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. The authors argue that contradictions in prior research reflect a missing distinction: not all multi-agent communication is equal. Different model families produce 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—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, while independent proposal generation avoids this collapse. Full-solution interaction mainly anchors agents to the first solution they see rather than encouraging different approaches; critique only helps when violated rules are easily discovered and fixed by LLMs. The findings suggest multi-agent performance depends more on what information is exchanged than on the number of agents, and interaction only works 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 (original)

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. We argue this contradiction partly reflects a missing distinction, because 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. We call this the interaction tax. We test 11 verifier-scored optimization tasks under matched budgets and find that full-solution interaction is a weak default. Independent proposal generation avoids this collapse. Full-solution interaction mainly keeps agents near the first solution they see rather than trying different approaches, and critique only helps when violated rules are easy for LLMs to find and fix. The results indicate that multi-agent performance depends more on what information is exchanged than on the number of agents, and interaction only works when agents share the right information at the right time.

Key Findings

Not all communication is equal: The contradiction in prior multi-agent research reflects a missing distinction between communication types.
Interaction tax: When agents read each other's full outputs, proposals converge within one round, destroying the diversity that motivates multi-model setups.
Independent generation wins: Under matched budgets across 11 verifier-scored optimization tasks, independent proposal generation avoids the convergence collapse.
Anchoring effect: Full-solution interaction mainly anchors agents to the first solution they encounter rather than encouraging exploration.
Limited critique value: Critique loops only help when violated rules are easily discoverable and fixable by LLMs.
Implication: Multi-agent performance depends more on *what* information is exchanged than on the *number* of agents.