Reproduce “How Emotion Shapes the Behavior of LLMs and Agents: A Mechanistic Study” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Moran Sun, Tianlin Li, Yuwei Zheng, Zhenhong Zhou, Aishan Liu, Xianglong Liu, Yang Liu
Researchers introduce E-STEER, a framework leveraging Valence-Arousal-Dominance (VAD) models and Sparse Autoencoders to directly steer continuous emotional states in Large Language Models' hidden representations. This intervention systemically impacts objective reasoning, subjective generation, safety, and multi-step agent behaviors, revealing non-monotonic relationships and improving task success rates by up to 14.5% and safety by 68.3% compared to neutral states.
Key insights
- Emotion signals can be directly embedded and manipulated in LLM hidden states, establishing an interpretable, mechanistic link between psychological emotion theories and AI behavior.
- LLMs and agents exhibit non-monotonic emotion-behavior relationships (e.g., inverted U-shaped curves) across VAD dimensions, consistent with human psychological findings, indicating optimal performance at moderate emotional states.
- Modulating internal emotional states can significantly enhance LLM capabilities, improve task success rates, reduce safety risks, and systematically shape multi-step agent planning and decision-making processes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.