Neuron Lens: A database of every neuron and attention head in Qwen3-8B and Gemma 4 12B
CatAstro_Piyush · x · 2026-10-31
- Two researchers released Neuron Lens, a searchable database covering every neuron and attention head in Qwen3-8B and Gemma 4 12B, with an in-browser explorer and a full data catalog on Hugging Face.
- The core method is the R-lens (a J-lens variant) plus their own S-lens: reading neurons' input/output projections makes most neurons' functions legibly interpretable, and the lens qualitatively predicts steering outcomes over 90% of the time.
- The same trick reads attention heads' QK and OV circuits, revealing both what they attend to and what they do with it. Known flaws: accuracy degrades in early layers and it can't separate the MLP block from the attention block; the authors say they found workarounds.
- Every unit is addressable by path (e.g. /qwen3-8b/L8/n344), with evidence and labels published openly.
More from Research
- Lacuna's Standing On traces any repo's research roots — nanoGPT stands on 140 papers — hugo_larochelle · 2026-10-31
- Three Mathematicians Join Lex Fridman to Discuss AI Solving Navier-Stokes, Quasi-Riemann — khademinori · 2026-10-31
- ANSER-Bench: 2,235 analytical search tasks trip up SOTA Deep Research in ~50% of cases — CShorten30 · 2026-10-31
- Stanford: AI reads routine ECGs to flag sudden cardiac death risk in 86% with no known warning signs — oziadias · 2026-10-31
- Zip-NeRF Gets Orders of Magnitude Faster to Train, but Rendering Still Crawls at 3-4 FPS — jonstephens85 · 2026-10-31
- Researchers reconstruct Meta Muse's hidden charter from 884 chain-of-thought samples — sethlazar · 2026-10-31