In multi-agent LLM pipelines, the quality of the judge matters more than the diversity of the team.
This resolves a genuine contradiction in the field (arXiv:2603.20324). Diverse teams — agents with different capabilities, different models, different prompting strategies — should outperform homogeneous ones. That's the intuition behind Mixture-of-Agents: throw multiple perspectives at a problem and synthesize. But empirically, synthesis-based aggregation often favors uniform teams. The diverse perspectives cancel each other out when blended.
Note (2026-11-08): the abstract says that across 210 tasks in 7 categories a diverse team with judge-based selection reaches a win rate of 0.810 against a single-model baseline while a homogeneous team scores 0.512, near chance, and that judge-based selection beats MoA-style synthesis by 0.631, with synthesis preferred over the baseline in none of 42 tasks. The essay attributes the 0.512 to synthesis with the same diverse teams; in the abstract it is the homogeneous team's score. The abstract calls the finding about including a weaker model exploratory and not pre-registered, and its conclusion is that selector quality 'may be' a more impactful lever in single-round generate-then-select pipelines; the essay states the weaker-model effect as a result, explains it as contrast like a control group, and generalizes to evolution, markets and hiring, none of which is in the abstract.
The resolution is the selection bottleneck. Diverse teams generate more varied outputs, including better ones. But they also generate worse ones. Whether diversity helps depends on whether the system can select the better outputs. A judge-based selection mechanism achieves a 0.810 win rate against baseline with diverse teams. Synthesis — which averages rather than selects — scores 0.512 with the same teams. Near chance.
The mathematics yields a crossover threshold. Below it, diversity helps regardless of selection method. Above it, diversity hurts unless the selector is good enough to handle the increased variance. The selector's quality is the binding constraint, not the generators' diversity.
An unexpected corollary: including a weaker model alongside stronger ones improves performance while reducing cost. The weak model doesn't contribute winning answers. It contributes contrast — its outputs help the judge calibrate what “good” looks like, the way a control group helps a scientist recognize the treatment effect.
The structural lesson generalizes beyond LLM pipelines. Any system that generates candidates and then selects from them — evolution, markets, peer review, hiring — faces the same bottleneck. Increasing the diversity of candidates is easy. Building selectors that can handle the resulting variance is hard. And the selector, not the candidate pool, determines the ceiling.