Goodfire Research Finds Neuron Communities in Language Models
Start here
Goodfire Research reports that small groups of neurons — often fewer than ten — implement common language-model behaviors such as grammatical agreement, copying and arithmetic. The group used gradient-based attribution over millions of prompts, then built graphs and applied community detection, without supervision. It ran the method on Llama-3.1-8B with 10M prompts from FineWeb-Edu, English Wikipedia and FineMath, and later replicated it on Qwen 3.6 35B-A3B. Ablating a community changed the model's predictions, while keeping it largely preserved them even after other MLP neurons in those layers were removed. Goodfire says its sufficiency claims are limited to the six layers that contain community neurons, and that the explanation remains partial: many predictions are still unexplained, and identifying a necessary and sufficient community does not yet show how its neurons compute or interact with the rest of the model.
Generated from reports · updated 6 h ago
Report timeline
Follow the reports to see the event from different sides.
- Goodfire ResearchGoodfire Research discovers neuron communities in language models
Goodfire Research has identified "neuron communities" — small groups, often fewer than ten neurons — that implement common model behaviors such as grammatical agreement, copying and arithmetic. Using gradient-based attribution across millions of prompts, graph construction and community detection, the unsupervised method was run on Llama-3.1-8B and replicated on Qwen 3.6 35B-A3B. Ablating a community changes predictions, while keeping it largely preserves them even after removing other MLP neurons in those layers.
Heat for this event
- Current, comparable range
- 4
- Peak, comparable range
- 5November 18 03:00
- Change in 24 hours
- –
The trend compares the same participants over hours observed continuously, so its range can be narrower than the current heat count. Hover or tap the chart for hourly heat; the arrow keys move through it.