.9%; 69% of neuron poles are dominated by a single pattern for at least 90% of the KL damage. After describing neurons in natural language, a judgment model can predict with 95% accuracy whether they activate on given text, and recovers 73% of the ablation damage in high-activation regions; 94% of neurons can be activated by synthetic text to levels exceeding the highest in a corpus of more than 10B tokens.
The study examines neurons as interpretable units on Qwen3.6-35B-A3B and provides quantitative results on sparse activation and description-predicted activation.