Sakana AI gets rival frontier models to team up

AB-MCTS is an inference-time search algorithm designed to make competing frontier models work together as a single problem-solving team. Instead of relying on a single AI to solve tough abstract reasoning tasks, it orchestrates multiple models to explore solutions through collaborative trial and error.

The system stands for Adaptive Branching Monte Carlo Tree Search. Rather than merging model weights during training, it scales intelligence at inference time by treating diverse frontier models as distinct specialists.

How it coordinates multiple models

Frontier models like ChatGPT, Gemini, and DeepSeek come with different reasoning styles and distinct blind spots. AB-MCTS turns those differences into an advantage by structuring problem-solving into a tree search.

During each step of the search, the algorithm adaptively branches out. One model might propose candidate actions, while another evaluates them or explores alternative paths. The tree search dynamically manages when to exploit promising leads and when to switch models to break out of dead ends.

Sakana AI noted that this builds on their 2024 work on evolutionary model merging. The difference is a deliberate shift from mixing weights to create a new model, to orchestrating existing, ready-to-use frontier models at runtime.

Benchmark results and open-source availability

On the challenging ARC-AGI-2 benchmark, Sakana AI tested a combination of o4-mini, Gemini 2.5 Pro, and DeepSeek-R1-0528. The cooperative system outperformed every individual model by a substantial margin.

The project is fully open source. The underlying search algorithm is available in the TreeQuest repository on GitHub, while the exact evaluation setup is published under the ab-mcts-arc2 repository. The technical paper is also live on arXiv.

To run it yourself, you will need API access to the respective frontier models and a Python environment to configure the search parameters.

Why it matters

We think this is one of the most compelling demonstrations of collective intelligence applied to inference-time scaling. While most labs are focused on making single models think longer, single models still suffer from persistent internal biases. Throwing radically different model families into a unified search tree breaks those echo chambers.

If you are building advanced reasoning agents or tackling complex planning benchmarks, TreeQuest is well worth experimenting with. The obvious trade-off is inference cost and latency from calling multiple frontier APIs, but for tasks where solving the problem matters more than speed, this multi-model framework delivers an immediate edge.