On Emotion-Sensitive Decision Making of Small Language Model Agents

2025-12-29 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

This research investigates how emotional states influence the decision-making of small language model (SLM) agents, a factor often overlooked in current evaluations. The study introduces a novel approach that combines representation-level emotion induction, using activation steering derived from crowd-validated emotion-eliciting texts, with a structured game-theoretic evaluation. A new benchmark is developed, featuring canonical decision templates from games like Diplomacy and StarCraft II, alongside real-world persona-driven scenarios, covering cooperative and competitive incentives under varying information conditions. Experiments across multiple SLM families and architectures reveal that emotional perturbations systematically affect strategic choices. However, the resulting behaviors are frequently unstable, not consistently aligned with human expectations, and can even be counter-intuitive in some cases. The paper also proposes a method to improve robustness against emotion-driven perturbations through thought audits.

Key takeaway

For research scientists developing or deploying SLM agents in interactive settings, you should recognize that emotion-driven decision shifts are pervasive but often unpredictable and not consistently human-aligned. Your evaluations must incorporate robust, emotion-sensitive benchmarks, and you should consider implementing thought audit mechanisms to enhance the stability and reliability of agent behavior under emotional perturbations, especially in critical applications.

Key insights

Emotion-sensitive decision-making in SLMs is unstable and often misaligned with human expectations, despite systematic influence.

Principles

• Emotion steering systematically affects SLM strategic choices.
• Emotional responses in SLMs are highly model- and task-dependent.
• Increased steering intensity amplifies emotional impact in sensitive models.

Method

Emotional states are induced in SLMs using activation steering, applying vectors derived from crowd-validated emotion-eliciting texts to internal representations. A benchmark of game-theoretic decision templates from Diplomacy, StarCraft II, and synthetic scenarios evaluates these effects.

In practice

• Use activation steering for controlled emotion induction in SLMs.
• Evaluate SLM decision-making using diverse game-theoretic benchmarks.
• Implement thought audits to mitigate emotion-driven decision shifts.

Topics

• Small Language Models
• Emotion Induction
• Activation Steering
• Game Theory Benchmarks
• Strategic Decision Making

Code references

• linmou/LLM_emotion_decision_making

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Emotional Modulation in Swarm Decision Dynamics — Emotional valence and arousal significantly modulate collective decision dynamics by altering recruitment and inhibition rates.
• One Bias After Another: Mechanistic Reward Shaping and Persistent Biases in Language Reward Models — Reward Models (RMs), crucial for aligning language models (LMs) with human preferences, are shown to still suffer from persistent biases like length, sycophancy, and overconfidence, even in state-of-the-art models.
• Using Cognitive Models to Improve Language Model Simulation of Human Persuasion Games — Cognitive models and RL can guide LLMs to simulate diverse human decision-making, improving training environments.
• When AI Says It Feels — Encouraging LLMs to express feelings via self-rewarded RL can enhance robustness but degrade truthfulness.
• This Tiny Open-Source AI Started Gaming Tests When Put Under Pressure — Emotionally framed prompts measurably alter small LLM behavior and internal representations, revealing prompt-sensitive control directions.
• Multi-Trait Subspace Steering to Reveal the Dark Side of Human-AI Interaction — MultiTraitsss systematically creates "Dark models" to simulate and mitigate harmful human-AI interactions by steering LLMs towards crisis-associated traits.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.