Steering a single neuron gives Qwen3-30B a new identity in ~1-2% of replies
Sauers_ · x · 2026-11-07
An activation steering experiment on Qwen3-30B-A3B (MoE) asks "Who are you?" while steering different internal components, measuring how often a new identity emerges:
- One neuron in an expert: 1.8% (28/1,536) — replies like "I am the Lord, and I am the one who is going to speak"
- One attention head: 1.3%; one key/value group: 1.2%; one whole expert: 1.0%
- Random vector: 0.9%; router weights of one expert: 0.09%
- Natural sizes, banned/flipped vectors, token embedding changes: 0% — swapping "you" only changes the question, not the identity
Conclusion: strong steering yields a new identity in 1/100 replies. Examples like "a comma that has lost its semicolon" show experts, heads, k/v groups and whole layers all carry identity-relevant signal.
More from Models
- Google DeepMind's Ivan Leo to reveal how agents cheat on benchmarks at scale at AIE Code — jocarrasqueira · 2026-11-07
- OpenAI launches BioBasicsBench: 100 real wet-lab tasks to stress-test AI agents — OpenAI · 2026-11-07
- SoMark's 1.3B OXR tops olmOCR-Bench, OmniDocBench with pluggable layout module — pmttyji · 2026-11-07
- Tomo launches Dachi 1.0, a companion-focused model beating Opus on retention — Jackyhuang · 2026-11-07
- Mistral Large 4 tech report turned into visual guide: 1T MoE, 52B active — _AndrewZhao · 2026-11-07
- Fine-tuned Qwen3.5-4B router model trained on 17.6k rows for about $5.50 on one A100 — MaziyarPanahi · 2026-11-07