Neuron Explorer for Qwen3.5-4B: See every neuron at work, make targeted edits

Neuron Explorer is an open-source toolkit for investigating and steering individual MLP neurons in Qwen3.5-4B. Rather than relying only on passive activation measurements, it pushes each neuron up or down across selected prompts, records the resulting model rollouts, and presents them in an interactive view of the model. The accompanying sweep scripts calculate per-neuron activation statistics, test effects on next-token probabilities, and compare units with random-direction controls. The project supplies commands for running these experiments on a CUDA GPU pod, along with tools to rank neurons, merge runs, check steering consistency, and build visualizations.

The browser-based explorer can load the collected runs and statistics, or load the model locally for live inference on a user’s text. Its map places neurons across layers; lists rank units by word-cluster signals, lens results, or activation on supplied text; and neuron pages show steered outputs, comparisons between negative, neutral, and positive steering, and connections to other units. An embedding view organizes rollout patterns, while an optional Claude CLI integration can inspect neurons and run model operations through the explorer. The repository also includes cached layouts and connection data, plus research notes. It is designed to make targeted interventions and their behavioral effects inspectable; the supplied description outlines the experimental and browsing tools but does not state particular research conclusions.