Can a team of people and AI models get smarter just by averaging their guesses, with no one specializing?

This explores whether a team can beat its members without a division of labor, where everyone attempts the same task and their answers are simply pooled. The corpus says yes, with one condition: the members have to be wrong in different ways. Choudhary and colleagues show that combining human and AI predictions on identical decisions beats either one alone whenever their errors are diverse enough, even when neither is very accurate by itself Can humans and AI improve decisions by predicting together?. The useful ingredient isn't who does what. It's how differently each member errs.

The same logic shows up inside a single model, which is a less obvious place to find it. A generative model trained on data from many imperfect experts can end up outperforming all of them. In effect it absorbs their consensus, and low-temperature sampling pushes its outputs toward that majority view, so idiosyncratic mistakes cancel out Can models trained on many imperfect experts outperform everyone?. So an ensemble doesn't need separate members at runtime. The averaging can be built in during training. Some reinforcement learning work turns the idea into a training goal: SPIRAL rewards a batch of reasoning traces for being useful as a set rather than for each trace being right, and finds that several different mediocre traces beat several near-identical strong ones, because whatever combines them needs real disagreement to arbitrate Can diverse mediocre traces outperform redundant expert traces?.

That also points to how this goes wrong. Diversity can be trained away. RL post-training tends to latch onto one dominant output format from pretraining within the first epoch and suppress the alternatives Does RL training collapse format diversity in pretrained models?. If your ensemble members share a base model and similar post-training, their errors may be more correlated than you'd expect, and most of the ensemble's benefit disappears. Pruning helps from the other side: DyLAN scores each agent's contribution during inference and switches off the ones adding nothing Can multi-agent teams automatically remove their weakest members?. Members that only repeat the others count as uninformative, much like the redundant traces in SPIRAL.

There's a floor, though. Diversity without competence doesn't help and can actively hurt. In multi-agent ideation, cognitively diverse teams lacking real domain expertise did worse than a single competent agent, because the back-and-forth generated noise instead of insight Does cognitive diversity alone improve multi-agent ideation quality?. And for tasks that truly need different skills, parallel work, and independent checking, pooling look-alike generalists hits a ceiling that only role specialization gets past Do single agents always hit organizational limits?.

The takeaway: for prediction and judgment tasks, you can often get accuracy gains without specialization, as long as you deliberately keep the errors uncorrelated. Choose members with different blind spots, and resist training or tuning steps that make them all think alike. Specialization becomes necessary when the work itself has to be split up, not just double-checked.