The standard way to catch an AI hallucination is to hand the answer to a second AI and ask it to check the work against the evidence. It fails for a reason nobody building these systems wants to say out loud: the checker reads the answer first.
Once it has read the answer, it isn't auditing anymore. It's looking for reasons the answer is right. The model that was supposed to be the safeguard has quietly joined the side it was meant to interrogate.
A new paper from Alibaba's Qwen team — MARCH, arXiv:2603.24579, accepted to ACL 2026 — gives this failure a name and a fix. The name is confirmation bias: an LLM-as-a-judge verifier "inadvertently reproduces the errors of the original generation." The fix is stranger and more interesting than a bigger model. They blindfolded the checker.
The judge who read the verdict first
Retrieval-augmented generation was supposed to make hallucination rare. Give the model the documents, ask it to answer from them, and it should stay grounded. It doesn't. It still invents figures, misquotes numbers, and asserts things the retrieved evidence never said.
So the field reached for the obvious patch: a second LLM pass. Take the answer, take the evidence, ask a judge model whether the first holds up against the second. This is now the default reliability layer in production RAG, and it is the layer MARCH is attacking.
The problem is that the judge sees the same thing the writer saw, plus the writer's conclusion. Anchoring does the rest. The answer frames the question. The judge starts from "is this right?" instead of "what does the evidence actually say?" — and those are not the same question. A judge who reads the verdict before the evidence isn't checking the work. It's co-signing it.
Take the answer away from the auditor
MARCH splits the job across three roles that share one trained policy. A Solver writes the initial RAG answer. A Proposer breaks that answer into small, checkable claims — atomic question-answer pairs, the paper calls it a "Response Atomizer." Then a Checker answers each of those questions from the retrieved documents alone, with the Solver's original answer deliberately withheld.
That withholding is the whole idea. The paper's term for it is "information asymmetry," and the Checker's nickname is the "Blinded Auditor." It cannot agree with the Solver out of habit, because it never sees what the Solver said. It can only go back to the documents and answer the question fresh. If its independent answer matches the claim, the claim survives. If it doesn't, the claim is a hallucination — caught not by a smarter critic, but by one that was denied the chance to anchor.
This is the move that matters, and it generalizes well past this one paper. A second opinion is only worth something if the second opinion couldn't see the first.
Why "close enough" gets a zero
MARCH trains this loop with multi-agent reinforcement learning, so the three roles co-evolve instead of being bolted together after the fact. The reward is where the design gets its teeth. They call it a Zero-Tolerance Reward, and it is exactly as unforgiving as it sounds: a whole answer scores as successful only if every claim in it matches the Checker's independent verdict. One hallucinated claim sends the entire response to a penalty. There is no partial credit for being mostly grounded.
That sounds harsh until you remember what a single fabricated number does to a financial summary or a medical answer. The paper specifically prioritizes numerical and quantitative claims — the misquoted figures and altered values that are the most expensive kind of confident wrong.
The headline result is the part worth sitting with. In the authors' words, "an 8B-parameter LLM equipped with MARCH achieves performance competitive with powerful closed-source models." Not a bigger model. A small one, wired so it cannot rubber-stamp its own output. The benchmarks span factual grounding — RAGTruth, FaithBench, Facts Grounding, ContextualJudgeBench — and multi-hop QA — HotpotQA, 2WikiMultiHopQA, MuSiQue.
The council most people are building is an echo
Here is the uncomfortable part for anyone running a multi-model setup. Most "AI councils" in the wild are arranged exactly the way MARCH says not to. One model drafts. The others see the draft and react to it. A judge reads everything and picks. Every member downstream of the draft has already been anchored by it. That isn't a council. It's one opinion, validated four times by models that all started from the same sentence.
The fix isn't more members. It's controlling who sees what, and when. Red Team vs Blue Team works because the red team is hunting for the failure, not reading the answer charitably. A blinded checker works because it never met the answer it's grading. The structure is doing the work that a bigger model can't — and MARCH is the cleanest demonstration of it we've seen this year, because it isolates the single variable. Same models. Same documents. The only change is whether the auditor gets to peek. That change is the difference between catching the lie and signing it.
It's the principle underneath every strategy in a real council: the value isn't in stacking models, it's in making sure at least one of them is reasoning from the evidence instead of from the answer. When the question is expensive to get wrong, that independence is the entire point — and it's the thing a single model, grading its own homework, can never give you.
Try it free — no signup. shingik.ai