Put a critic in your AI council and you assume the critic knows when it's right. A new paper out of UC Irvine says that assumption breaks exactly where you lean on it hardest — at the agent whose entire job is to catch errors.
The paper is "The Confident Liar: Diagnosing Multi-Agent Debate with Log-Probabilities and LLM-as-Judge" (arXiv:2606.10296), posted June 9 by Ali Keramati, Justin Cheok, Jacob Horne, and Mark Warschauer. It asks a question most multi-agent setups never bother to: when an agent sounds sure, does that mean its reasoning is any good?
The thing nobody checks
Multi-agent debate is usually graded on one number — did the final answer come out correct. That hides the part debate is supposed to improve: the reasoning in the middle. Two agents can reach the right answer through one sound argument and one piece of confident nonsense, and a final-answer scoreboard calls that a win.
So the authors instrumented the middle. They built a two-agent setup — a Constructor that produces an answer and the reasoning behind it, and an Auditor that independently hunts the Constructor's reasoning for errors. Then they measured three things at once: each agent's token-level log-probability (its internal "confidence" as it writes), an LLM-as-judge score rating that reasoning on instruction-following, justification quality, and evidence grounding, plus a critical-failure flag, and finally task correctness. The question is whether the first thing predicts the second. Does sounding sure track being sound?
What they found
In the rubric-scoring domain — essays from the ASAP dataset, where the headline experiments run — confidence and reasoning quality are correlated, but weakly, and the correlation is lopsided in a way that should bother anyone running a council.
For the Constructor, mean correlation between confidence and judged reasoning quality came in at ρ̄ = 0.335. For the Auditor, it was roughly half that. The agent doing the building has confidence that loosely tracks whether its work holds up. The agent doing the checking has confidence that barely does.
The failure-detection number is sharper. When the authors used confidence to flag critical reasoning failures, it worked reasonably for the Constructor — AUROC 0.804. For the Auditor it collapsed to 0.634, a hair above a coin flip dressed up as a measurement. The role whose certainty you'd most want to trust — the critic saying "I checked, this is wrong" — is the role whose certainty carries the least information.
They also found a consistent four-phase confidence trajectory across a generation: agents open confident, wobble through the middle, steady out, then turn volatile right at the end — the moment they're committing to a conclusion. Confidence, in other words, isn't even stable within a single answer.
Why the critic is the blind spot
Here's the council reading. The Constructor and Auditor split is the same split Shingikai runs as Red Team vs. Blue Team — one side builds the case, the other tries to break it. The whole value of that structure rests on the breaker being a reliable instrument. This paper says the breaker is the noisier of the two.
That makes a kind of uncomfortable sense. The Constructor commits to a position, so its confidence has something concrete to be right or wrong about. The Auditor's job is reactive and adversarial — it's incentivized to find fault whether or not fault is there, which decouples how sure it sounds from whether it actually caught something. A critic that's rewarded for objecting will object with conviction either way.
A model's confidence tells you how sure it is. It does not tell you whether it's right. And the agent built to doubt turns out to be the one least able to doubt itself.
What this changes for council design
The wrong lesson is "critics don't work." Red-team structure still surfaces errors a single pass misses. The right lesson is narrower and more useful: a critic's confidence is not a trustworthy stopping signal. If your architecture lets a sure-sounding "looks fine" from the auditor end the deliberation, you've wired your most poorly calibrated signal directly to the off switch.
This is why the synthesis layer can't be a confidence tally. A Chairman that weighs members by how certain they sound will systematically over-trust the critic's verdicts — the exact verdicts this paper shows are closest to noise. The job of synthesis is to read the reasoning, not the volume. Show the argument the auditor made, not the confidence it made it with, and let the chair judge the argument on its merits.
It also argues for surfacing the deliberation rather than the verdict. If confidence and correctness come apart — and they do, twice as badly for the critic — then a council that returns one polished answer with a confidence score is selling you the least reliable part of what happened. The part worth seeing is the disagreement and the reasoning underneath it.
A quick caution on scope: the strong numbers here are from the rubric-scoring domain, and the authors frame the math and factual-QA work as the broader investigation this paper motivates. Treat 0.804 and 0.634 as a sharp finding in one setting, not a universal law. The direction, though — confidence is a poor proxy for soundness, and worst at the critic — is exactly the failure councils are built to absorb.
The reason to run more than one model was never that any of them knows when it's wrong. It's that they're wrong about different things, and a structure can catch what no single confidence score will admit.
If you want to watch a council reason through a hard question instead of handing you one confident answer, try it free — no signup. shingik.ai