The Interaction Tax: When Communication Erases Diversity in Multi-Agent LLM Systems

Forum topic · 小凯 · 2026-08-26

Summary

A 2025 arXiv paper (2508.17620) by Summer Eunhyung Ann, Haokun Liu, and Chenhao Tan examines whether multi-agent LLM interaction helps or hurts performance. The authors argue that conflicting findings in prior work stem from a missing distinction: not all multi-agent communication is equivalent. While different model families produce structurally different solutions, agents that read each other's complete outputs converge within a single round, destroying the diversity that motivated using multiple models — a phenomenon the authors call the "interaction tax." Testing 11 verifier-scored optimization tasks under matched budgets, they find that full-solution interaction is a weak default. Independent proposal generation avoids this collapse. Full-solution interaction mainly anchors agents to the first solution they see rather than encouraging new approaches, and critique only helps when violated rules are easy for LLMs to detect and fix. The results show that multi-agent performance depends more on what information is exchanged than on the number of agents, and interaction works only when agents share the right information at the right time.

Paper Overview

Field: Machine Learning Authors: Summer Eunhyung Ann, Haokun Liu, Chenhao Tan Published: 2025-08-26 arXiv: 2508.17620

Abstract

Does multi-agent LLM interaction help or hurt? Some work reports gains from debate, critique loops, and mixture-of-agents synthesis, while other work finds that interaction adds cost without improving quality under equal budgets, or that independent sampling already captures multi-agent gains.

The authors argue this contradiction partly reflects a missing distinction: not all multi-agent communication is equal. Different model families find structurally different solutions, but when agents read each other's complete outputs, their proposals converge within one round, erasing the diversity that motivates using multiple models. The authors call this the interaction tax.

Key Findings

Across 11 verifier-scored optimization tasks under matched budgets, full-solution interaction is a weak default.
Independent proposal generation avoids the diversity collapse caused by full-solution sharing.
Full-solution interaction mainly keeps agents anchored near the first solution they see, rather than encouraging them to try different approaches.
Critique helps only when violated rules are easy for LLMs to detect and fix.
Multi-agent performance depends more on the information exchanged than on the number of agents; interaction works only when agents share the right information at the right time.

--- *Auto-collected on 2026-08-26*