A new paper argues AI models reason better when you exploit their disagreements instead of averaging them away.
Researchers describe Divergent-Convergent Reasoning (DCR), a two-phase method for squeezing more out of test-time compute. First, a model generates multiple candidate solutions to a problem. Then a reconciliation phase reviews those candidates and produces a final answer, rather than simply taking a majority vote. That distinction matters because majority voting fails when the correct answer is a minority pick among the generated candidates, while DCR's reconciliation step can still recover it. A recursive version of the method, which repeatedly re-analyzes disagreement and directs extra compute to the hardest cases, reached 93.3% on the AIME 2024 math competition and 92.0% on AIME 2025, using roughly 27% less compute than fixed-budget baselines.
The interesting part is what disagreement is doing here. Instead of treating varied model outputs as noise to be smoothed over, the researchers use a dispersion metric showing that more disagreement among candidate answers predicts bigger gains from reconciliation. That reframes disagreement as a signal for where compute should go, a more targeted approach than just generating more guesses and hoping a majority vote sorts it out.
AIME is a competition math benchmark, tidy and cleanly scored by design, so whether disagreement-driven reconciliation holds up on messier, real-world reasoning tasks without a single clean answer to converge on is still an open question.