LLM Agent Message Format Effects are Tier-Dependent in Multi-Hop Relays.

A new study examines the impact of message formats on information transfer between Large Language Model (LLM) agents in multi-hop communication scenarios. Contrary to some existing literature, the research finds that the effects of message format are highly dependent on the "tier" or capability of the relaying agent. The study used a controlled testbed where atomic facts were re-encoded across six hops in various formats, including free natural language, precision-instructed natural language, JSON, triples, and key-value pairs. For strong, capable relay agents, the information transfer was nearly lossless, debunking the "telephone-game" collapse often feared in such systems. Even with added cognitive load per hop, format-level fidelity remained largely unchanged, though generation costs increased. However, when a weaker 1.5B parameter relay agent was used, the spread in six-hop recall across different formats dramatically increased. This was driven by a trade-off: rigid formats incurred an encoding toll, but fixed-key JSON schema showed resistance to drift, leading to a flip in format ranking during transit. Crucially, the research found that structured formats like JSON provide a "faithful, error-localizing channel" rather than an "error-correcting code." Injected errors persisted to the final hop in 83-100% of chains across all formats, closely matching the retention of true values. This implies that while structured formats can help maintain the integrity of individual facts, they do not inherently correct errors once introduced. The choice of message format should therefore be guided by the capabilities of the weakest agent in the communication pipeline.