Representational Similarity in LLM Interactions: A Double-Edged Sword for Cooperation and Novelty
This research explores the impact of representational similarity on multi-agent interactions, specifically between large language models (LLMs). Findings indicate that while similar representation spaces foster higher cooperation, they simultaneously inhibit novelty and creativity. This effect holds true across various game types and persists even when controlling for factors like performance disparity and model size. Early layer similarity appears particularly crucial for both cooperation and novelty, suggesting a strong link to shared lexical and semantic grounding. These insights are vital for designing effective multi-agent systems.