New research by Zayx Shawn, published on June 12, 2026, examines how message formats impact multi-hop agent relays, revealing that these effects are tier-dependent. The study investigates whether structured messages influence accuracy when large language model (LLM) agents pass information to one another.
Understanding Message Format in Multi-Hop Relays
The study introduces a controlled relay testbed where twelve programmatically generated atomic facts are encoded hop-by-hop in five formats: free natural language (NL), precision-instructed NL, JSON, triples, and key-value. The findings indicate that under faithful-relay conditions, a strong relay operates nearly losslessly, contradicting the commonly known 'telephone-game' phenomenon.
As the research shows, the addition of cognitive load per hop does not significantly affect format-level fidelity, maintaining a range within +/-1.8 points. However, the generation cost increases by 24-53%, suggesting a trade-off between fidelity and cost when using structured formats.
Insights from Tier-Dependent Performance
In a weaker relay scenario, specifically with a 1.5 billion parameter model, the study found that the recall across different formats could vary dramatically, with a spread growing by a factor of 8.7 from 2.3 to 20.5 points. This variation is attributed to two competing mechanisms: a toll incurred by rigid formats and drift resistance linked to the fixed-key JSON schema.
