Random peer pairing beats learned routing in agent debate

Multi-agent debate systems have leaned on learned or adaptive topologies to decide which agents talk to which. A new paper tests a much dumber approach: each round, every agent gets paired with two distinct peers, sampled fresh and at random, no training involved.

That random pairing consistently beats the learned and adaptive topology methods on the accuracy-cost trade-off, according to the paper. Adding a lightweight stopping rule, ending debate early once agents converge, cuts cost further without giving back the accuracy gain.

The stakes are practical before they're theoretical. Teams building debate or voting pipelines for LLM agents have been spending training runs and engineering time on topology selection, the idea being that smarter routing between agents should pay for itself. If two random peers per round gets you there for less, that spend was buying complexity, not performance.

The authors' own framing is worth sitting with. They're not arguing random routing is secretly optimal. They're arguing the field hasn't been checking its work: benchmark any new topology method against random sampling before you accept the added complexity, because right now most of that literature doesn't.

What the paper doesn't settle is why learned routing underperforms here. Is it overfitting to whatever benchmark it was trained on, or is topology just a weaker lever than people assumed once you're already running multiple rounds of debate? Worth checking is whether this holds at agent counts and round counts beyond what the paper tested, since sparse random sampling gets less random as the peer pool shrinks.