A council of nine models just learned to stop talking the moment five of them agree. That's the whole idea behind EMS — Efficient Majority-then-Stopping — a new paper out of the University of Science and Technology of China. Run nine LLMs on a hard question, and once the leading answer mathematically can't be overturned by the votes still outstanding, the system quits and returns it. No sense paying four more models to rubber-stamp a result already decided.
The numbers are clean. Across five benchmarks — GSM8K, GPQA Diamond, CommonsenseQA, AQuA, and MMLU — EMS matches the accuracy of full majority voting exactly (87.01% average, same as running all nine every time), while cutting the average number of models invoked from 9.00 to 5.84 and token consumption from 7,381 to 4,123 per query. That's 35% fewer agents and 44% fewer tokens for the same answer. On easy math, 90% of GSM8K questions are settled after the first five models even open their mouths.
If you run councils, you feel this problem in your bill. Querying a room of frontier models is expensive, and most of that spend is confirmation — models agreeing with a conclusion the room already reached. EMS is a genuinely smart fix, and the mechanism is more careful than "count to five." It orders the models by task-conditioned reliability: for each new question, it looks up how each model has done on semantically similar past questions and calls the most reliable ones first, so consensus arrives sooner. That's a quiet but real concession — EMS has already abandoned one-model-one-equal-vote. Some voices are weighted more than others before the counting even starts.
So the paper is good. It's also optimizing the exact wrong thing for anyone using a council to make a decision.
Here's the tension. EMS is built to reach agreement as cheaply as possible. A council exists to surface disagreement before it resolves. Those are opposite objectives wearing the same nine models. The dissenting votes EMS prunes as redundant are the votes a council is built to hear.
Watch it show up in the paper's own error table. EMS reports how often early stopping produces a wrong answer that full voting would have caught. On AQuA — easy algebra — that harmful-stop rate is 0.39%. Negligible. On GPQA Diamond — graduate-level science, the hard set — it climbs to 3.54%, and the overall false-positive stop rate jumps from 6% to 19%. The pattern is exactly what you'd expect and exactly what should give you pause: the harder the question, the more often stopping early costs you the right answer. On GPQA, only a third of questions resolve after the first five models; the rest genuinely need the room. Cheap consensus works beautifully on the questions where you didn't need a council in the first place, and gets riskiest on the ones where you did.
That's not a knock on EMS. It's honest about this — the authors frame it as an efficiency method for majority voting, full stop. The knock is on treating majority voting as the thing a council does.
Because majority voting, efficient or not, throws away the one artifact worth keeping. It reduces nine reasoning traces to a tally and reports the tally. EMS makes the tally cheaper to compute. Neither the fast version nor the full version ever asks why the two models that disagreed disagreed. And on a real question — should we take this deal, is this diagnosis right, does this contract have a hole — the why is the entire product. The minority answer isn't noise to be outvoted and dropped. It's often the model that saw the edge case the other eight missed.
Run the two side by side. Ask nine models a factual question with a verifiable answer, and EMS is the right tool: get to the certified majority, stop, save the tokens, move on. Ask nine models whether to fire a founder, and a majority vote — stopped early or run to completion — hands you "six to three, go" and buries the three most useful sentences anyone said. You wanted the trade-off mapped. You got a headcount.
This is why the interesting work in multi-model AI is moving away from the vote as the consensus primitive, not toward optimizing it. A vote counts heads. It can't tell you that the lone dissenter was the only one who read the fine print. Structured deliberation is built for the opposite job: Red Team vs. Blue Team, where models attack each other's reasoning instead of casting parallel ballots; a Chairperson layer that reads the disagreement and decides what to do with it rather than declaring whichever pile of votes is taller. Those strategies deliberately refuse to collapse to a majority, because the collapse is where the signal dies.
EMS asks how cheaply a council can agree. A council's whole value is what surfaces before it does.
None of which means efficiency doesn't matter — it obviously does, and any council that ignores cost won't survive contact with a real budget. It means efficiency and deliberation are answers to different questions. If the question is "what's the answer, and I can check it," make the vote as cheap as EMS makes it. If the question is "what should we do, and getting it wrong is expensive," the last thing you want is the fastest path to a tally. You want every reason the room didn't agree, laid out where you can see it.
That second question is the one Shingikai is built for — councils that surface the disagreement instead of counting past it, across 200+ models. Try it free, no signup: shingik.ai.