Three mediocre craftsmen surpass a Zhuge Liang: OAI/Google/DeepSeek team up for the first time "AI Dream Team"! Power increased by 30%, crushing all single models.

Recently, a study from Sakana AI attempted to integrate the abilities of three different models in the reasoning process, rather than in the construction process. The results were surprising, as the integrated model's abilities far exceeded those of individual models, and the performance of the three-model combination was better than that of just two models combined. Sakana AI used a new reasoning scaling algorithm, the adaptive branch Monte Carlo tree search AB-MCTS. By combining the three cutting-edge AI models, o4-mini, Gemini-2.5-Pro, and R1-0528, using AB-MCTS, they achieved remarkable results in the ARC-AGI-2 benchmark test. The scores of the multi-model far surpassed those of the individual o4-mini, Gemini-2.5-Pro, and DeepSeek-R1-0528 models.