In multi-agent debate, what do extra rounds buy beyond a single critique pass?
answer
- gains front-loaded, then flatten
- round two carries most of the lift
- convergence is not correctness
- agents herd toward the confident voice
- rounds are sequential, so latency stacks
basics
~20 sExtra debate rounds let each agent see the others' revised arguments and change its own, which corrects confident first-shot errors. Most of the lift lands in the first one or two exchanges; after that agents converge and mostly restate a consensus that may be wrong.
solid answer
~50 sA single critique pass gives the proposer one outside opinion; debate iterates it. Each agent answers, reads the other answers, and revises, so a claim has to survive repeated challenge instead of one review. The published multi-agent debate work (Du et al., framed as a "Society of Minds") shows the gain concentrated in the early exchanges on reasoning and factuality tasks, then flattening. Extra rounds mostly buy convergence, not correctness: agents drift toward whichever position is stated most confidently, and if they share a base model and prompt they can herd onto the same wrong answer. Practically, I cap rounds at two or three, stop early when answers stop changing, and force diversity by giving debaters different prompts or different base models. Each round multiplies both tokens and wall-clock latency, because rounds are strictly sequential.
go deeper
Be able to say what a debate round is: several agents answer separately, then read each other and revise. Knowing that this costs one full generation per agent per round already puts you ahead of a vague "multiple agents check each other" answer.
Explain the mechanism and the shape of the curve — most of the improvement arrives in the first exchange and flattens after two or three rounds — and name herding as the reason agreement does not prove correctness. Interviewers expect you to state the token and latency multiplier without prompting.
Show you have run this: a round cap, an early-stop rule when answers stop changing, deliberate decorrelation of debaters, and a rule that a deterministic verifier always beats an argument. Be ready to say how you would detect a debate that converged on a confident wrong answer in production.
Own the framing that debate is one way to spend a token budget, competing with a stronger single model, better grounding, or a verifier. Expect to defend applying it only to a selected slice of traffic and to explain what you would measure before making it the default path.
## What multi-agent debate is In multi-agent debate, several model instances are asked the same question and produce answers independently. Their answers are then shown to each other, and every agent is asked to reconsider in light of what the others said. That cycle — answer, read peers, revise — is one round. After a fixed number of rounds, the final answers are collapsed into one output, either by taking the majority position or by handing the transcript to a separate judge agent. The idea entered the field with work on improving factuality and reasoning through multi-agent debate (Du et al.), often described under the "Society of Minds" framing: reasoning emerges from disagreement between independent processes rather than from one process thinking longer. ## A single critique pass versus a debate A single critique pass is asymmetric: one agent proposes, one agent reviews, and the proposer either accepts the review or does not. Only one opinion ever enters the system, and if the critic is wrong the error is now laundered as a correction. Debate is symmetric and iterated. Every participant is both proposer and critic. A wrong claim has to survive being contradicted, then survive the contradiction being contradicted. That is the mechanism: repeated challenge filters claims that only one agent's particular sampling path supported. ## Where the gain actually comes from The first exchange carries most of the value, because that is the moment an agent first encounters evidence it did not generate. An agent that hallucinated an intermediate step in an arithmetic or multi-hop factual chain frequently abandons it once two peers show a different chain. This is why debate helps most on tasks where errors are idiosyncratic — each run fails differently — and helps least on tasks where the model has a systematic bias that every instance shares. By the second and third round, the answers being exchanged are already heavily influenced by the previous round, so the marginal new information per round shrinks. Published curves flatten quickly. Beyond three rounds you are usually paying full price for restatement. ## Convergence is not correctness The most important failure mode is herding. Debate reliably produces agreement; it does not reliably produce truth. Agents tend to defer to the most assertively phrased position, and models that share training data and prompt framing share priors, so a confident wrong answer can pull the whole panel. A transcript that ends in unanimous agreement therefore tells you the process terminated, not that the answer is right. If you report confidence to a downstream consumer, do not derive it from unanimity alone. A related pathology is instant collapse: on round one every agent already agrees, so the remaining rounds are pure cost. Detect this and terminate. ## Keeping disagreement real Several levers keep the panel from being one voice repeated: - Have every agent commit to a private first answer before seeing any peer output, so the first round is genuinely independent. - Vary the debaters: different system prompts, different personas with different priorities, or genuinely different base models. Different models are the strongest source of decorrelated error; temperature alone is the weakest. - Require each revision to name the specific claim it is disputing, rather than allowing a bare concession. - Give the debaters access to a retrieval or execution tool so at least some claims can be settled against evidence instead of rhetoric. ## When debate is the wrong tool If the task has a cheap deterministic verifier — the code compiles or it does not, the SQL runs, the JSON validates, the invariant holds — run the verifier. Verification is orders of magnitude cheaper than argument and is actually correct rather than persuasive. Debate earns its keep on judgement calls and on reasoning chains where no checker exists. Debate is also weak on open-ended generation with no ground truth. There, the panel converges on the most conventional phrasing, which reads like consensus and is really regression to the mean. ## The cost, stated plainly Debate multiplies generations. Three agents over three rounds is nine full generations before you even run a judge, and each round's prompt carries the previous rounds' transcript, so per-call input tokens grow as the debate proceeds. Latency is worse than the token count suggests, because rounds cannot be parallelised — only the agents within a round can. A structure that raises accuracy by a few points while multiplying cost by an order of magnitude has to be justified against simply spending that budget on a stronger single model or better grounding. ## What a good answer sounds like Name the mechanism (independent answers, then mutual revision), say where the gain concentrates and why it flattens, name herding as the reason convergence is not evidence, and finish on the round cap plus early-stop rule and the sequential-latency cost. That is the full shape an interviewer is listening for.
- What signal would you use to stop a debate early?Stop when the round-over-round answers stop changing — no agent revised its position, or the diff between rounds is cosmetic. Some systems ask each agent for a short structured verdict alongside its argument so consensus is machine-checkable rather than judged by a model. Always pair the early-stop rule with a hard round cap, so a pair of agents that keep flip-flopping cannot run forever.
- How do you stop a debate collapsing into agreement on the first round?Make the first answer private: every agent commits before seeing any peer output. Then decorrelate the debaters — different system prompts, different priorities, ideally different base models, since same-model agents share priors and agree too cheaply. Requiring each revision to cite the specific claim it disputes, rather than allowing a bare concession, also keeps the disagreement substantive.
- Would you use debate on a task that has a unit-test suite as ground truth?No. When a deterministic verifier exists, run it — the tests settle the question for a fraction of a single generation, and they are actually right rather than merely persuasive. Debate is for judgement calls and reasoning chains with no checker. A sensible hybrid gives the debaters the verifier as a tool so their arguments are grounded in run results instead of rhetoric.
It is like a review panel that keeps talking until everyone agrees: the first exchange surfaces the real objections, and the later rounds mostly manufacture unanimity — including, sometimes, unanimity about something false.
saying these in an interview costs you the question
- Claims accuracy keeps rising with every additional debate round
- Treats unanimous agreement as evidence the answer is correct
- Uses debate on tasks that already have a cheap deterministic verifier
- Gives every debater the same model and prompt and still expects diversity
- Ignores that rounds are sequential, so latency scales with round count