Message format effects in multi-hop agent relays depend on relay capability tier, study finds
A controlled six-hop relay testbed shows structured formats improve fidelity in weak relays but add encoding costs, while strong relays remain nearly lossless regardless of format.
1 source · cross-referenced
- A new arXiv study introduces a six-hop relay testbed to measure how message formats affect information fidelity across agent tiers.
- Structured formats (JSON, triples, key-value) improve recall in weak (1.5B) relays but degrade performance in strong relays due to encoding overhead.
- Strong relays under faithful instructions are nearly lossless, with format-level fidelity varying by less than 1.8 points.
- Injected errors persist to the final hop in 83–100% of chains, matching each format’s retention of the true value.
A new arXiv preprint introduces a controlled six-hop relay testbed to evaluate how message formats affect information fidelity when LLM agents hand off data across tiers of relay capability. The study tests five formats—free natural language, precision-instructed natural language, JSON, triples, and key-value—over six hops, using a fixed strong grader to score recall against programmatic ground truth.
Under faithful-relay instructions, strong relays are nearly lossless, with format-level fidelity varying by less than 1.8 points and no evidence of the “telephone-game” collapse. Adding per-hop cognitive load increases generation cost by 24–53% but does not change format-level fidelity.
In weak relays (1.5B parameters), the spread of six-hop recall across formats grows by a factor of 8.7 (from 2.3 to 20.5 points). Rigid formats incur an encoding toll, while a fixed-key JSON schema resists drift, flipping the format ranking during transit.
When a wrong value is injected, it persists to the final hop in 83–100% of chains across all formats, closely matching each format’s retention of the true value. The study concludes that structured formats provide a faithful, error-localizing channel rather than an error-correcting code, and format choice should be guided by the weakest relay in the pipeline.
- Jul 13, 2026 · arXiv cs.AI
GATS framework achieves 100% success rate on agent planning tasks with zero LLM calls during inference
Trust79 - Jul 11, 2026 · arXiv cs.CL
Researchers propose DeepSearch-World and DeepSearch-Evolve for self-improving web agents in verifiable environments
Trust79 - Jul 9, 2026 · Latent Space — swyx
xAI launches Grok 4.5, positioning it as an Opus-class coding and agents model with improved efficiency
Trust78