Beyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent Systems

Yang, Chih-Hsuan; Li, Xinyang; Thakur, Rajeev; Yan, Feng

Left: first-pass accuracy is about equal with and without roles (A), while recovery after a failed first answer is highest for the role-free team with feedback (B). Top right: one deliberation cycle, in which only the Reviewer’s approval sends an answer to the evaluator. Bottom right: gemma-4’s role-structured team spends most of its six cycles without submitting; forcing a submission every cycle uses the budget on submissions (C). — Without roles, the first answer is just as good. What roles change is when the team asks to be corrected.Only the Reviewer’s approval sends an answer to the evaluator. Recovery is oracle-assisted; forced submission is one exploratory run on 400 problems.

What do roles change beyond the division of labor?

We hold the agents, topology, call budget and ground-truth evaluator fixed, then delete every role instruction from a Planner–Executor–Reviewer team. Specialization largely disappears, yet first-pass accuracy does not drop. Roles also allocate decision rights: a role-structured Reviewer often withholds approval, so its team submits less and receives less corrective feedback. In exploratory comparisons, the role-free team recovers from failed first answers more often in all five settings, and forcing a submission every cycle closes the gap.

Supplementary material ↗Experiment reproduction guide ↗

Citation

Download BibTeX ↗

Chih-Hsuan Yang, Xinyang Li, Rajeev Thakur, and Feng Yan. Preprint, 2026.

BibTeX

@misc{yang2026beyond,
    title = "Beyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent Systems",
    author = "Yang, Chih-Hsuan  and
      Li, Xinyang  and
      Thakur, Rajeev  and
      Yan, Feng",
    year = "2026",
    note = "Preprint"
}