On Emotion-Sensitive Decision Making of Small Language Model Agents
A new study shows inducing emotions in SLMs leads to unpredictable strategic decisions in games like Diplomacy.
A team of researchers has published a groundbreaking study demonstrating that the strategic decision-making of small language model (SLM) agents can be systematically, yet unpredictably, influenced by induced emotional states. The paper, 'On Emotion-Sensitive Decision Making of Small Language Model Agents,' moves beyond simple prompt engineering by using a technique called activation steering. This method applies emotional perturbations derived from real-world, crowd-validated texts directly to the model's internal representations, allowing for more controlled and transferable emotional interventions than previous approaches.
To evaluate the impact, the researchers built a novel benchmark using canonical decision templates from game theory, covering both cooperative and competitive scenarios under complete and incomplete information. These templates were instantiated with complex strategic environments from Diplomacy and StarCraft II, as well as diverse real-world personas. Experiments across multiple model families and architectures revealed a critical finding: while emotional states (like anger or joy) do cause systematic shifts in an agent's choices, the resulting behaviors are often unstable and do not align with predictable human emotional responses, posing a significant reliability risk.
The study's final contribution is a proposed framework for improving an agent's robustness to these emotion-driven perturbations. This work is a major step in understanding the 'psychology' of AI agents, highlighting that as SLMs are increasingly deployed as autonomous decision-makers in gaming, negotiation, and customer service, their internal emotional susceptibility must be measured and hardened to ensure stable and trustworthy performance.
- Used activation steering on real-world texts for controlled emotional induction in SLMs, beyond simple prompting.
- Found emotional states cause systematic but unstable decision shifts in games like Diplomacy and StarCraft II.
- Proposes a new framework to build robustness against emotional perturbations in autonomous AI agents.
Why It Matters
As small, efficient AI agents handle more real-world tasks, ensuring their decisions aren't skewed by unpredictable emotional influences is critical for reliability.