Neuron Lens adds circuits view: tracing how activations flow across features in Qwen3 and Gemma
burny_tech · x · 2026-11-12
Interpretability tool Neuron Lens now shows circuits — how activations flow across features — and the authors found these flows are often very legible, i.e. the internal activation pathways can be presented in human-understandable form.
Neuron Lens is a neuron/attention-head explorer for Qwen3-8B and Gemma 4 12B: every unit gets an address (e.g. /qwen3-8b/L8/n344 is neuron 344 of layer 8), drawn in-browser with JavaScript, showing what each neuron and attention head responds to and what it makes the model say. The full per-unit data catalog is hosted on Hugging Face.
More from Research
- Verona Declaration sets principles for generative AI use in computing research — luislamb · 2026-11-12
- Intology Founder at YC Paper Club: Drug Failures Stem From Picking Wrong Targets, Not Bad Molecules — yusufroohani · 2026-11-12
- DeepMind's FlipTest benchmark: best video world models score below random guessing — AlexGDimakis · 2026-11-12
- AI Predicts Tooth Decay Risk from Brazilian Survey Data for Public Health Planning — Brighter-Side-News · 2026-11-12
- Meta FAIR open-sources e-means: 10–45x faster k-means with lower loss than Lloyd's algorithm — generativist · 2026-11-12
- Google and Artificial Analysis launch MKA benchmark to verify agent deliverables — OnlineInference · 2026-11-12