Language-model agents' behavior depends on state encoding, study finds

A new research paper explores how language-model agents interact with their environment through state encodings. The study found that the way an environment's state is encoded can significantly influence the collective behavior of these agents, even when the physical system remains unchanged. Specifically, using low-order circular moments as state encodings led to synchronization in GPT agents, while histogram encodings did not. This effect was observed to reverse direction in Claude agents, indicating that state encodings are not neutral interfaces but rather form part of a model-dependent interaction law. AI

IMPACT Demonstrates that the internal representations of AI models can fundamentally alter their emergent behaviors, impacting how we design and deploy multi-agent systems.

RANK_REASON Research paper published on arXiv detailing experimental findings. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Claude
• Gemini
• generative pre-trained transformer

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →