There is a decades-old result in the study of how groups think that says a group can never out-reason its best member. When people — or agents — talk and nudge each other toward agreement, every updated opinion is just a weighted average of the opinions already in the room. Average things together and you stay inside the range you started with. The group can't land on an answer nobody brought.
A new paper out of the University of North Texas watched AI models deliberate and caught them breaking that rule. (arXiv:2606.19494, "Hidden Anchors in Multi-Agent LLM Deliberation," Apurba Pokharel and Ram Dantu, June 17.)
Why the ceiling matters
"A council beats its best single member" is the entire promise — and the entire vulnerability — of multi-model AI. If a group of models can only ever settle somewhere between their starting answers, then a council is at best a fancy averaging machine: good for canceling noise, useless for finding an answer none of them had. Worth knowing whether that ceiling is real.
The classical models — DeGroot, Friedkin–Johnsen, the standard toolkit for how opinions move through a network — all enforce the ceiling by construction. Each round, an agent replaces its view with a blend of its neighbors' opinions, and in Friedkin–Johnsen a tug back toward its own opening position. Blend, blend, blend. No opinion can ever leave the convex hull — the geometric "box" drawn around everyone's initial positions. (Convex hull: the smallest region that contains every starting point.)
The escape
The paper's finding is that LLM deliberation leaves the box. An agent's confidence in the correct answer can climb past where any agent started — outside the hull, which the classical math forbids outright. Something is pulling the group that isn't coming from its neighbors.
The authors name that something a hidden anchor: an internal belief each model carries and keeps getting tugged toward, round after round, no matter what the other agents say. Two forces, not one. There's the herd pull toward consensus that the old models already captured — and there's the anchor pull toward a private conviction that they didn't.
The old math only modeled the room. This paper models what each member walked in already believing.
The anchor is recoverable
The sharp part: the anchor isn't a metaphor. The paper recovers it from the deliberation transcript alone. Fit the trajectory of an agent's stated positions across the rounds, and the hidden belief it's being pulled toward falls out of the math.
Then a test. Does that recovered anchor predict the agent's behavior in held-out runs it wasn't fit on? When it does, the model is genuinely anchor-driven. When it doesn't, it's mostly just following the herd. That's a model-selection test you can run, not a vibe you can assert.
Anchors are a spectrum
Run the test across three open-weight model families and the result isn't uniform. Anchor influence is a spectrum, not all-or-nothing. The strength of the pull is roughly comparable across families — but where the anchor sits differs. And the location decides everything.
When a model's anchor sits right on top of its opening answer, the old averaging math describes it fine, and the group never escapes the box. Only when the anchor sits far from the initial opinions does deliberation break out of the hull and require the full two-force model.
Read that twice, because it's the whole game for council design. The escape-the-box behavior — the good kind, the council-beats-its-best-member kind — happens precisely when a model is being pulled by a conviction that is not its first answer. The internal belief and the opening move come apart.
Both sides of the anchor
Here's the side-by-side. Ask one model and you get one anchor, and no amount of re-prompting will pull it off — there's no second force in the room. Put several models in a structured deliberation and you get the possibility of a result better than any single starting point. That possibility is the upside, and this paper is the cleanest dynamical-systems case for it published so far.
The same mechanism is the warning. A hidden anchor that has nothing to do with the evidence — a model's self-preference, an order effect, the answer it happened to say first — pulls just as hard as a good one. The math doesn't care whether the conviction is wise. It pulls the group regardless.
Which is why the structure around the deliberation is not decoration. Anonymizing which model said what, randomizing who answers first, forcing a critique round before a vote — these aren't manners. They're attempts to make sure the anchor that wins is the one attached to the strongest argument, not the loudest prior. Karpathy's llm-council anonymizes responses before ranking for exactly this reason. Same instinct, now with a dynamical-systems model underneath it.
What the council actually buys
The skeptic's version of the council story is that several models just average out each other's noise. This paper says something more interesting is on offer — and more dangerous. A council can genuinely arrive somewhere none of its members began. Whether it arrives somewhere better depends entirely on which hidden belief does the pulling.
A single model can only ever hand you back the answer it already had. A council can move past every starting point in the room — toward the right answer or the confident wrong one, depending on how you built it.
That "depending on how you built it" is the whole job. It's why a council earns its keep on the questions where being wrong is expensive — and why the design of the deliberation, not the headcount of models in it, is the part that matters.
Try it free — no signup. shingik.ai