Preprint finds independent model panels improve reasoning with fewer tokens

A preprint reports that selectively asking different AI models to solve a question independently improved reasoning results while using up to 47.2% fewer tokens than a multi-agent baseline. The method, called DIA, outperformed full chain-of-thought reasoning in 12 of 14 settings across seven benchmarks and two base models, including a 7.00% gain on MuSiQue. DIA activates when intermediate and final predictions disagree, then aggregates answers from a separate panel that sees neither prediction. It activated on approximately 17% of examples. Ablation tests found that both independent solving and differences among the panel's models were necessary for the gains.