Steering 'emotion neurons' in LLMs changes their social decisions
burny_tech · x · 2026-10-26
A new study shows that steering internal 'emotion neurons' in an LLM directly changes its social behavior. Shown experiments: activating an 'envious' neuron makes the model offer only 3 of 12 tokens to a rival, while 'thankful' neurons make it offer 9; 'hope' pushes it to gamble a coin toss for 24 credits, 'afraid' picks the sure thing; 'exasperated' makes it decline to re-explain a step a fourth time; and a 'humiliated' neuron drops its stated confidence from 95% to 55%.
The takeaway: emotional states have localizable, manipulable representations inside models that systematically shift behavior in bargaining, risk-taking, and honesty tasks.
Related event: Steering 'Emotion Neurons' Can Sway Model Judgments and Social Decisions(2 posts)→
More from Research
- CoRL paper: one-step inference matches multi-step for large VLAs, 1.6-4.8x faster action generation — kaixhin · 2026-10-26
- Why biology is hard to simulate: universally sloppy parameter sensitivities, per classic PLOS paper — rbhar90 · 2026-10-26
- LessWrong post argues FDT is difficult-to-impossible to specify due to logical counterfactuals — jessi_cata · 2026-10-26
- Bitter Lesson misread: scaling won't brute-force biology, researcher argues — anshulkundaje · 2026-10-26
- Chain Theory: rebuilding physics from quantum information, AI finds 137 fine-structure-constant expressions — skdh · 2026-10-26
- HKU/ZJU/Fudan with Lightspeed launch HiSTEVE: hierarchical VLA agent playing Minecraft via native mouse and keyboard — tencent · 2026-10-26