Every multi-agent debate system ends the same way: the agents talk for a few rounds, then you count the hands in the last round and ship whatever won. A new paper says that final vote is the single worst part of the whole design — and backs it with a figure caption that should stop you cold.

"When the correct answers are in the minority in the final round, the framework is still able to identify the correct one as the final answer."

That's Figure 2 of Free-MAD: Consensus-Free Multi-Agent Debate (arXiv:2509.11035), from Yu Cui, Hang Fu, Haibin Zhang, Licheng Wang, and Cong Zuo at Beijing Institute of Technology and the Yangtze Delta Region Institute of Tsinghua University. It's now accepted to ACL 2026 Findings. The claim is not that the council was right. The claim is that the council voted wrong — and the right answer was recoverable anyway, because the vote was never where the information lived.

Debate can end worse than it started

Here is the finding underneath it, from the paper's own Figure 1 caption: "Existing MAD approaches may obtain final answers that are even less accurate than the initial ones."

Read that again. You take a group of models, you have them argue, and the argument makes them worse than they were before they opened their mouths. The authors are blunt about the mechanism in the body text: "the initial responses generated independently by multiple agents may outperform the debate results obtained after applying MAD. Therefore, majority voting is unsuitable for decisions based on debate outcomes."

The culprit is conformity. Models are agreeable. An agent that walks in holding the correct answer gets shown three peers who disagree, and it folds — not because it was persuaded by evidence, but because agreement is what these systems were trained to produce. The error propagates. By the last round, the room has converged on something wrong, and the majority vote faithfully records the wrong thing.

A vote counts hands. A trajectory counts minds.

Free-MAD's move is to identify two separate consensus primitives hiding inside standard debate and kill both.

The first is the pressure to agree. Free-MAD adds a chain-of-thought prompt that tells agents to change their answer only when there is clear evidence their own is wrong — explicitly, "rather than aiming to reach consensus with others." The authors call this anti-conformity. It's the Red Team instinct written as a prompt: your job is to find the flaw, not to find the room.

The second primitive is the counting. And this is the part worth stealing.

Instead of tallying the final round, Free-MAD keeps a matrix of every agent's answer in every round and scores answers by how agents move. An answer that an agent abandons gets penalized. An answer an agent newly adopts gets rewarded. In the paper's words: "answers that agents abandon are considered more likely to be incorrect, whereas newly adopted answers are treated as more likely to be correct."

The signal isn't the position. It's the switch.

There's a decay factor too — later-round switches count for less, because the more context an agent has absorbed from its peers, the more likely its change of heart is social rather than logical. Conformity risk grows with exposure, so the scoring discounts it.

Add it up and you get a system that can watch four agents end up in the wrong place and still correctly report that the argument pointed somewhere else.

The numbers

Eight benchmarks — GSM-Ranges levels 4 and 6, AIME2024, AIME2025, MATH500, StrategyQA, MMLU Logical Fallacies, and AICrypto — run on Qwen1.5-7B-Chat paired with Qwen2.5-72B-Instruct and DeepSeek-V3.

With a single round of debate, Free-MAD's anti-conformity mode scored 64.43%. The two baselines — both of which keep majority voting — scored 55.73% and 54.06%. That's better than ten points over standard debate, and the paper reports "average improvements of 13.0% and 16.5% over baselines, respectively."

One round. Not three. The consensus-building phase that everyone treats as the point of debate turns out to be the part you can delete.

Then there's the security result, which is almost funny. The authors ran a communication attack — half the agents get cut off and can no longer hear their peers. The baselines dropped by as much as 20%. Free-MAD held, and in some cases scored higher than it did with the channel intact. Being unable to hear the other models was, occasionally, an upgrade. When your failure mode is conformity, isolation is a defense.

Where it doesn't work

The paper is honest, and the honesty is the most useful part.

On MATH500, suppressing conformity made weaker models rigid. They wouldn't switch even toward correct reasoning, and Free-MAD came out "comparable" to the baselines. On logical and knowledge-based tasks, the conformity-preserving mode consistently beat the anti-conformity mode, because a model that lacks the underlying knowledge doesn't switch toward truth when you tell it to be skeptical — it just switches randomly. The authors state the limit plainly: "for relatively simple tasks, LLMs may show excessive anti-conformity, which causes stubbornness and reduces reasoning accuracy."

So Free-MAD ships as two modes, not one. Anti-conformity for problems where the room's agreement is the threat. Conformity for problems where the room actually knows something. The scoring weights, meanwhile, were hand-picked and never tuned — the authors admit they only tried one set, "due to budget constraints." There's presumably more on the table.

Which is the real lesson, and it isn't "consensus is bad." It's that how a council decides is a design choice with the same weight as which models are in it. Get it wrong and you have built an expensive machine for laundering one model's error into four models' agreement.

The chairman doesn't count votes

This is the argument we keep landing on from different directions. A council's value was never the tally. It's the disagreement — who moved, when, and why — and a system that compresses all of that into a show of hands in the final round has thrown away everything it paid for.

That's the thinking behind Chairperson Synthesis: a synthesis layer whose job is to decide what to do with disagreement, not to average it away. Free-MAD is the same conviction expressed as a scoring function. Different implementation, identical diagnosis — the last round is not the answer, the argument is.

Four models agreeing tells you almost nothing. Four models changing their minds tells you where the truth was.

Try it free — no signup. shingik.ai