Sakana AI unveiled a new open-source algorithm on Tuesday designed to facilitate collaboration among various artificial intelligence (AI) models for addressing intricate challenges. Named Adaptive Branching Monte Carlo Tree Search (AB-MCTS), this inference-time scaling algorithm introduces a novel dimension to the existing AI model framework. When encountering a new challenge, the algorithm not only evaluates whether to prioritize extended reasoning or broader exploration but also identifies the most appropriate AI model for the task. Furthermore, in cases of particularly complex issues, it has the capability to utilize multiple AI models simultaneously.
Sakana AI Releases Algorithm That Promotes Collaborative AI Thinking
In a recent post on X (previously Twitter), the Tokyo-based company emphasized that its new algorithm fosters a collaborative environment for AI, enabling advanced models like Gemini 2.5 Pro, o4-mini, and DeepSeek-R1 to work together effectively.
The initiative aims to address a considerable challenge in the AI sector: effectively merging the distinct strengths and mitigating the unique biases inherent in various AI models to enhance overall performance. Sakana AI has dedicated several years to researching this issue and presented findings on “evolutionary model merging” in a paper published in 2024.
Building on its research, the company has developed an algorithm that allows AI models to perform computations according to specific constraints, generate multiple outputs for diverse insights, and collaborate with other suitable AI models to achieve superior results.
During testing at the ARC-AGI-2 benchmark, researchers utilized a combination of o4-mini, Gemini-2.5-Pro, and R1-0528 with the AB-MCTS system, which outperformed the individual models. According to Sakana AI, while o4-mini independently solved 23 percent of the problems, it managed to solve 27.5 percent when operating within the AB-MCTS framework.
Sakana AI has made the TreeQuest algorithm available on its GitHub repository and also published separate ARC-AGI experiment results. Further details from the research can be found in a paper uploaded to arXiv.
