Researchers find full-answer sharing collapses AI agent diversity

One round of full-answer sharing erases the solution diversity that separates different AI model families, per a study testing 11 optimization tasks under matched compute.

Different model families start out finding structurally different solutions to the same problem. That diversity is the entire reason to run more than one model instead of one, and it's what teams pay extra compute for when they build debate or mixture-of-agents pipelines rather than ship a single model.

Critique loops helped only when the violated rule was easy for a model to spot and fix. Independent proposal generation, where each model never sees another's full draft, kept the diversity intact through the same benchmark.

That reframes years of contradictory results on debate and mixture-of-agents setups as a question of what agents share, not how many run. Full answers before voting optimizes for consensus; withholding full drafts until judging time keeps the search spread across the solution space.