Loop-Back Authority Degrades Report Quality in LLM Agent Teams
ORIGINAL / Loop-Back Authority in LLM Agent Teams: A Paired Experiment on Flat and Hierarchical Coordination
The novelty lies in showing that in an open-ended business-intelligence reporting task, hierarchical teams where a manager can send back worker output for revision produce lower-quality reports; they score lower on Utility and Writing Clarity and use more hedging words than flat teams. This provides counter-evidence to the default manager-revision authority in multi-agent LLM frameworks.
01 ABSTRACT
A paired experiment with 43 laptop products and 86 runs compares hierarchical and flat LLM agent teams writing open-ended business-intelligence reports. Fact: flat teams score higher on Utility and Writing Clarity, while hierarchical reports have the same length but use 53% more hedging words, and each revision is linked to a 0.14-point drop in Writing Clarity. Author view: authority improves quality when the manager can verify work, but has a negative effect when it can only provide feedback.
02 KEY FINDINGS
- Flat teams outperform hierarchical teams on Utility (d = 0.42, p = 0.009) and Writing Clarity (d = 0.34, p = 0.030).
- The hierarchical team's first draft is indistinguishable from the flat team's report; the quality gap can be traced to revision.
- Hierarchical reports use 53% more hedging words, and each revision is associated with a 0.14-point drop in Writing Clarity on a 1 to 5 scale.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.CL · ARXIV Q-FIN