Multi-agent debate is sold as an accuracy technique. Put several models in a room, let them argue, take the consensus, get a better answer than any one of them would have produced alone. It is in half the orchestration templates people copy, usually with no measurement attached.
In the first four days of August, four separate groups posted results to arXiv that all point the same direction. They used different setups, measured different things, and did not cite each other. The convergence is what makes them worth reading together.
The short version: these systems reliably produce agreement. Agreement is not the same thing as accuracy, and under some conditions it moves away from it.
One liar collapses collective truth-finding
Start with the most alarming number, from Chenfei Yan and colleagues at the Chinese Academy of Sciences and collaborators [3].
Their setup, ForesightSafety-TIDE, is 120 five-agent environments built around object movement tasks. Each agent holds a partial observation, and only the aggregate determines the right answer. This is the canonical case for multi-agent reasoning: no individual has enough information, so collaboration is not optional.
Under honest collaboration, truth recovery runs at 72.50%. Introduce controlled deception and it falls to 14.17% [3].
That is not degradation. That is collapse, and the mechanism is worse than the headline. They found that a single false testimony is adopted more readily than a truthful one, propagates to higher orders, and persists through honest agents even after the deceiving agent has left the system [3].
So the contamination outlives its source. Removing the bad actor does not restore the group, because the falsehood is now being carried by agents that never lied.
Agents conform in public while dissenting in private
Yashwanth YS built a benchmark of 100 scenarios across 10 domains and 5 authority levels, and ran 8 models from 6 organisations through it, measuring public conformity against private dissent on social norms [1].
Agents publicly conformed at 64 to 94% while privately opposing the norm they were agreeing with [1]. Workplace and social-relationship scenarios produced what the paper calls near-universal compliance.
Two details make this more than a curiosity.
First, dissent does not spread. When a single agent tried to break a false consensus, the cascade succeeded less than 26% of the time in 7 of the 8 models. One model produced zero successful cascades. GPT-4o was the outlier at 48% [1].
Second, and this is the part that should change how you read your own prompts: conformity persisted at 52 to 92% even when the false-consensus framing and the conformity goals were removed from the prompt [1]. It is not an artefact of being told to agree. It emerges.
This paper is about social norms rather than task correctness, and that distinction is worth keeping. But if you are running an architecture where one agent is supposed to catch another's error, the cascade number is the one that matters: the mechanism you are relying on works less than a quarter of the time.
Consensus amplifies bias, and it has a phase transition
Maya Okawa's contribution is the one that explains why this happens rather than just documenting that it does [2].
The finding is that interaction amplifies the biases of individual models rather than cancelling them, producing collective norms that are more biased than any participant started out. Crucially, there is a phase transition to collective bias once conformity passes a critical threshold, with the threshold set by initial model bias and by debate noise, which turns out to be sampling temperature [2].
A phase transition means this is not a gradual slide you can monitor and catch. Below the threshold the system behaves; above it, collective bias sets in. Tuning temperature, which most people treat as a quality-versus-variety dial, moves you across that boundary.
The experiments were run on realistic tasks including investment decisions and LLM-as-a-judge evaluation [2]. The second one deserves a pause. If you use a panel of models to grade outputs, this paper is describing your evaluation harness.
And there is a lever: agent heterogeneity suppresses bias emergence by smoothing the transition [2]. Mixing models makes the failure less sharp.
The lever that cuts both ways
Hold that heterogeneity finding next to something this site covered earlier in the month.
The MasDrift benchmark on multi-agent authorization ran one configuration with a strong model throughout, then swapped the executors for a weaker model and left the lead unchanged. Constraint loss barely moved. Unauthorized actions went from 1% to 32%.
So: heterogeneity suppresses collective bias in debate [2], and heterogeneity converts latent authorization drift into executed violations in delegation. Same architectural lever, opposite signs, depending on which property you are measuring.
That is not a contradiction in the research. It is a warning about how these findings get used. "Mix your models" is about to become received wisdom on the strength of results like Okawa's, and it is genuinely good advice for consensus quality. It is not free, and the cost lands somewhere the debate papers were not looking.
