Most AI councils run on a hidden assumption: more talk is better. Every model answers, every answer gets sent to every other model, and the room reasons its way toward something. A new paper from Deakin University says that default is the bug. Broadcasting everything, every round, doesn't sharpen the debate. It floods it.
The fix isn't a smarter model. It's a better filter on what each model is allowed to hear.
The paper is called "Hear Both Sides" (arXiv:2603.20640), and its one-line thesis is worth pinning to the wall: what agents hear is as important as what agents say.
The problem is the wire, not the roster
The entire pitch of a multi-model council is the roster. Two hundred models, five frontier labs, pick your panel. Roster is the thing every council product markets, and it feels like the obvious lever. Add a better model, get a better answer.
The last two years of research keep landing somewhere less flattering. When every member hears every other member's full answer at every round, the redundant messages pile up. A confident-but-wrong answer gets repeated, echoed back, and quietly voted in. The noise doesn't cancel out. It compounds. You didn't build a council. You built an echo with a headcount.
The paper is blunt about why the usual patch fails. Existing systems try to trim the noise by scoring each message's confidence and dropping the low-confidence ones. But confidence scores are miscalibrated, and the whole thing hinges on picking a threshold. A model can be confidently wrong, and often is. Filtering on how sure a model sounds just launders the noise through a number.
The move: keep the disagreement, drop the duplication
Here's what DAR — Diversity-Aware Retention — does instead. Each round, it looks at every agent's answer and keeps only the subset that maximally disagrees — both with each other and with the majority vote. Those are the messages it broadcasts. Everything redundant gets dropped before it ever reaches the other models.
Consensus is the residue. Disagreement is the payload.
And it forwards those dissenting messages untouched. An index-based mechanism passes the original text through without modification, so the disagreement the other models see is the real thing, not a summary that already smoothed off the edges. The point is to keep the sharpest opposing case intact long enough for the room to actually reckon with it.
The numbers
DAR was the top performer in 42 of 48 settings — four models (Qwen2.5-1.5B, Qwen2.5-3B, Falcon3-7B, Llama3.1-8B) run across six benchmarks (Arithmetics, GSM8K, HH-RLHF, Formal Logic, Professional Medicine, CSQA). Its average edge over the second-best method was 1 to 3 percent.
Small in isolation. But the shape is the story. The gains got bigger as the council got bigger — most pronounced at eight agents, exactly where naive debate gets worse. Noise accumulates fastest when there are more voices in the room, and that is precisely where a retention policy earns its keep. Scaling the member count without one isn't scaling the council. It's scaling the interference.
It's also cheaper. Dropping the redundant messages cut the message count by up to 32% (Qwen2.5-3B on Arithmetics), for a modest latency increase — 3 to 14 percent, about 9 on average. You send less across the wire and you get more out of it.
Why this reframes what a council is for
Put DAR next to how most "multi-agent" products actually work and the gap is obvious. The standard design treats disagreement as a problem to resolve — average the answers, take the vote, hand the user a clean consensus. DAR treats disagreement as the one thing worth transmitting, and the majority vote as the thing to route around. That isn't a tuning tweak. It's an inversion of what the council is for.
It also lands on a distinction the field has been circling. An earlier controlled study found the sequencing knobs — who speaks first, whether models see each other's confidence, how many rounds — barely move the needle. This paper says a different knob does: not the order of the messages, but which messages propagate at all. Sequencing is cosmetic. Channel policy is structural.
That instinct is how we think about running councils at Shingikai. A council isn't a group chat. The interesting design surface was never "how many models" — it's what crosses the wire between them, and what the synthesis layer does with the disagreement it finds. When we run Red Team vs Blue Team, the whole job is to keep the strongest opposing case standing long enough to stress-test it, not to sand it down toward agreement. DAR is the measured, formal version of that: keep the dissent, drop the duplication.
One honest limit. This was tested on small open models and reasoning benchmarks, and it's still a preprint — no conference stamp yet. Don't read it as a law. Read it as one more entry in a pattern that's getting hard to ignore: more than one lab this year has landed on a version of the same point. The lever in a multi-model system is the channel between the members, not the size of the cast.
The close
The reflex when a council underperforms is to reach for a smarter model. The research keeps answering back with the same thing: the problem isn't the models. It's everything they're forced to hear.
The best thing a council can do with most of what its members say is not repeat it.
You can watch one actually argue — models pushing back on each other in real time, the disagreement kept visible instead of averaged away — at shingik.ai. Free to try, no signup.