Beyond Blockage: Distinguishing Information Compression from Distortion in LLM Multi-Agent Organizations

Read full text Summarise with VIRAT
Authors: Koh Shinoda
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This study investigates the role of information flow in LLM-based multi-agent systems, focusing on the distinction between harmful information blockage and useful information compression. Prior work identified Compression Ratio (CR) < 0.3 as an indicator of information blockage. This study tests whether CR can be used as a real-time trigger for dynamic structural adaptation in multi-agent organizations. Through 384 experimental runs across 14 tasks, we find that CR-based adaptation significantly degrades output quality compared to the best fixed structure. The root cause is the misclassification of useful compression (especially in final synthesis steps) as harmful blockage. We introduce the compression–blockage distinction, formalized as a two-axis model using Compression Ratio (CR) and Semantic Similarity Decay (SSD). We also identify inter-phase blockage as a new mechanism affecting multi-phase task execution. The study provides design guidelines for multi-agent systems, including the use of flat hub structures, unified prompts, and role-aware adaptation strategies.

Fields of Study & Subjects

Multi-Agent Systems and Negotiation Topic Modeling Natural Language Processing Techniques

Publication details

Publication date
1 January 2026
Language
EN
Document type
Preprint