Most multi-agent debate is an argument about memory. Several models take a question, push back on each other, revise — but every one of them is reasoning from what it absorbed in training. When they're wrong, they tend to be wrong together. The debate gets more confident without getting more correct.

A January arXiv paper changes what the models are arguing about. Tool-MAD (2601.04742), out of Beakcheol Jang's ANDlab at Yonsei University, gives each agent in the debate its own external tool — one queries a search API, another pulls from a RAG module — so the argument runs on evidence the agents go and fetch, not just the weights they shipped with.

That sounds like a plumbing change. It's the difference between a confident answer and a checkable one.

Why fact verification is the hard case

Fact verification is where confident-but-wrong does the most damage. A model asked to verify a medical claim or a contested news fact will produce a fluent verdict either way. Add more models and let them debate, and you get the failure mode the whole field has been naming all year: agents reinforce each other into a shared wrong answer, and the chorus reads as certainty.

The standard fix was to give the debate something external to stand on. MADKE — the framework Tool-MAD builds against — did exactly that: it retrieved evidence once, before the debate started, and then let the agents argue over that fixed packet. Better than arguing from memory alone. But the evidence never updated. When a new claim surfaced halfway through the debate — the kind of turn that's the entire point of deliberation — the agents had nothing fresh to check it against. They were debating live with a frozen reference book.

The reframe: a council that can be corrected mid-sentence

Here's the move worth borrowing. A council that argues from memory can only ever be as right as its training. A council that argues from evidence can be corrected mid-sentence.

Tool-MAD makes three changes to get there, and each maps onto a question anyone designing a real council has to answer.

1. Heterogeneous tools, not just heterogeneous models. The diversity argument in multi-agent systems usually stops at the models — put a Claude and a GPT in the room because they fail differently. Tool-MAD extends diversity to what each agent can reach. One agent's search API and another's RAG retriever surface different evidence for the same claim, which means the agents disagree about facts, not just framing. That's a sharper kind of disagreement than two models reciting different priors.

2. Adaptive query formulation. This is the real upgrade over MADKE. Instead of one retrieval up front, Tool-MAD reformulates its queries as the debate moves — when an agent raises a new sub-claim, the system goes and looks for evidence on that. Retrieval follows the argument. It's the difference between researching a question once and researching it the way a careful person actually does: each answer raises the next thing you didn't know to look up.

3. A Judge that scores faithfulness, not votes. Most debate frameworks resolve by tallying — majority wins, or the last model standing. Tool-MAD's Judge agent folds two quantitative signals into the final call: a Faithfulness score (does the response actually follow from the retrieved evidence?) and an Answer Relevance score (does it address the question asked?). The Judge isn't counting hands. It's checking whether each argument is grounded — and using that to detect hallucination directly.

That third piece is the one that should feel familiar. A synthesis layer whose job is to weigh grounding rather than average opinions is the Chairman pattern by another name.

What the numbers say, and what they don't

Across four fact-verification benchmarks, Tool-MAD reports up to a 5.5% accuracy improvement over state-of-the-art multi-agent debate frameworks. In medically specialized domains, it holds up across different tool configurations — the robustness check that matters most, because medicine is exactly where a confident wrong answer is expensive.

5.5% is not a number that ends an argument. It's a number that tells you the architecture is doing something the alternatives weren't — and the something is the interesting part. The gain doesn't come from a smarter model or more debate rounds. It comes from changing what the agents are allowed to consult, and when. The lift is structural.

That's the recurring lesson of the 2026 council literature, stated again in a new key. Adding agents doesn't help. Adding rounds doesn't reliably help. What helps is structure — who retrieves what, when the evidence refreshes, and what the synthesis layer is actually scoring.

The part Tool-MAD makes concrete

There's a quiet reframe buried in this paper that's worth pulling out. We tend to talk about an AI council's diversity in terms of which models are in the room. Tool-MAD is a reminder that a council is also defined by what each member can see. Two instances of the same model with different tools will surface different facts and argue to different verdicts. Heterogeneity isn't only a property of the weights — it's a property of the evidence each seat can reach.

That's not an abstraction for us. When a question turns on facts — is this claim true, does this source support it, what does the current evidence actually say — the useful council isn't four models reciting their training. It's a structured deliberation where members consult different sources, the evidence updates as the argument develops, and a synthesis step asks whether each answer is grounded before it commits to one. The seven strategies exist because "have them debate" isn't a design; who checks what, and how the result gets adjudicated is the design.

Tool-MAD is one more piece of evidence that the field is converging on the same conclusion from the research side: the value was never in the number of voices. It was in giving them something real to argue over, and someone whose job is to check the work.

If you've got a question that turns on facts you can't afford to get wrong, that's the kind of thing a council is for. Try it free — no signup. shingik.ai