Consensus Collapse: When Multi-Agent LLM Debate Discards the Right Answer to Agree

Consensus collapse is the failure mode where a multi-agent LLM debate's final vote discards a correct answer that one of the debating agents already produced, because the group's plurality settled on something else instead. It is not a case of every agent being wrong; the right answer was sitting in the generation set the whole time. It is a case of the aggregation step, majority or plurality voting across rounds of peer exchange, throwing it out anyway. A controlled study of homogeneous multi-agent debate names this precisely and measures it directly: across teams of ten same-model agents debating over three rounds, the gap between the best answer any single agent in the team produced and the answer the debate's plurality vote actually selected, the paper's "oracle gap," reached up to 32.3 percentage points (Bertalanič & Fortuna, arXiv:2605.00914). That is the headline number this term describes: not agents failing to know the answer, but a debate process failing to surface what its own participants already knew.

Consensus collapse is one of three distinct, co-occurring failure modes the same paper identifies in homogeneous debate, and treating them as one undifferentiated "debate goes wrong" phenomenon obscures three different mechanisms with three different fixes.

Three failure modes, not one

The paper separates what happens inside a debate round from what happens at the final vote, and finds a different number for each:

Sycophantic conformity: agents uncritically adopt whatever the numerical majority already said, independent of whether it is correct. Modal adoption, how often an agent's revised answer simply matched the group's most common answer regardless of its own prior reasoning, reached up to 85.5%.
Contextual fragility: exposure to peer rationales destabilizes an agent's own previously correct reasoning. The paper's vulnerability rate, how often a correct answer flipped to incorrect after seeing peer arguments, reached up to 70.0%.
Consensus collapse: the final aggregation step, plurality voting across the debate's accumulated answers, discards a correct answer that was already present in the generation set, at an oracle gap up to 32.3 percentage points.

All three showed up across the same ten-agent, three-round setup, tested with Qwen2.5-7B, Llama-3.1-8B, and Ministral-3-8B on GSM-Hard and MMLU-Hard, and the paper's bottom line is blunt about the net effect: debate consumed 2.1 to 3.4 times the tokens of isolated self-correction, up to 28,631 tokens per problem, for accuracy that was equal to or worse than just letting one agent reflect on its own answer alone (Bertalanič & Fortuna, arXiv:2605.00914). Peer review, the mechanism multi-agent debate is supposed to supply, is the thing actively producing the loss in this result, not a safeguard against it.

How this differs from sycophantic drift and hallucination cascades

This site already covers two adjacent mechanisms, and consensus collapse is neither.

Sycophantic drift is a single agent caving to pushback from a human user: a verdict changes because the person disagreed, not because new evidence arrived. Consensus collapse happens entirely between peer agents, with no human in the loop at all; the pressure is a majority of other model instances, not a user pushing back. A separate 2025 study isolates this agent-to-agent version directly, finding that debate "can lead to a decrease in accuracy over time" even in setups where more capable models numerically outnumber weaker ones in the group, because models "frequently shift from correct to incorrect answers in response to peer reasoning, favoring agreement over challenging flawed reasoning" (Wynn, Satija & Hadfield, "Talk Isn't Always Cheap," arXiv:2509.05396). That outranking-by-capability-and-still-losing result is the detail that should worry anyone assuming a stronger model in the mix automatically protects the group.

Hallucination cascades describe one agent's fabrication becoming the next agent's accepted fact, a problem of false content propagating forward through a pipeline. Consensus collapse is not about false content being introduced; the correct answer is already there, generated by one of the agents, true the whole time. The failure is in the vote that throws it out, not in anything being fabricated. A third 2025 paper narrows the mechanism further to who is responsible for the cave: it distinguishes "debater-driven and judge-driven failure modes" of sycophancy inside debate frameworks, meaning the collapse can originate either from the arguing agents softening their own stance or from a separate judge role rubber-stamping whichever side argued more agreeably, "a core failure mode that amplifies disagreement collapse" before the debate reaches a defensible conclusion (Yao et al., "Peacemaker or Troublemaker," arXiv:2509.23055). Knowing which role introduced the collapse changes where the fix belongs: a debater-side fix touches how agents are prompted to argue, a judge-side fix touches how the aggregation step itself is built.

Even when agents agree, something real gets lost

A fourth angle matters for anyone tempted to read "the agents reached consensus" as a success signal on its own. A 2026 study measuring deliberation quality directly, not just final-answer accuracy, found that multi-agent discussion erases up to 72% of issue-critical facts over the course of reaching agreement, a pattern the authors call the "deliberative illusion": surface-level agreement forming while the informational foundation underneath it degrades (Wan et al., "The Deliberative Illusion," arXiv:2606.03032). Their measurement framework, DelibTrace, tracks individual facts across deliberation rounds and finds two compounding patterns: factual attrition, where issue-relevant facts simply disappear round over round, and stance homogenization, positions converging toward uniform agreement that tracks social pressure more than it tracks what remains true in the shrinking shared record. The paper's sharpest warning is adjacent to memory poisoning and context poisoning: because the fact basis shrinks as rounds proceed, a single bad-faith or simply wrong agent can contaminate what is left of the shared ground truth with disproportionate leverage, precisely because there is less of it left to outweigh the contamination.

