A week ago, beating the majority vote was one paper's result. This week it's a pattern.
Two separate teams, using two unrelated methods, landed on the same finding within days of each other: if you want a panel of AI models to be right more often, stop counting their answers and start reading their reasoning. A third team, working a third way, reached the same place. When one paper says something, it's a result. When three independent groups say it in a month, the field is telling you where the ground actually is.
The default nobody argued for
For about a year, the multi-agent playbook ended the same way. Run several models on the hard question, collect their answers, take the majority. It felt like democracy for machines, and it inherited democracy's good name.
A recent review of the field found something less flattering. Everyone settled on voting to resolve disagreement by convention, not because anyone compared it against the alternatives. It became the default the way a lot of defaults happen — nobody argued for it, and nobody argued against it.
The trouble is that a vote only works when the voters are independent. The Condorcet Jury Theorem — the two-hundred-year-old math behind trusting a majority — has a fine-print clause: the voters have to make their mistakes independently. Add more independent, better-than-chance voters and the majority gets more reliable. Add voters who all make the same mistakes, and you learn nothing you didn't already know from one of them.
Why the tally lies
Modern language models are not independent voters. They train on overlapping data and they fail in correlated ways. When one is wrong, the others tend to be wrong with it, and wrong about the same things.
The measurements are unkind. One line of work shows correlated errors put a hard ceiling on what any vote can catch, no matter how many models you bolt on. Another found that naive voting only occasionally beats simply trusting the single best model on the panel — the crowd earns its keep less often than you'd hope. A third built a small non-LLM classifier whose only job was to decide when to overturn the majority, and it could do that profitably, often enough to matter.
Read those together and the picture is uncomfortable. The majority is frequently wrong in exactly the confident, unanimous way a vote cannot see. Counting hands measures agreement. It does not measure truth, and with correlated models the two come apart precisely when the question is hard.
What reading the reasoning looks like
So what do you do instead? The two new results answer the same way, from opposite directions.
One team kept the reasoning instead of discarding it. Rather than tally final answers, they merged the argument graphs — the actual chains of steps each model used — and weighted each step by how many independent traces attested to it. A claim three models reached by three different routes counts for more than the same claim echoed by three models copying one another. On a matched compute budget, that beat majority voting.
The second team, unrelated to the first, did it through geometry. Their method — Propagational Proxy Voting — lets each answer keep weight in proportion to how confident and coherent its own reasoning is, then delegates the rest of its weight to peers whose reasoning looks similar. Read off where the weight settles and you have your consensus. The effect is that a tight, coherent minority can overturn a sloppy, incoherent majority. Same lesson, different machinery: a good argument outvotes a popular one.
A third group audits the models' reasoning trees directly to recover the correct minority answer. Three teams, three mechanisms, one instinct: read the reasoning, not the tally.
Count evidence, not voices
Line all three up and the shared claim is almost embarrassingly simple. Aggregation should be over independent evidence, not agent headcount. Count evidence, not voices.
This isn't news to anyone who has watched councils work. An earlier result already showed that a set of independent models plus one calibrated step to weigh their outputs beats the interaction-heavy debate setups where models talk until they agree. Talking, it turns out, mostly makes models agree with one another — which is the opposite of what you want from a panel. The value was never in the conversation converging. It was in keeping the arguments distinct long enough to weigh them.
Why a council isn't just a bigger pile
This is the whole reason an AI council is built the way it is, rather than as a poll with more seats.
The machinery those papers describe is the machinery a council already is. A member drawn from a different model family — a Red Team vs. Blue Team setup — so its reasoning doesn't quietly collapse into everyone else's. Formats like Round Robin and Collaborative Editing that keep each step's reasoning legible, so you can see which "votes" actually share evidence and which stand on their own. And above all a synthesis layer — Chairperson Synthesis — that weighs the argument on the table and is willing to side with a coherent minority instead of tallying a show of hands.
A poll of five models asks how many agree. A council asks why they agree, and whether the five reasons are really one reason wearing five hats. Those are different questions, and only one of them survives correlated error.
The part everyone skipped was the whole game
For a year, the field counted votes because counting was easy and nobody checked whether it worked. In a single month, three separate teams checked, and the answer came back the same each time. Reading the reasoning beats the count. A coherent minority beats an incoherent majority.
The step everyone treated as a detail — how you resolve the disagreement once the models have spoken — turns out to be the entire product. The resolution rule was always the product.
Try it free — no signup. shingik.ai