Research map / Large language models / Language, safety and interpretability
Do LLMs “Feel”? Emotion Circuits Discovery and Control
Chenxi Wang, Yixuan Zhang, Ruiji Yu, Yufei Zheng, Lang Gao, Zirui Song, Zixiang Xu, Gus Xia, Huishuai Zhang, Dongyan Zhao, Xiuying Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| modulate the assembled global emotion circuit instead of prompting or a single steering vector | LLaMA-3.2-3B-Instruct · held-out test set | emotion expression accuracy (%): 91.22 -> 99.65 (steering vs circuit) | Improved |
| use circuit modulation instead of steering for surprise | LLaMA-3.2-3B-Instruct · held-out test set, surprise | emotion expression accuracy (%): 67.71 -> 100 | Improved |
| ablate top-ranked emotion neurons (k=2, 4) per layer (ablation) | LLaMA-3.2-3B-Instruct · emotion score on test set | emotion score change | Hurt |
| ablate randomly chosen neurons or heads instead of top-ranked ones (ablation) | LLaMA-3.2-3B-Instruct · emotion score on test set | emotion score change | No clear effect |
| enhance top-k neurons with emotion difference vectors (ablation) | LLaMA-3.2-3B-Instruct · emotion score without emotion instruction | emotion score change | Improved |
| enhance a few top attention heads (ablation) | LLaMA-3.2-3B-Instruct · emotion score without emotion instruction | emotion score change | Improved |
| ablate top-2 neurons on Qwen2.5-7B (ablation) | Qwen2.5-7B-Instruct · anger emotion score | delta s: -100.30 (top-2) vs -0.27 (random) | Hurt |
| ablate random attention heads on Qwen2.5-7B (ablation) | Qwen2.5-7B-Instruct · emotion score | delta s: -322.77 (random k=8) vs -544.44 (anger k=8) | Hurt |
| apply emotion steering vectors to Qwen2.5-7B negative emotions | Qwen2.5-7B-Instruct · held-out test set | emotion expression accuracy (%): anger 0.60, fear 0.00 vs happiness 93.30 | Mixed |
| circuit-based steering at scale 2.0 on Qwen2.5-7B | Qwen2.5-7B-Instruct · held-out test set | emotion expression accuracy (%): disgust 100.00, anger 18.33, average 66.18 | Mixed |
Closest papers by meaning
- Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow (ICML 2026 · similarity 0.88)
- MME-Emotion: A Holistic Evaluation Benchmark for Emotional Intelligence in Multimodal Large Language Models (ICLR 2026 · similarity 0.88)
- Unveiling the Cognitive Compass: Theory-of-Mind–Guided Multimodal Emotion Reasoning (ICLR 2026 · similarity 0.87)
- Apathetic or Empathetic? Evaluating LLMs' Emotional Alignments with Humans (NeurIPS 2024 · similarity 0.87)
- Nano-EmoX: Unifying Multimodal Emotional Intelligence from Perception to Empathy (CVPR 2026 · similarity 0.87)
- The Good, The Bad, and Why: Unveiling Emotions in Generative AI (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
