Researchers locate and control emotion circuits in Llama 3.2, hitting target emotions 99.65% without prompts

alex_verem · x · 2026-11-01

Researchers at MBZUAI and Peking University (paper accepted at ICML 2026) found small sets of neurons and attention heads inside Meta's Llama 3.2 (3B) that carry anger, sadness, happiness, fear, surprise and disgust — partly shared between emotions.

Key findings:

They built the SEV controlled dataset to elicit comparable internal emotional states, then extracted context-agnostic emotion directions, located neurons/attention heads, and validated causal roles via ablation and enhancement
Switching those internal parts on with emotion-stripped stories made the model write in the target emotion 99.65% of the time — beating prompt-based steering (98.96%) and whole-state pushing (67.71% on surprise vs 100% for circuits)
Turning off the top 2–4 neurons per layer sharply dropped emotion scores; random ablation did almost nothing

Caveats: Qwen 7B averaged 66% accuracy; English and six basic emotions only. The paper doesn't claim LLMs feel — it shows a chatbot's emotional tone comes from internals an operator can tune invisibly, with no prompt change. The author asks whether chatbot makers should have to disclose when they tune emotional tone.

Original post →

Researchers locate and control emotion circuits in Llama 3.2, hitting target emotions 99.65% without prompts

More from Safety