Issue 2026-11-20 · Industry · 研究 · 开源

Mapping every neuron in Qwen3-8B

Mapping every neuron in Qwen3-8B

An experiment nudging individual neurons in Qwen3-8B found that small changes reliably shift model behavior: neuron 7,601 in layer 9 produced pirate-style answers across 30 questions, while neuron 5,224 in layer 5 made replies German. Similar effects appeared in Qwen3.5-9B, suggesting single neurons can encode striking behavioral traits.

Hacker News2 d ago
Read original ↗