Useful information
Prime News delivers timely, accurate news and insights on global events, politics, business, and technology
Do you want smarter ideas in your entrance tray? Register in our weekly newsletters to obtain only what matters to the leaders of AI, data and business security. Subscribe now
Japanese laboratory Samán It has introduced a new technique that allows multiple models of large language (LLM) to cooperate in a single task, effectively creating a “dream equipment” of AI agents. The method, called Multiple AB-MCTIt allows models to perform tests and errors and combine their unique strengths to solve problems that are too complex for any individual model.
For companies, this approach provides a means to develop more robust and capable systems. Instead of being locked in a single supplier or model, companies could dynamically take advantage of the best aspects of the different border models, assigning the correct AI for the right part of a task to achieve higher results.
The power of collective intelligence
The border models are quickly evolving. However, each model has its own different strengths and weaknesses derived from their unique training data and training architecture. One could excel in coding, while another stands out in creative writing. Sakana AI researchers argue that these differences are not a mistake, but a characteristic.
“We see these varied biases and skills not as limitations, but as precious resources to create collective intelligence,” says the researchers in their Blog. They believe that just as the greatest achievements of humanity come from various teams, artificial intelligence systems can also achieve more working together. “By grouping their intelligence, AI systems can solve unsurpassed problems for any unique model.”
Thinking more time at the time of inference
The new AI Sakana algorithm is a “inference time” scale (also known as “test time scaling”), a research area that has become very popular in the last year. Although most of the AI approach has been in the “scale in training time” (making the models larger and trained in larger data sets), the inference time scale improves performance by assigning more computational resources after a model is already trained.
A common approach implies the use of reinforcement learning to boost models to generate longer and detailed sequences of the thought chain (COT), as seen in popular models such as Openai O3 and Deepseek-R1. Another simplest method is repeated sampling, where the model has the same notice several times to generate a variety of potential solutions, similar to a rain of ideas. Sakana’s work combines and advances these ideas.
“Our frame offers a more intelligent and more strategic version of Best-OF-N (also known as repeated sampling),” Takuya Akiba, a research scientist from Sakana AI and co-author of the newspaper, told Venturebe Tabeat Takuya. “It complements reasoning techniques such as Long Cot through RL. By dynamically selecting the appropriate search strategy and LLM, this approach maximizes performance within a limited number of calls LLM, offering better results in complex tasks.”
How the search for adaptive branch works
The nucleus of the new method is an algorithm called Mount Carlo Tree of Adaptive Ramification (AB-MCTS). It allows a LLM to effectively perform the test and error by intelligent balance two different search strategies: “Looking more deeply” and “looking broader.” Searching more deeply implies taking a promising response and refining it repeatedly, while seeking broader means generating completely new solutions from scratch. AB-MCTS combines these approaches, allowing the system to improve a good idea, but also to pivot and try something new if it reaches a dead end or discovers another promising direction.
To achieve this, the system uses Monte Carlo Tree Search (MCT), a decision -making algorithm by Deepmind Alphago. In each step, AB-MCTS uses probability models to decide whether it is more strategic to refine an existing solution or generate a new one.
The researchers took this one step further with the AB-MCT Multi-Llm, which not only decides to “” refine in front of generating) but also “what” LLM should do it. At the beginning of a task, the system does not know which model is the most appropriate for the problem. It begins by testing a balanced mixture of LLM available and, as you advance, learn which models are more effective, assigning more than the workload over time.
Testing the ‘dream team’ of AI
The researchers tested their Multi-Llm AB-MCTS system in the ARC-AGI-2 reference. ARC (abstraction and reasoning corpus) is designed to prove a human being to solve new visual reasoning problems, which makes it notoriously difficult for AI.
The team used a combination of border models, including O4-mini, Gemini 2.5 Pro and Deepseek-R1.
The models collective could find correct solutions for more than 30% of the 120 test problems, a score that significantly exceeded any of the models that worked alone. The system demonstrated the ability to dynamically assign the best model for a given problem. In the tasks where there was a clear route to a solution, the algorithm quickly identified the most effective LLM and used it more frequently.
More impressive, the team observed instances in which the models solved problems that were previously impossible for any of them. In one case, a solution generated by the O4-mini model was incorrect. However, the system passed this defective attempt from Deepseek-R1 and Gemini-2.5 Pro, which could analyze the error, correct it and, ultimately, produce the correct answer.
“This shows that Multi-LLM AB-MCT can combine border models in a flexible way to solve previously insoluble problems, which drives the limits of what can be achieved by using LLM as a collective intelligence,” the researchers write.
“In addition to the pros and individual cons of each model, the tendency to hallucinate can vary significantly among them,” Akiba said. “By creating a set with a model that is less likely that Alucine, it could be possible to achieve the best of both worlds: powerful logical capacities and a strong base. Since hallucination is an important problem in a commercial context, this approach could be valuable for its mitigation.”
From research to real world applications
To help developers and companies to apply this technique, Sakana AI has launched the underlying algorithm as an open source framework TrebyquestAvailable under an Apache 2.0 license (usable for commercial purposes). Treequest provides a flexible API, which allows users to implement Multiple LLM AB-MCT for their own tasks with personalized qualification and logic.
“While we are in the early stages of applying AB-MCT to specific problems oriented to companies, our research reveals significant potential in several areas,” Akiba said.
Beyond the ARC-AGI-2 reference point, the equipment was able to successfully apply AB-MCT to tasks such as complex algorithmic coding and improve the accuracy of automatic learning models.
“AB-MCT could also be highly effective for problems that require iterative tests and errors, such as optimizing existing software performance metrics,” Akiba said. “For example, it could be used to automatically find ways to improve the latency of a web service.”
The launch of a practical and open source tool could pave the way for a new class of more powerful and reliable business applications.
