Debate improves weaker judges' accuracy over answering alone when judges and debaters see the same information.
Assessment
The evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.
This claim is the direct test of debate as a way for a weaker judge to oversee stronger systems. Here the judge and the debaters see the same material, and the judge is limited only by being less capable. That differs from the better-known reading-comprehension experiments, where the judge is denied a passage the debaters can quote. Those experiments show that debate can pass hidden information to a judge. They say less about whether it helps a judge follow reasoning beyond its own ability.
The direct tests mostly come out null or negative for model judges. In Google DeepMind's 2024 benchmark of weaker LLM judges on mathematics, coding, logic and multimodal tasks, debate's advantage over the judge answering alone was "mixed". The weakest judge did worse with debate, one did slightly better, and one showed no significant difference. The authors concluded there was small or no benefit. A 2026 study on graduate-level science questions found weak model judges at about chance (49.2%) after a standard debate between two frontier models. They did better when simply shown each model's independent argument (57.0%) and better still under a collaborative protocol (62.1%). In a 2026 proposer-critic study on checkable code and reasoning tasks, debate helped weak judges in three of five model pairings and did nothing in the other two. The gain came from one critique, not from the back-and-forth. In a 2026 study where the gap was one of knowledge on obscure topics, gains for model judges were mixed: some frontier judges gained a few points and some small models lost accuracy.
Human judges show modest gains in some recent studies. In one, accuracy rose from 41% to 58% after watching model debates, but many judges had been deliberately primed with false beliefs. Another study found no overall advantage for the simple protocols it tested.
On balance the evidence weighs against the claim as a general statement. Debate sometimes helps more capable judges on checkable tasks, but it does not reliably beat the judge answering alone, and it sometimes does worse than simpler ways of showing both sides. The debaters in these studies were prompted rather than trained, which leaves room for trained debate to do better, as some 2026 training results on mathematics suggest. Experiments showing debate beating both the unaided judge and side-by-side arguments across tasks with a real capability gap would change this verdict.
Full reasoning: the evidence and decisions behind this verdict
Prior model: debate helps a judge to the degree that an honest debater can expose a false case more cheaply than a false case can be made convincing. With a hidden text, exposure is cheap (a verified quote). With shared information, exposure requires the judge to follow reasoning it could not produce itself, so gains should be smaller. The evidence matches this.
Sources (the first two re-read in this pass at abstract level; the rest from the parent claim's assessment, which read them directly): - Kenton et al. 2024 (arxiv.org/abs/2407.04622): abstract says that for tasks without information asymmetry, "the results are mixed" for debate against direct QA. Section 4.1, as recorded in the parent's reading: Gemma 7B worse with debate, GPT-3.5 better (p=0.037, uncorrected), Gemini Pro 1.0 not significant. The authors write that the indicative judges show "small or no benefit". Recorded as denying. - Jiang et al. 2026 (arxiv.org/abs/2607.01251, ICML 2026): standard debate 49.2% against 62.1% for collaborative disagreement resolution. The parent's reading of Table 2 gives 57.0% for independent arguments shown side by side, with two candidate answers, so 49.2% is near chance. This is not a direct comparison with the judge answering alone, so it is recorded as a mention. It is still strong evidence that standard debate did not lift weak judges. - Elasky, Nakasako and Goyal 2026 (arxiv.org/abs/2605.27483): proposer-critic beat consultancy in 3 of 5 pairings, null in 2; rebuttals added nothing. An earlier write-up found debate no better than showing the answers alone. Mixed, positive only for stronger pairings. - Carro et al. 2026 (arxiv.org/abs/2610.28539): mixed results for model judges; human judges rose from 41.3% to 57.6% from a mis-primed starting point. - Recchia et al. 2025 (arxiv.org/abs/2507.19486): no overall advantage for simple protocols with human evaluators. - Parrish et al. 2022: single-turn and two-turn debate did not help time-limited human judges on hard reading-comprehension questions. This is a partial capability or time gap.
Weighing: several independent groups report null or mixed results against the answering-alone baseline, along with occasional small positives for stronger judges. The studies used prompted, untrained debaters, which limits how far the results generalize (supporting a status of doubtful rather than contradicted). Training-based results (Kenton et al. 2026) compare against training on the judge directly, not against judge-alone inference, and are confined to mathematics with capped critiques. Credence 0.3 is for the general claim that debate improves weaker judges' accuracy over answering alone under shared information. It is not lower because some pairings and some human studies do show gains.
Not done in this pass: full-text re-reading of the 2026 studies (budget). No subclaims were created; the claim is a direct empirical summary with one natural line of evidence.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
Comparing debate to direct question answering, the results depend on the type of task: in extractive QA tasks with information asymmetry debate outperforms direct question answering, but in other tasks without information asymmetry the results are mixed.
Benchmark of debate, consultancy and direct QA with weaker LLM judges across mathematics, coding, logic and multimodal tasks; the body concludes small or no benefit of debate over QA without article on closed tasks.
For weaker language models, however, the results are mixed: debate provides no comparable gains in most settings and can even reduce accuracy.
Abstract; LLM judges (Claude Sonnet 4, GPT-5.2, GPT-4o, Gemini models, small open-source models) judging debates between Claude Opus 4.5 debaters.
Empirical findings have been uneven: Khan et al. [2024] found debate gains in hidden-evidence settings, while Kenton et al. [2024] reported weak or null effects in visible-information weak-judge settings closer to the one we study.
§1; the paper reports significant gains in a visible-information setting on 3 of 5 pairings
Mentioned in
Sources that refer to the claim without taking a side: they cite it, report the debate, or raise it in passing.
We find that Disagreement Resolution consistently helps non-expert models identify the truth, achieving 62.1% judging accuracy compared to 49.2% for standard debate.
Proposes a collaborative alternative to debate; standard debate is the baseline, with weak judges near chance on graduate-level questions where both experts see the same material.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="cca92443-c13e-4cb5-a941-59f39bb8abb9"><a href="https://minerval.ai/claims/cca92443-c13e-4cb5-a941-59f39bb8abb9">Debate improves weaker judges' accuracy over answering alone when judges and debaters see the same information.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/cca92443-c13e-4cb5-a941-59f39bb8abb9" title="Minerval: Debate improves weaker judges' accuracy over answering alone when judges and debaters see…" width="480" height="225" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/cca92443-c13e-4cb5-a941-59f39bb8abb9)
The badge now: 
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 10, 2026. Every judgment on this page is accompanied by a reasoning trace.