At a Glance
Exchanging full candidate solutions among multiple AI agents often makes them converge in one round and erases the useful diversity that helped find better solutions; keep proposals independent or share low-bandwidth signals instead.
ON THIS PAGE
What They Found
A diverse set of models covers more solution space and avoids complete failure on tasks where same-model teams can score zero. Letting agents read each other’s full outputs causes rapid convergence: after one exchange agents’ solutions become substantially more similar and performance can fall. Workflows that keep proposals independent or only share targeted critique (when the error is obvious) preserve diversity and often perform better. Mutual Verification Pattern
Data Highlights
1Diversity advantage: having diverse proposers gave a coefficient of +0.188 (95% CI [0.073, 0.299], p < 0.001) in controlled tests.
2Collapse in one round: mean pairwise solution distance dropped from 0.315 before interaction to 0.229 after a single full-output exchange.
3Debate interaction tax: when proposers were diverse, the Debate workflow’s multi-agent interaction gain fell by −0.078 (with same > diverse probability ≥ 88%).
What This Means
Engineers building multi-agent AI systems should care because sharing entire solutions can undo the main benefit of using different models—coverage. Technical leaders and product owners evaluating multi-agent trust or agent-to-agent evaluation should prefer designs that preserve independent exploration or use narrow, repair-focused feedback channels. For design considerations, see Emergence-Aware Monitoring Pattern
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: (a) Independent proposals
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Results come from 11 verifier-scored optimization tasks (not open-ended writing or dialogue), so effects may differ in other domains. The factorial diversity check used a subset of tasks and modest seed counts, and the coverage benefit is task-dependent (removing a single task greatly reduces the measured advantage). The study shows full-solution exchange is risky but does not identify the single best low-bandwidth signal to substitute for full sharing. The use of limited signals aligns with coordination approaches like Blackboard Pattern
Methodology & More
Teams made of different models often find complementary solutions, and a diverse team in these experiments never scored zero on any test task while same-model teams did. Experiments ran multiple multi-agent workflows on 11 optimization problems where a deterministic verifier assigns scores. Workflows that let agents read each other’s full candidate solutions (for example, debate or sequential revision) caused solutions to converge immediately: representation distances dropped sharply after one exchange, and synthesis often just copied a single proposer’s output rather than recombining components. That rapid convergence—what the authors call an "interaction tax"—erased the practical benefit of initial model diversity. Tree of Thoughts Pattern Not all communication is bad. Keeping proposers independent and only synthesizing later preserved the diversity advantage, and critique helps when it points to a clear, fixable mistake (for example, Knapsack solutions where capacity violations are easy to spot). Dynamic Task Routing Pattern Critique was less helpful or harmful on problems where the error is hard to localize (for example, a combinatorial constraint in the 3-term arithmetic-progression task). Practical takeaway: avoid sharing full candidate solutions by default; instead use independent proposal generation, low-bandwidth signals (scores, short method descriptions, identified failure causes), or evolutionary selection-style mechanisms to preserve coverage while coordinating agents.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Includes a recognizable researcher (Chenhao Tan) with established reputation; despite missing affiliations in metadata and arXiv venue, author reputation warrants higher credibility.