Copies or Sources? Measuring How LLM Aggregators Count Restated Evidence in Multi-Agent Systems

2026-10-05

AI Summary

This study addresses the issue in LLM-based multi-agent systems where aggregators frequently misinterpret repeated statements as independent evidence, leading to decision biases. To quantify the model's over-weighting of redundant information, this work introduces the concept of "replica weight," which transforms probabilities into equivalent units of independent observation. Furthermore, by integrating multi-source evaluations—including a Bayesian oracle, web documents, and agent logs—a communication protocol is designed to eliminate duplication interference. The results demonstrate that optimizing assertion and citation rules significantly reduces the early commitment error rate from 11.2% to 1.1%. These findings provide an effective approach for enhancing the robustness of multi-agent collaboration.

Abstract

Multi-agent systems built on large language models (LLMs) restate observations as a matter of course: relays forward them, shared boards repeat them and discussion rounds echo them. An aggregator that pools such messages should count sources, not statements. We convert a reported probability into units of independent readings, which assigns every restatement a copy weight, 0 for an aggregator that counts sources and 1 for one that counts every statement, and yields the implied decision under any cost structure. Three testbeds hold the evidence fixed and vary how it is restated: message logs with an exact Bayesian oracle, web documents with appended copies, and logs written by LLM agent teams under four communication protocols. Across four models from three providers, a forwarded copy counts for 0.06 to 0.42 of a new reading, mostly because some replies count every statement. On 5% to 40% of logs that state one reading three times, the reported belief implies an early commitment that the oracle never makes. The models that count copies least and most on controlled logs do so on web copies and agent-written logs as well. A one-paragraph declaration of what a copy contributes brings the copy weight on controlled logs to 0.08 or less. A rule that has agents refer to readings instead of restating them cuts belief-implied early commitment from 11.2% to 1.1% and preserves genuine corroboration.

Problem

• multi-agent systems
• large language models
• evidence aggregation
• restated evidence
• copy weight

Innovation

• Multi-agent systems
• Copy weight
• Evidence aggregation
• Large language models
• Bayesian oracle