Study: structured messages help only weak AI agents in multi-hop relays

A new paper on arXiv titled "Faithful, Not Corrective: Message-Format Effects in Multi-Hop Agent Relays Are Tier-Dependent" by Zayx Shawn examines how different message formats impact information fidelity when LLM agents hand off data across multiple hops. The study used a controlled relay testbed: 12 programmatically generated atomic facts were re-encoded hop-by-hop in five formats (free natural language, precision-instructed NL, JSON, triples, and key-value) over six hops, scored against ground truth by a fixed strong grader. Two relay-capability tiers (strong vs. weak), a cognitive-load condition, and a paired-fork error injection were tested.

The key findings are tier-dependent. Under faithful-relay instructions, a strong relay is nearly lossless—the documented "telephone-game" collapse does not occur. Adding per-hop cognitive load leaves format-level fidelity unchanged (within ±1.8 points) but raises generation cost by 24-53%. Under a weak 1.5B relay, the across-format spread of six-hop recall grows by a factor of 8.7 (from 2.3 to 20.5 points), driven by two opposing mechanisms: an encoding toll paid by rigid formats and drift resistance specific to fixed-key JSON schema that flips the format ranking during transit. In error injection experiments, an injected wrong value persists to the final hop in 83-100% of chains across all formats, closely matching each format's retention of the true value with no detectable collateral damage. This means structured formats provide a faithful, error-localizing channel—not an error-correcting code—and format choice should follow the weakest relay in the pipeline.