Study finds induced anger reduces LLM penalty sensitivity in sequential decisions

A study investigates whether induced emotions influence the sequential decision-making of Large Language Models (LLMs) using the Iowa Gambling Task and an imagination-based emotion induction procedure.

• The research validates that LLMs can sense distinguishable emotions and learn from sequential interactions at a human-like pace.
• Unlike humans, induced emotions do not significantly bias LLM decision dynamics on average.
• Inducing anger specifically makes LLM agents less sensitive to penalties for bad decisions.
• In early game stages, anger lowers exploration, locking decisions into a few choices.

These findings reveal subtle effects of emotion on LLM behavior compared to humans and provide a tool for future research on affective modulation.