David Ha’s Post

Inference-Time Scaling and Collective Intelligence for Frontier AI Blog: https://sakana.ai/ab-mcts We developed AB-MCTS, a new inference-time scaling algorithm that enables multiple frontier AI models to cooperate, achieving promising initial results on the ARC-AGI-2 benchmark. Frontier AI models like ChatGPT, Claude, Gemini, Grok, DeepSeek, and Qwen are evolving at a breathtaking pace. While each model requires enormous resources to develop, fierce market competition has made them become low-cost commodities that can be deployed at scale, at the same time ensuring that there will never be a single “winner” in AI. This is because no matter how advanced each model becomes, each model retains its own individuality stemming from its unique training data and methods. We see these biases and varied aptitudes not as limitations, but as precious resources for creating collective intelligence, further enhancing the performance of frontier models. Just as humanity’s greatest achievements arise from the collaboration of diverse minds, we believe the same principle applies to AI. Like a team of human experts tackling difficult problems, AIs should also collaborate by bringing their unique strengths to the table. The AB-MCTS (Adaptive Branching Monte Carlo Tree Search) algorithm we are introducing is a concrete step toward realizing this vision. AB-MCTS is a method for inference-time scaling that enables frontier AIs from commoditized, low-cost providers like OpenAI, Google and DeepSeek to cooperate and efficiently perform trial-and-error. In particular, our AB-MCTS combination of o4-mini + Gemini-2.5-Pro + DeepSeek-R1-0528, current frontier AI models as of writing, achieves strong performance on the ARC-AGI-2 benchmark, outperforming individual o4-mini, Gemini-2.5-Pro, and DeepSeek-R1-0528 models by a large margin.