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

A new neuron explorer tool has been launched for the Qwen3.5-4B model, enabling researchers to visualize and manipulate the neural activations of the model in real-time. This steering-first approach allows users to interactively adjust individual neurons within the model's layers and view the outcomes across a 3D visualization. Users can engage with various features such as neuron activation filtering, UMAP embeddings for layer rollouts, and a chat interface powered by Claude’s command line interface, enhancing their ability to analyze and understand model behavior under different prompts. This development is significant for the AI/ML community as it provides a novel way to study and refine model dynamics at a granular level, fostering better interpretability and control over neural networks. The interconnected tools—including scripts for neuron statistics, steering sweeps, and depth analysis—help generate comprehensive insights into neuron performance and clustering. Such capabilities are crucial as the field seeks to improve model reliability, mitigate biases, and ultimately advance the understanding of complex AI systems.