"Neuron"pedia and ICA Lens

Neuronpedia released 8.6 million neuron dashboards and added ICA Lens support for 6 models, letting researchers find interpretable directions in the residual stream without dictionary training. The platform also imported explanations from five neuron-explaining platforms and introduced Token Lists, which show tokens associated with a neuron or latent and are directly linkable, with inference enabled for most ICA lenses and several hosted models. Neuronpedia said it offers a gated upload API for explanations available on request.

What's New Other Updates Newsletter Bug : Our email sender had a bug which caused some unsubscribe links to not work as expected. This is now fixed, our apologies. Browse 8.6 million neuron dashboards via the MLP Neurons https://www.neuronpedia.org/neurons release page, or jump to specific models here: Inference is enabled for several of these models, so you can do the usual activation testing, steering, search by topK, and more. Here's a TopK search for Qwen3 8B https://www.neuronpedia.org/search-topk-by-token?modelId=qwen3-8b&source=20&text=In+his+%27I+Have+a+Dream%27+speech%2C+Martin+Luther+King+Jr.+envisioned+a+future+where+people+would+be+judged+by+their+character%2C+not+the+color+of+their+skin.+This+speech+is+iconic+in+the+civil+rights+movement.&densityThreshold=0.5&ignoreBos=false&sortBy=frequency 's neurons: And here's a steer on a Qwen "airplanes" neuron https://www.neuronpedia.org/qwen3-8b/steer?saved=cmwa7g3gp000zveaeo159qnvz&source=24&index=2665&strength=152.25 : Neuronpedia now hosts imported explanations from five neuron-explaining platforms onto Neuronpedia, with permission from their authors and links back to the original sources: Here's the same Qwen3 8B "airplanes" neuron https://www.neuronpedia.org/qwen3-8b/24/2665 , as explained by three different sources: Notice in the above example that clicking certain explanations changes the top logits to different lists more relevant to that explanation or explanation method. These are "Token Lists", a new feature on Neuronpedia that which show lists of tokens associated with a neuron or latent. Token lists are directly linkable. For example, here's the link showing tokens found via the fitted Jacobian Lens for that same Qwen airplanes neuron: https://www.neuronpedia.org/qwen3-8b/24/2665?tokenlist=jacobian-lens:J-Lens https://www.neuronpedia.org/qwen3-8b/24/2665?tokenlist=jacobian-lens:J-Lens 🙋 Opportunity to Contribute : Want to upload your labels? We have a gated upload API for explanations, available upon request. Contact us mailto:johnny@neuronpedia.org?subject=Neuron%20Explanations and we'll set you up. ICA Lens revisits independent component analysis as a way to find interpretable directions in the residual stream, with no dictionary training needed. Neuronpedia hosts ICA Lenses for 6 models. Each lens is for one layer, and the number of directions per lens is the model's hidden size d model . Each direction has its own full dashboard, like SAE latents and neurons. Using the dropdown above the logits, you can switch to viewing the tokens through J-Lens or R-Lens as well. Most ICA lenses are inference enabled, too. For example, here's steering on a "legal system" direction https://www.neuronpedia.org/gpt2-small/8-icalens-res/51 in GPT2-Small steer link https://www.neuronpedia.org/gpt2-small/steer?saved=cmw8mpwvb001qhnlw96cvp9o5&source=8-icalens-res&index=51&strength=208.98 : Just like with SAEs and neurons, you can use Inference Search to find top directions in a model's ICA lenses using custom text. Here's an inference search with text about food in Italy https://www.neuronpedia.org/gemma-2-2b-it/?sourceSet=icalens-res&selectedLayers=%5B%5D&sortIndexes=%5B15,16%5D&ignoreBos=true&q=In%20the%20heart%20of%20Tuscany%2C%20a%20family-owned%20vineyard%20pairs%20its%20robust%20Chianti%20with%20a%20hearty%20wild%20boar%20rag%C3%B9%2C%20creating%20an%20unforgettable%20taste%20of%20Italian%20countryside%20cuisine. in Gemma 2 2B IT: Inspired by the previously featured NP-Feature-Compare https://np-feature-compare.netlify.app/?left=gemma-3-4b%2F9-gemmascope-2-res-16k%2F83&right=gemma-3-4b-it%2F9-gemmascope-2-res-16k%2F83 by Melissa Wessel, you can now compare two dashboards side by side on Neuronpedia directly. Here, we compare a neuron https://www.neuronpedia.org/gemma-4-12b/32/3?compare=gemma-4-12b-it/32/3 in Gemma 4 12B with the same neuron in its instruct variant: The Neuronpedia Workspace Lenses https://huggingface.co/neuronpedia/jacobian-lens Huggingface has been updated with R-Lens and J++ lens from Bobby Faber. Two J++ lenses are enabled for custom chats on Neuronpedia: gemma-3-12b-it https://www.neuronpedia.org/gemma-3-12b/jlens and llama-3.1-8b-it https://neuronpedia.org/llama3.1-8b-it/jlens . As always, please contact us https://www.neuronpedia.org/contact with your questions, feedback, and suggestions.