A new study says AI agents should stop judging their teammates by whether they got the right answer.
Researchers tested a method called Diverse Hypothesis Deliberation, which generates five independent answers to a problem, then reruns a downstream "integrator" model with each answer either shown or hidden. This isolates what the researchers call trajectory value: whether a message actually helps later reasoning, separate from whether it was correct. Testing two open-source model families, gpt-oss-120b and gemma-4-31B-it, across five math and science benchmarks, they found wrong answers that improved the final result in every single benchmark-model pairing. Of the wrong answers that changed the final outcome at all, more than 40 percent changed it for the better, and repeat trials show this isn't noise (p=0.0002).
This matters because most multi-agent AI setups use confidence scores or majority agreement to decide which messages to trust, on the assumption that likely-correct answers are the only ones worth keeping. This research suggests that assumption quietly discards useful work. A wrong answer can still hand the system a solid decomposition of the problem or a correct constraint, and the study finds keeping the reasoning behind a wrong message helps more than keeping just its final (wrong) answer.
It's a small, open-model study, not a production system, but it points at the same lesson code review has always taught: a wrong conclusion built on sound reasoning is often more useful than a lucky right answer with no shown work.