Existing emotion-aware studies mainly treat emotion as a surface-level style factor or a perception target, the paper states. The work introduces E-STEER for direct representation-level intervention by embedding emotion as a structured variable in hidden states (https://arxiv.org/abs/2604.00005).

The framework examines emotion's impact on objective reasoning, subjective generation, safety, and multi-step agent behaviors. Results indicate non-monotonic emotion-behavior relations (https://arxiv.org/abs/2604.00005).

Specific emotions enhance LLM capability, improve safety, and systematically shape multi-step agent behaviors, consistent with established psychological theories, according to the primary source (https://arxiv.org/abs/2604.00005).