Emotion2Skill: Using LLM Internal Emotion Signals for Adaptive Skill Selection

A new framework, Emotion2Skill, proposes to leverage the internal emotional representations of large language models (LLMs) to improve skill selection and evolution in skill-based agents. The framework, detailed in arXiv paper 2608.09248, extracts a 27-dimensional emotion state from the model's residual stream at each decision step and injects a confidence-gated summary into the routing prompt. This approach contrasts with traditional methods that rely solely on text-level signals such as task descriptions and verbal reflections. The authors argue that while LLMs maintain linear emotion representations that causally influence behavior, these have only been used for post-hoc analysis or direct output steering, not for agent-level decision-making. Emotion2Skill aims to fill this gap by incorporating these internal signals into both skill selection and skill evolution processes. The paper was announced as a new submission on arXiv.