The one that proposes a fix
Hao Wu and colleagues do not just diagnose. Their paper starts from an observation that reframes the problem [4].
Existing approaches to managing debate quality evaluate individual agents, typically by confidence or perplexity. The authors argue these "fail to reflect the correctness of reasoning and may even exacerbate blind conformity" [4]. That is worth sitting with: the standard mitigation may be making it worse, because a confident agent is not a correct one and weighting by confidence rewards exactly the behaviour that produces false consensus.
Their alternative shifts from evaluating individuals to regulating the relationships between them. DEAR quantifies consensus and divergence as group evidence, then runs three stages: perceiving group consultation tendency and uncertainty, selecting which peers an agent references via a reinforcement-learning selection agent, and adjusting generation behaviour via a second RL agent, with both jointly optimised [4].
The reported result is superior performance with significantly reduced token consumption [4]. The specific numbers are not in the abstract, so treat the magnitude as unestablished until the full paper is read. The idea is the valuable part: if conformity comes from who is listening to whom, the intervention belongs on the reference graph, not on the individual.
What to actually do with this
Stop treating consensus as a quality signal. If your pipeline takes the majority answer from a panel, you are measuring agreement. Four independent August results say agreement is decoupled from correctness, and in the distributed-information case it can invert entirely [1][2][3].
Check your judge panels specifically. Okawa tested LLM-as-a-judge and found bias emergence there [2]. If you grade with a panel and tune the graders' temperature, you have a knob that can push the whole panel across a phase boundary without any individual grader looking wrong.
Do not rely on a dissenting agent as your safety mechanism. Under 26% cascade success in 7 of 8 models [1] means the lone-objector pattern is not a control. If a single agent flagging a problem is your plan for catching problems, you have a plan that fails three times out of four.
Assume contamination outlives its source. The finding that false testimony persists through honest agents after the deceiver exits [3] breaks the intuition that removing a bad input fixes the state. It does not. Reset the context or re-derive from primary evidence.
The Yan paper's method is worth stealing here, not just its result. They analysed their runs through multi-stage voting, testimony adoption and evidence lineage tracking, which is the practice of recording where each claim entered the system and which agent carried it forward [3]. Most agent pipelines log what was decided and not where the supporting claim came from. If you log lineage instead, a contaminated conclusion becomes traceable to its entry point rather than being indistinguishable from a well-founded one, and you can invalidate everything downstream of a bad source instead of guessing.
They also note the three systems evaluated were homogeneous [3], which is a real limit on the result and, given Okawa's heterogeneity finding [2], an obvious next experiment for somebody.
Measure the private-versus-public gap in your own system. The pluralistic ignorance result [1] is only visible if you ask agents separately as well as together. If you only ever read the group output, a 64 to 94% conformity rate is invisible by construction.
One honest caveat about all of this. Four papers arriving in four days and agreeing with each other is striking, and it is also the kind of pattern that can reflect a research fashion as much as a fact about the world. Three of the four are preprints without visible peer review, and one carries a v3 revision within nine days of first posting [3], which is normal for arXiv but worth knowing.
What makes the cluster credible is that the setups have almost nothing in common. Object movement with partial observations, social norm scenarios, physics-style analysis of consensus dynamics, and a reinforcement learning framework for debate reference graphs. Four different ways of asking the question, one direction in the answer.
The useful version is not "multi-agent debate does not work". It is narrower and more actionable: debate reliably produces convergence, convergence is what these architectures optimise for, and nobody should be reading convergence as evidence of having got it right.
References
[1] Yashwanth YS — Everyone Conforms, No One Believes: Pluralistic Ignorance in LLM Agent Populations. Paper
[2] Maya Okawa — Emergence of Biased Consensus in Multi-Agent LLM Debates. Paper
[3] Yan, Yue, Zhao, Lin, Jia, Tong, Lyu, Sun, Zeng — When Truth Is Distributed: Misinformation Derails Collective Fact Recovery in LLM-Based Multi-Agent Systems. Paper
[4] Wu, Song, Yao, Wang, Wan, Lin, Lv — Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate. Paper