Steering just 100 of Qwen3-8B's 440k neurons matches persona vector effect
maxsloef · x · 2026-11-01
Research by wmayner shows persona vector steering is sparse in the neuron basis: adjusting just 100 of Qwen3-8B's 440k neurons has the same effect as a persona vector, while steering top directions of a randomly rotated or PC basis fails — suggesting the neuron basis is privileged. maxsloef adds the neurons even outperform the persona vector's PCs.
More from Research
- Nostalgia: early AI researchers dared to map RL concepts to brain neuromodulators — fleetwood___ · 2026-11-01
- Superintelligence guiding existing cancer therapies could beat new drug approvals — hardimanjames · 2026-11-01
- Will the replication crisis slow AI progress? Burned-in intuitions may stay wrong — moultano · 2026-11-01
- CMU paper shows pretraining, SFT and RL can run as one unified likelihood-based algorithm — omarsar0 · 2026-11-01
- Frontier LLMs still fail basic molecular cloning, researcher says — anshulkundaje · 2026-11-01
- After42: experts rate 45 of AI's 135 proposed research questions genuinely interesting — james_y_zou · 2026-11-01