AB-MCTS: Enabling Collective Intelligence Among Frontier AI Models

At Sakana AI, we've been exploring how to leverage the collective intelligence of multiple AI models to achieve better performance during inference. Building on our 2024 research on evolutionary model merging, which combined existing open-source models through evolutionary computation and model merging, we now present AB-MCTS (Adaptive Branching Monte Carlo Tree Search). This new algorithm allows multiple frontier AI models to cooperate effectively, significantly outperforming individual models on the ARC-AGI-2 benchmark.

What Changed Technically?

AB-MCTS is a novel inference-time scaling algorithm designed to enable multiple AI models to work together in a coordinated manner. Here’s how it works:

• Adaptive Branching: AB-MCTS dynamically adjusts the branching factor of the search tree based on the confidence and diversity of responses from different models. This ensures that the most promising paths are explored more thoroughly.
• Monte Carlo Tree Search (MCS): MCTS is a best-first search algorithm that uses random sampling to explore the search space. AB-MCTS enhances this by incorporating feedback from multiple models, allowing it to make more informed decisions.
• Collective Intelligence: By combining the outputs of multiple models, AB-MCTS can leverage the strengths of each model, effectively creating a "super-model" that performs better than any individual component.

Why It Matters

The ability to combine multiple frontier AI models during inference is significant for several reasons:

• Improved Performance: The combined intelligence of multiple models often outperforms single models, especially on complex tasks.
• Flexibility: AB-MCTS can be applied to a wide range of models and tasks, making it a versatile tool in the AI practitioner's toolkit.
• Scalability: As new and more powerful models are developed, they can easily be integrated into the AB-MCTS framework, ensuring that the system remains cutting-edge.

Initial Results

We tested AB-MCTS on the ARC-AGI-2 benchmark, which is designed to evaluate AI systems' ability to solve complex reasoning tasks. The combination of o4-mini, Gemini-2.5-Pro, and R1-0528-all current frontier AI models-achieved impressive results:

• Pass@k: AB-MCTS significantly outperformed individual models in terms of the number of correct answers (Pass@k) as a function of the number of LLM calls.
• Efficiency: The algorithm efficiently balances exploration and exploitation, ensuring that the most promising solutions are found with fewer computational resources.

Implementation Details

For those interested in the technical details, here’s a breakdown of the AB-MCTS implementation:

• Search Tree Construction: The search tree is constructed using MCTS, where each node represents a state in the problem space.
  • Node Selection: Nodes are selected based on the Upper Confidence Bound (UCB) formula, which balances exploration and exploitation.
  • Expansion: New nodes are added to the tree by expanding the most promising leaf nodes.
• Model Integration:
  • Querying Models: Each model is queried for its response to the current state.
  • Aggregation: The responses are aggregated using a weighted average, where weights can be adjusted based on the model's performance and confidence.
• Adaptive Branching:
  • Confidence Estimation: The algorithm estimates the confidence of each model’s response using statistical methods.
  • Branching Factor Adjustment: Based on the confidence and diversity of responses, the branching factor is dynamically adjusted to focus on promising paths.

Conclusion

AB-MCTS represents a significant step forward in leveraging collective intelligence among AI models. By enabling multiple frontier models to cooperate during inference, we can achieve better performance and solve more complex problems. The initial results on the ARC-AGI-2 benchmark are promising, and we look forward to further advancements in this area.

For more details, you can check out our paper and implementation:

• Paper: https://arxiv.org/abs/2503.04412
• Algorithm Implementation: https://github.com/SakanaAI/treequest
• ARC-AGI-2 Experiment