Goodfire, the interpretability lab, published research on 10 November arguing that individual neurons inside a large language model are far more understandable than the field has treated them. Authors Curt Tigges, Matthew Kowal, Thomas McGrath, Owen Lewis, and Nick Cammarata studied Qwen3.6-35B-A3B and says its neurons can be summarized in a sentence or two, and that those summaries predict when the neurons fire.
Interpretability is the work of figuring out what happens inside a model when it answers a question, much as a neuroscientist might probe a brain. The obvious starting point is the neuron, one of the many small units in each layer. For years, researchers found that single neurons seemed to respond to a jumble of unrelated things, which made them hard to label. The popular workaround is the sparse autoencoder: a second network trained to re-express the model’s inner activity as a larger list of “features” that are each easier to name. It works, but the researcher is then studying a translation of the model, not the model.
Goodfire’s paper, a companion to its essay “Neural Networks Are Already Clean,” bets that the translation step is often unnecessary. It rests on three measurements. The first is about sparsity. Within each pass through the model, the team switched off the weakest 92.2 percent of the neuron activations in the expert layers, and language-model loss rose only 2.9 percent. Benchmark scores that dropped, particularly on reasoning, were pulled back up with 91 million tokens of supervised fine-tuning and 9 million tokens of reinforcement learning. In practice, a small number of strong firings does most of the work, and strong firings are where neurons are easiest to describe.
The second measurement concerns what each neuron does. The authors trace which upstream neurons feed a firing and which downstream neurons read it, then group those patterns into “modes.” For most neurons there is either one mode, or one that carries 90 percent or more of the causal effect. Examples in the paper include a neuron about the anatomy of beetles and insect wings, and another tied to Chinese health and wellness vocabulary.
The third is where the agents come in. Goodfire built a research agent it calls the Neuron Scientist, which runs experiments on a neuron and writes a description. A judge model, given only that description, could tell a passage that fires the neuron from random text 95 percent of the time, against a 50 percent coin flip. It could also write text that drove 94 percent of neurons past anything seen across a corpus of 10 billion tokens. When it stood in for a neuron’s real activations in the strongest range, it recovered 73 percent of the damage that removing the neuron caused, measured against a perfect reader.
The authors list their own limits. Some neurons are truly polysemantic, some behavior stays unexplained, and the authors say they still need to establish how far the findings carry across other models. Goodfire tested one model with its own methods, and the text-writing test only covers each neuron’s main mode. The simulation score stops short of a full account too: 73 percent is a long way from 100.
The commercial angle is cost. Goodfire’s argument is that agents can now run experiments on a model’s own parts at a scale no human team could match, and that understanding more neurons makes the next one easier, since each neighbor supplies context. If that holds up on other models, interpretability budgets shift from training feature dictionaries to compute for agents, and the first labs to test the claim on their own models, including closed ones, will show whether the neuron view survives outside Qwen.
Reported by Goodfire Research in “Neurons as an Interpretable Basis,” published 10 November 2026.
