Goodfire finds most Qwen neurons interpretable without feature extraction
Goodfire’s study of Qwen3.6-35B-A3B suggests researchers can interpret individual MLP neurons directly, instead of first translating the model into learned feature spaces. Removing the weakest 92.2% of mixture-of-experts neuron activations raises language-model loss by only 2.9%. Agent-generated descriptions predict when a neuron fires with 95% accuracy and recover 73% of the ablation damage in simulation, making those descriptions useful for testing and steering model behavior.