Jun 2026
A senate of language models
One model is a confident liar. Five hundred of them, voting, are a deliberative body that can disagree with itself.
A single language model is a confident liar. Ask it a hard question and it returns one fluent answer in one voice with one set of blind spots, and it will defend a hallucination with the same calm prose it uses for the truth. The fix everyone reaches for is a bigger model. The more interesting fix is more models. Not one oracle. A senate.
The idea is older than transformers. In 1785 the Marquis de Condorcet proved a theorem about juries: if each voter is even slightly better than a coin flip, and they vote independently, the majority gets more accurate as you add voters, and the accuracy climbs toward certainty. With 51 percent per juror and 1,000 jurors, the majority is right essentially always. The two load-bearing words are independent and better than chance. Break either one and the theorem turns on you.
Sampling the same model is not a senate
Here is the trap people fall into first. They take one model, set temperature=1.0, sample it 500 times, and take a majority vote. This is "self-consistency," and it genuinely helps on math and reasoning. DeepMind and Google showed it lifting GSM8K accuracy by double digits. But it is not Condorcet. Five hundred samples from gpt-4 are not 500 independent jurors. They are 500 draws from one brain with one prior. When that brain is confidently wrong, every draw is wrong the same way, and the majority ratifies the error with a huge fake margin. You have not built a senate. You have built one senator shouting in an empty room and counting the echoes as votes.
Correlation is the enemy. The whole reason a vote beats an individual is that errors cancel, and errors only cancel when they point in different directions. Two models trained on overlapping web dumps, fine-tuned with similar RLHF recipes, fail on the same trick questions. Their mistakes are correlated, so averaging them buys you almost nothing. The math is brutal here: the variance reduction from N voters scales like the average pairwise correlation, not like N. Add a thousand near-identical models and you get the accuracy of about three good independent ones.
Where the disagreement has to come from
So a real senate needs diversity that is structural, not cosmetic. Different base models from different labs. Different prompts that frame the problem from different angles. Different tools, where one senator can run code and another can search the web and a third only reasons from memory. The point is to manufacture uncorrelated errors on purpose. This is the same instinct that made random forests beat single decision trees: you deliberately weaken and decorrelate each voter so the ensemble gets strong.
Anthropic published a concrete version of this in 2025. Their multi-agent research system put one orchestrator model in charge of spinning up subagents that explore different branches of a question in parallel, then synthesized the results. It beat a single Claude Opus by 90 percent on their internal research eval. The honest line in the same writeup: it burned about 15 times the tokens of a normal chat. A senate deliberates, and deliberation is expensive. You are paying for the redundancy that makes the errors cancel.
Voting is the weakest way to aggregate
Plain majority vote throws away almost everything the senators know. It ignores confidence, it ignores who is usually right about what, and it lets a loud cluster of correlated wrong answers win. The upgrades are where it gets good. Weight each model by its track record, so the senator who is reliable on chemistry counts more on chemistry. Let them debate: several papers show that when models critique each other across two or three rounds, the wrong ones often revise toward the right answer, and the group converges. Or skip voting entirely and have a judge model read every argument and rule, the way a panel of forecasters beats any single forecaster in Philip Tetlock's superforecasting work.
And debate has its own failure mode worth naming: models are sycophantic, so a confidently stated wrong answer in round one can drag the honest ones toward it instead of away. A senate of yes-men is just a louder liar. The diversity has to be real enough to survive contact with a persuasive mistake.
The deeper reason this matters is that intelligence at the frontier is starting to look less like a single huge weight matrix and more like a protocol for combining many fallible reasoners. One model gives you an answer. A senate gives you a distribution, a margin, a dissent, and a reason. When the vote is 499 to 1 you can trust it and move on. When it splits 260 to 240 you have learned the most valuable thing of all, which is that the question is genuinely hard and your one confident oracle was about to lie to you with a straight face.
Our senate cannot lie about its vote
Verse adds one thing to the recipe: the senators are replay-verified. Pin a model by its weight hash, fix the seed, set temperature to zero, and its decision becomes a deterministic program anyone can re-run. So a senate like that cannot lie about what it ruled — you just run it again. And to be exact about what runs today: the council that actually allocates our grants is not a senate of models but a deterministic, karma-weighted quadratic-funding fold — replayable for the same reason, votes signed by the people who verified real work. That karma gate is the decorrelate-on-purpose move aimed at sybils: a hundred fake zero-karma voters add nothing to the result. The model senate is the same primitive aimed at harder questions, and it is not the thing signing grants yet. You still cannot prove a senate has taste. You can prove exactly what it ruled.