Sakana AI’s TreeQuest: Deploy multi-model teams that outperform individual LLMs by 30%

Japanese EU Laboratory Straw Very large language models (LLS) presented a new technique that allows the EI agents to work in a position to effectively create a “dream team”. Method, called Multi-LLM AB-MCTSModels allow you to experiment and make a mistake and combine their unique strengths to solve problems with any individual model.

This approach for enterprises provides a tool to develop more robust and skilled AI systems. Instead of locking a single provider or model, enterprises can properly determine the correct AI to achieve superior results using different border models.

The power of collective intelligence

Border AI models are developing rapidly. However, each model has different strengths and weaknesses obtained from their unique teachings and architecture. In coding, the other can correct in a way that is another superior in the creative writing. Sakana EU researchers claim that these differences are wrong, but it is a feature.

“We see these biases and various abilities as valuable resources to create a collective intelligence, but also in themselves Blog Post. They believe that when it comes to various teams of humanity, the AI ​​systems can achieve more by working together. “By collecting intelligence, AI systems can solve solutions for any model.”

To think longer than infertility on time

Sakana Ai’s new algorithm “result time scale” technique (also called “Test time scale“), A very popular area of ​​research in the last year. In the AI,” teaching training models in a larger and larger database), allocating more sources after a model, improves the performance by allocating more sources.

A common approach covers the use of the reinforcement learning that does not want models to prepare longer, more detail philosophy (COT), Openai O3 and sequences as seen in popular models DeepSeek-r1. Another, simple method, the model of the model is the same proposal to create various potential solutions similar to the brainstorm session. Sakana EU’s job combines and develops these ideas.

“Our frame, the best N (aka repetition),” Takuya Akiba, “Takuya Akiba, research scientist Sakana AI and co-author of the platform of the paper of the paper” The author of the author of the co-author of the paper “offers a more strategic version. “This complements justification methods as a long cot through the RL.

How do adaptive branch searches work

The nucleus of the new method is an algorithm called Monte Carlo Tree Work (AB-MCTS) called Adaptive Branch. Balancing two different search strategies, test-error and “deeper” and “more extensive search” and “wider search”. In a deeper search, a promising response and repeatedly clears, and when conducting a wider investigation, it means to get completely new solutions from scratch. Ab-MCTS combines these approaches, and trying to improve a good idea of ​​the system, but also strikes something new if you have a dead end or detect another promising direction.

The system uses the system to implement it Monte Carlo Wood Search (MCTS) is used popular by the algorithm of a decision Deepmind’s Albago. In every step, the EUMTS uses probability models to decide whether it is more strategic to purify or create a new solution.

Researchers do this step by many-LLM AB-MCTs only “what to do” or create “how much to do” or create rafinat, but “which” should “do” LLM. At the beginning of a task, the system does not know which model is best for the problem. As this is a balanced mix of existing LLMs and progresses, as the models are more effective, learn more of the workload.

Put the AI ​​’Dream Team’ test

Researchers tested the multi-LLM AB-McTS system ARG-AGI-2 EVALUATION. ARC (Abstraction and Rationale Corpus) is designed to test a person similar to a person to solve visual thinking problems, makes it notoriously difficult for AI.

The team used the combination of border models, including o4-mini, Gemini 2.5 Proand Deepseek-R1.

The staff of the models was able to find the correct solutions for more than 30% of the 120 test problems, the correct solutions for the significantly preferred any of the odd models. The system demonstrated the ability to determine the best model for a particular problem dynamically. In positions, the algorithm in the clear path of a solution, the algorithm quickly identified the LLM and used it more often.

More effectively, the team observed the cases that the models have solved the problems that are not possible for any of them. In one case, a solution created by the O4-mini model was wrong. However, this is a defective attempt by DeepSEEK-R1 and Gemini-2.5 Pro that can analyze the error and eventually provides the correct answer.

“This demonstrates that multinational EUMTs can agile border models to solve pre-resolved issues, not developing problems with collective intelligence.

“In addition to the individual positive and disadvantages of each model, it can change significantly between the tendencies of the hallucinat,” he said. “With the creation of an ensemble with a model with a model, it may be possible to get the best of both worlds: strong logical opportunities and powerful reasonable.

From the study to real world applications

Sakhana AI, which applies this technique and helps enterprises and enterprises, left the main algorithm as an open source frame TreequestAvailable under an Apache 2.0 license (can be used for commercial purposes). Treequest provides a flexible API that allows users to apply numerous ab-mcts to special goals and logic.

“Although the EU-MCTS is in the first stages of application of special business-oriented problems, our research reveals its significant potential in several areas,” he said.

Outside the ARG-AGI-2 assessment, the team was able to successfully impose EU-MCTs to successfully increase EU-MCTs to increase the accuracy of complex algorithmic coding and machine learning models.

“AB-MCTS can also be highly effective for problems that require an iterative test and error, such as optimizing the performance measurements of existing software,” the EUMS said. “For example, a website can be used to find a way to improve the delay in response.”

The release of a practical, open source vehicle can lead to a new class of AI apps, which is more powerful and reliable.