Sakana AI Unveils TreeQuest: A Game-Changing Approach to Multi-Model LLM Collaboration

If you think AI is all about building bigger and bigger models, Sakana AI is here to shake things up. The Tokyo-based startup has just introduced TreeQuest, a clever new method that lets teams of large language models (LLMs) work together—kind of like an orchestra led by a conductor, rather than relying on a single soloist to do all the work.

So what’s the secret sauce? TreeQuest borrows a technique from the world of game-playing AIs—Monte-Carlo Tree Search (MCTS), the same strategy that powered AlphaGo. But instead of a chess board or Go stones, TreeQuest uses this approach to organize multiple LLMs in real time, picking the best “lead” model for each part of a problem and letting the others chip in where their strengths are needed most. Rather than sticking with whatever answer a single model spits out, TreeQuest treats problem-solving as a team sport, making decisions collaboratively at every step.

And the results are worth paying attention to. According to Sakana AI’s testing, model groups coordinated by TreeQuest outperformed the best single LLMs by a solid 30% on a range of benchmarks. That includes tough tasks like twisting reasoning puzzles, generating computer code, and answering complex, multi-step questions—areas where LLMs often stumble when working alone. Instead of one model getting stuck, another might run with an imperfect answer and drive the team to a correct solution.

Why should this matter to anyone outside the lab? If you care about making AI more affordable and efficient—think business, research, consumer tech or pretty much anything that uses smart software—TreeQuest offers an intriguing path. Blending the talents of smaller or specialized models could lead to better performance for less computational cost, sidestepping the expensive arms race of endlessly scaling up a single model. In other words: it’s collaboration over brute force.

This is still early days for TreeQuest, but Sakana AI’s approach is already getting attention in the industry. Their plan is to keep improving the method, explore how it interacts with other AI systems, and see how far the “many minds, one task” concept can go. If they’re successful, AI developers may soon see teamwork as a core principle when designing the most advanced models, rather than always chasing the biggest possible ones.

Read the full story here: VentureBeat.