Lates News

A recent study from Sakana AI attempts to integrate the capabilities of three models during the reasoning process, rather than during construction. Surprisingly, the integrated model's capabilities far exceed those of individual models, outperforming even the performance of only two models combined. Sakana AI uses a new reasoning time Scaling algorithm Adaptive Branching Monte Carlo Tree Search (AB-MCTS). By using AB-MCTS to combine the o4-mini, Gemini-2.5-Pro, and R1-0528, three of the most advanced AI models currently available, the study achieved impressive results in the ARC-AGI-2 benchmark test. The multi-model score far surpassed the individual o4-mini, Gemini-2.5-Pro, and DeepSeek-R1-0528 models. (Synced).