A 15,336-question study found multi-agent systems generate the right answer and then vote for a wrong one.

The authors separated multi-agent reasoning into candidate generation, peer communication and terminal selection, held two fixed to isolate the third, and replayed 81,390 fixed candidate pools drawn from 16,278 questions across five benchmarks (arXiv 2608.25937). A correct answer is frequently already in the pool while the system converges on a wrong one, a failure they call memetic drift. Judge reliability turns out not to be a fixed model trait but to vary with task, generator, and how rare the correct answer is. Combining answer frequency with the judge's verdict, changing only the selection rule, lifts accuracy.