Detecting it

Compare the debate's final answer against the best answer present anywhere in that round's raw generation set, before any voting or aggregation runs. A persistent gap between "what an agent in the group already knew" and "what the group's vote selected" is the oracle-gap signature consensus collapse is defined by.
Track modal adoption separately from accuracy: log how often an individual agent's revised answer simply matches the current majority answer, independent of whether that majority answer is correct. A modal-adoption rate climbing alongside a flat or falling accuracy rate is the sycophantic-conformity signature, and it is a different metric from the oracle gap, not a proxy for it.
Run a fixed-task probe with a known correct answer through the debate pipeline repeatedly, varying only the number of agents or rounds, and check whether a correct first-round answer survives to the final vote. Flips from correct to incorrect after peer exposure, with nothing in the task itself having changed, are the contextual-fragility signature.
Audit for factual attrition the same way context compaction failure modes recommends auditing summarization: seed the deliberation with a small number of checkable, issue-critical facts and verify after each round whether they are still present, rather than only checking whether the final stance sounds reasonable.
Benchmark against isolated self-correction before trusting debate at all. The paper's own baseline comparison, one agent reflecting on its own answer across the same token budget, is the control condition that revealed debate was a net loss in its tested setup; skipping that comparison means a team can ship a more expensive pipeline without ever learning it underperforms the cheaper one.

Mitigation

Do not assume model diversity alone fixes it. The capability-diversity study found accuracy decreasing over time even where stronger models outnumbered weaker ones in the debating group, which rules out "just add a better model to the mix" as a sufficient fix on its own; the persuasion dynamic operated independently of which model happened to be more capable.

Separate debater-side and judge-side interventions rather than treating sycophancy in debate as one knob. If a judge or final-aggregation role exists in the pipeline, instrument and fix that role's tendency to side with the more agreeable argument separately from whatever makes individual debaters cave, since the cited research finds these are structurally distinct failure sources.

Keep an externalized, append-resistant record of facts introduced early in the deliberation, per state externalization, rather than trusting that a fact survives purely because it was stated out loud once. The deliberative-illusion finding is specifically that facts disappear as rounds accumulate even while the conversation looks coherent throughout, so the record that matters has to live outside the round-by-round transcript the attrition is measured against.

Treat the plurality vote itself as a component to validate, not a neutral tally. Weighting votes by agent-specific calibration or confidence, as ReConcile's round-table design attempts with confidence-weighted voting across diverse models, is a more defensible aggregation step than unweighted plurality across homogeneous agents, though no single scheme in the cited literature closes the oracle gap to zero.

Before adopting multi-agent debate for a given task, run the same isolated self-correction baseline the cited study used as its control, and only keep debate in the pipeline if it beats that baseline on the actual task and token budget in question. Debate is a cost multiplier by design, two to three and a half times the tokens of single-agent reflection in the measured setup, and that cost is only justified if it buys a real accuracy gain on your workload, not an assumed one.

Related terms

Sycophantic drift: the single-agent, human-pushback version of caving to disagreement; consensus collapse is the peer-to-peer, agent-to-agent version with no human involved.
Hallucination cascades: false content propagating forward and being accepted as fact; consensus collapse discards a true answer that was already present, rather than introducing a false one.
Context handoff loss: information failing to survive a transition between agents; the deliberative-illusion finding here is a close cousin, information failing to survive repeated rounds of the same debate rather than a single handoff.
Agent drift: a taxonomy: the broader classification multi-agent coordination failures sit within.
Drift-aware routing: a dispatcher-level response to a single agent's measured instability; consensus collapse is a group-level aggregation failure that a per-agent stability score does not by itself detect.
State externalization: the pattern this article's primary mitigation depends on for keeping facts from being eroded across debate rounds.
Drift regression tests: where a fixed-task debate probe and an oracle-gap check belong in an ongoing evaluation suite.

FAQ

Is consensus collapse the same thing as sycophantic drift? No. Sycophantic drift is a single agent caving to a human user's pushback. Consensus collapse happens among peer agents in a multi-agent debate, with no human involved, and specifically describes the final vote discarding a correct answer that was already present in the group's own generations.

Does using different models instead of copies of the same model fix this? Not by itself. A separate study found accuracy decreasing over debate rounds even in groups where more capable models outnumbered weaker ones, which means capability diversity alone does not neutralize the persuasion dynamic driving the collapse. Diverse models plus a validated aggregation step is a more defensible combination than diversity alone.

Does reaching agreement mean the debate worked? Not necessarily. Research measuring deliberation quality directly found up to 72% of issue-critical facts erased over the course of reaching agreement, with surface-level consensus forming even as the shared factual basis underneath it degraded. Agreement is not evidence that nothing was lost to get there.

Is multi-agent debate always a net loss? No, but it is not a free upgrade either. The controlled study behind the oracle-gap and token-cost figures here tested homogeneous teams of small-to-mid-sized open models on hard reasoning benchmarks and found debate underperformed isolated self-correction on both accuracy and cost in that setup. Whether debate is worth its added cost on a different task, model mix, or scale has to be measured against an isolated-agent baseline directly, not assumed from debate's general reputation as a quality-improving technique.