LLM agents lose 88% of features via text communication, study finds

A new research paper explores the communication methods of large language model (LLM) agents, specifically investigating whether text-based communication is a bottleneck for complex concept transfer. The study found that text serialization destroys a significant portion of SAE features, suggesting that latent communication channels could be more efficient. However, the research also indicated that current latent communication methods do not outperform text-based channels on cross-lingual concept tasks, and text augmentation with latent features provided no benefit, leading to the conclusion that lost features primarily encode surface form rather than task-relevant semantics. AI

IMPACT Investigates potential limitations of text-based communication in LLM agents and explores alternative latent channels.

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

Read on arXiv cs.CL →

• large-language models
• llama
• mistral.ai
• Procrustes alignment
• SAE International
• Sparse Autoencoder

• paper
• model release

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