Sakana AI is teaching models to work as a team

July 25, 2025

The Japanese lab Sakana AI has developed an algorithm called AB-MCTS that enables different AI models to collaborate and solve complex problems together. Instead of relying on a single powerful model, the system uses multiple models—such as ChatGPT, Gemini, and DeepSeek—and dynamically assigns tasks based on each model’s strengths. Some models focus on strategy, others on coding, and all contribute to the final solution.

This approach has shown notable progress: the team-based models solved 30% of the ARC-AGI-2 benchmark tasks, while the best-performing individual models reached only 23%. Interestingly, the models are also capable of recognizing and correcting each other’s errors, which further improves the accuracy of results. Sakana has made this technology open-source through a tool called TreeQuest, allowing other researchers and developers to build their own multi-AI collaboration systems.

These kinds of experiments suggest that the future of artificial intelligence may not lie in a single all-knowing model, but in a network of specialized models that work together as a team.

In Brief: Tech World Highlights

NASA’s Mars rover Perseverance discovered a mysterious rock made up of hundreds of tiny spheres that resemble fossilized spider eggs, puzzling scientists.
Chinese researchers have developed WHERE-Bot, a new soft robot without wheels that safely navigates chaotic environments without using obstacle detection sensors.
Doctors in Gurugram, India, performed a complex heart surgery on a patient over 1,000 miles away in Bengaluru using the SSI Mantra surgical robot.
Baidu’s Apollo Go plans to launch the first 100 fully autonomous robotaxis in Dubai by the end of 2025, with plans to expand the fleet to at least 1,000 vehicles by 2028.
China’s Pudu Robotics unveiled the FlashBot Arm, a semi-humanoid robot designed for commercial environments such as hotels, restaurants, and retail spaces.