Researchers develop benchmark to test how emotional states influence decision-making in small language model agents

Study combines emotion induction at the representation level with game-theoretic evaluation to measure how emotions affect SLM behavior in decision-making tasks

Emotional states are induced using activation steering derived from real-world, crowd-validated emotion-eliciting texts, providing more controlled and transferable interventions than prompt-based methods

Benchmark uses canonical decision templates covering both cooperative and competitive scenarios with complete and incomplete information, instantiated through strategic games like Diplomacy and StarCraft II

Experiments across multiple model families and architectures demonstrate that emotional perturbations systematically affect agent decision-making behavior