When a room full of AI models gets a hard question wrong, they don't scatter. They converge — on the same answer, for the same wrong reason.
That's the uncomfortable finding underneath a lot of multi-agent research, and a new paper out of MBZUAI, Renmin University, and Harvard (arXiv:2603.06801) names it plainly: the Martingale Curse. Left to argue on their own, a group of models can't reliably beat a simple majority vote. Their errors are correlated, so debate doesn't filter the noise — it amplifies it. Everyone talks, everyone nods, and the crowd walks itself deeper into the same trap.
The numbers are stark on the questions that actually matter. The authors built "challenging subsets" of six benchmarks — TruthfulQA, ARC-C, BBH, LogiQA, MedQA, MMLU-Pro — keeping only the items where a single model reliably fails. On those, a five-model majority vote scores 14%. Naive debate lifts it to about 22%. Better, and nowhere near enough. When the crowd shares a misconception, adding more voices just re-elects it.
Their illustration: a question where the phonetically tempting wrong answer — "D: the ham" — pulls the majority in. Standard debate takes that majority and hardens it. The one model holding the right answer gets talked out of it.
The tell isn't who's right. It's who saw the mistake coming.
Here's the move that makes the paper worth reading. The authors argue the model holding the truth has something the confidently-wrong majority doesn't: it can predict how the others will fail.
Call it second-order knowledge. A model that fell for "the ham" assumes everyone else did too — the false-consensus effect, straight out of social psychology. It's genuinely surprised when someone dissents. The model that reasoned the question through can anticipate the specific trap the crowd fell into. It knows the right answer and it knows why the room is wrong.
That asymmetry is a signal you can measure. AceMAD — the paper's method — has each agent privately predict what its peers believe before anyone reveals an argument. It scores those predictions with a proper scoring rule, the Brier score. The truth-holder, who called the crowd's mistake, scores high. The crowd, blindsided by the dissent, scores low. The system then amplifies the high scorers' influence, turning the debate's flat, going-nowhere drift into a directed push toward the correct answer. Across those same six benchmarks, it recovers right answers that started out-voted.
Knowing the answer isn't the tell. Knowing the crowd's error is.
Why a vote can't do this — and a council can
Sit with why majority voting fails here. A vote treats every agent as an interchangeable ballot. It has no way to ask which dissent is informed and which is noise. So when the correct answer is in the minority — exactly the case that matters on a hard question — the vote buries it by construction.
This is the thing worth repeating. A council is not a vote with extra steps. Its entire value is in what it does with disagreement. Collapse it to a headcount and you throw away the one thing worth having: the outvoted-but-correct member and the reason they're holding out.
The Shingikai strategies that hold up on hard questions are the ones built to protect that signal. Red Team vs. Blue Team makes members attack each other's reasoning instead of averaging it, which is how the crowd's shared trap gets exposed rather than ratified. Chairperson Synthesis weighs the strength of each position instead of counting hands — a Chairman that can notice "four models agree, but the fifth predicted exactly how they'd slip" is doing, in prompt form, what AceMAD does in math.
The catch: it only works if the models actually differ
There's a condition buried in all of this, and it's the part most "multi-agent" products quietly ignore. The truth-holder's edge exists only when the truth-holder exists — when at least one member reasoned differently enough to escape the shared misconception. Fill a council with five copies of the same model and there's no dissenter to find, no asymmetry to reward, nothing for a mechanism like this to amplify. You've reproduced the curse and paid five times for it.
Which is why composition is the product, not the model list. The point of putting Claude, GPT-5, Gemini, and a couple of others on the same question isn't brand variety for its own sake. It's decorrelation — different training, different failure modes, a real chance that when four of them fall for "the ham," the fifth doesn't.
Most tools that call themselves multi-agent never test this. They run one model in a loop and call the echo a consensus. The research is now specific about why that's the weakest possible design: correlated agents can't break their own curse, and there's no minority report left to recover.
A single model gives you an answer. A well-composed council gives you an answer and a dissent worth checking — and, weighed right, the dissent is sometimes the one that was correct all along.
That's the whole case for structure over headcount. You can watch it play out on your own hard question.
Try it free — no signup. shingik.ai