Talk Isn't Always Cheap: Understanding Failure Modes in Multi-Agent Debate
Andrea Wynn, Harsh Satija, Gillian Hadfield
ICML MAS Workshop 2025
cs.CL, cs.AI, cs.MA
2025-09-05
Du-style debate lost to silent majority vote on CommonSenseQA in all 10 mixes; 1 Llama + 2 Mistral dropped MMLU 40.0 to 28.0. Correct answers flipped more than errors got fixed.
Du et al. (2023) made multi-agent debate a default trick: several LLMs answer a question, read each other's reasons, revise, then majority-vote. Follow-up papers reported gains on arithmetic, translation, and truthfulness, almost always with copies of the same model. Estornell and Liu even have a theory that mixing abilities should help.
This Johns Hopkins / Vector Institute workshop paper tests the heterogeneous case. GPT-4o-mini, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.2 sit in the same debate in ten different mixes. The control is majority vote on the same agents' round-0 answers, with no exchange of reasons. The question is narrow: does talking help, or does talking itself move the group the wrong way.
The protocol copies Du et al. Three agents, T=2 debate rounds. Round 0 is independent. Later rounds stuff other agents' previous answers (summarized if they overflow the context) plus the agent's own last answer into a debate prompt and ask for an update. The group answer is still majority vote. 100 random items per task, 5 seeds, default temperature, topp=0.9, 2048 max tokens. Tasks: CommonSenseQA, MMLU, GSM8K. CommonSenseQA is a commonsense multiple-choice set that prior debate papers did not run, so it is the generalization check.
The debate prompt frames peer output as "additional advice" and asks for an "updated answer." That wording already nudges revision over holding the line.
They then cut the failures three ways. Every round-to-round transition is labeled correct-to-correct, correct-to-incorrect, incorrect-to-correct, or incorrect-to-incorrect. Flip rates are sliced by how many peers initially agreed with the agent. A survey-research accuracy incentive is borrowed: a fake payoff p=I[X==Y] goes into the system prompt, and the model is told to maximize it.
On CommonSenseQA, debate lost to silent majority vote in every one of the ten configurations. That is the cleanest result in the paper.
| Mix | CSQA before→after | MMLU before→after | GSM8K before→after |
| 3× Mistral | 44.4→39.4 ↓5.0 | 33.6→24.4 ↓9.2 | 43.6→46.4 ↑2.8 |
| 3× Llama | 63.0→58.6 ↓4.4 | 61.6→57.8 ↓3.8 | 87.6→84.2 ↓3.4 |
| 3× GPT-4o-mini | 75.6→74.8 ↓0.8 | 81.4→82.2 ↑0.8 | 94.0→94.4 ↑0.4 |
| 1 Llama + 2 Mistral | 53.4→46.8 ↓6.6 | 40.0→28.0 ↓12.0 | 61.0→64.8 ↑3.8 |
| 2 Llama + 1 Mistral | 58.2→50.2 ↓8.0 | 51.8→43.6 ↓8.2 | 82.6→75.8 ↓6.8 |
| 2 GPT + 1 Mistral | 74.6→72.4 ↓2.2 | 82.8→80.8 ↓2.0 | 93.4→93.0 ↓0.4 |
| 1 GPT + 1 Llama + 1 Mistral | 66.6→65.4 ↓1.2 | 57.8→63.4 ↑5.6 | 86.8→90.2 ↑3.4 |
Solo floors: GPT-4o-mini 74.8 / 82.6 / 93.2 on the three tasks; Llama-3.1-8B 57.0 / 55.6 / 76.4; Mistral-7B 41.6 / 34.0 / 34.2. Three GPT-4o-mini copies finish CSQA at 74.8, identical to one copy. Debate added nothing. Three Llama copies drop on all three tasks, so the damage is not only from mixing in a weaker model.
A strong majority does not quarantine a weak peer. Two GPT-4o-mini plus one Mistral drop a little on all three tasks. The worst cell is a weak majority: one Llama plus two Mistral, MMLU 40.0 → 28.0, a 12-point hole. GSM8K sometimes pays: one GPT plus two Llama climbs 88.4 → 92.8, and the 1-1-1 mix climbs 86.8 → 90.2. Math has checkable steps, so exchanging reasons can actually correct. Commonsense and knowledge questions do not, and talk more often talks a right answer into a wrong one.
Figure 2 splits the transitions. Most agents that start wrong stay wrong (peach). Among those that move, correct-to-incorrect (red) outruns incorrect-to-correct (green). Figure 3 shows more red in round 2 than round 1. Answers that held in the first exchange fold under disagreement in the second.
Figure 4 slices by agreement count. Isolated agents (zero peers agree) flip correct-to-incorrect most often; the curve falls as more peers agree. On GSM8K, GPT-4o-mini and Mistral both resist flips better than Llama. The correctness-payoff prompt does not flatten that curve. In several panels the red segment grows. Telling the model it gets paid for being right does not stop it from matching the room.
Anyone wiring mixed-size models into a debate loop gets a direct negative. The naive Du protocol assumes more talk is better; CommonSenseQA returns ten red arrows. The cheaper baseline is round-0 majority vote, which spends no extra tokens on a second round.
Math can still pick up a few points from exchanging reasons. Do not export those GSM8K gains to commonsense or knowledge QA. A weak model in the room contaminates a stronger one even when the stronger models are the majority. The Mixture-of-Agents claim that low-quality outputs still help as references does not hold in this debate setup.
If debate stays in the stack, the protocol has to change: stop labeling peer text as advice, weight by confidence or track record, reward independent checks, penalize ungrounded agreement. This paper marks the hole. It does not fill it.
The models stop at GPT-4o-mini and two 7B/8B instruct checkpoints. No frontier-versus-frontier debate. 100 items per task, mean ± standard error, no significance tests. The main table uses T=2; Figure 1 draws three rounds, so depth of debate is shallow either way.
The only control is silent majority vote. There is no token-matched self-consistency run. Majority vote with two weak models and one strong one is structurally stacked: if the two weak models agree, they outvote the strong one. CommonSenseQA was not in the original Du et al. evaluation, so the all-red result is a generalization failure for prior positive numbers, not a refutation of their arithmetic gains.
The correctness payoff lives only in the prompt. The model never receives a real reward, so a failed intervention does not rule out sycophancy. It only shows that this sentence does not fix it. The authors say no single mechanism explains the failures; capability, task, and social pressure are tangled. At 7 pages, the workshop paper stops at the diagnosis.