Three planners comparing the generated plans and actual rollouts: diffusion, diffusion forcing, and Monte Carlo Tree diffusion (MCTD). Diffuser and diffusion forcing cannot create a successful trajectory plan, but MCTD succeeds by adaptively improving the plan. Credit: Monte Carlo Tree Spread
Diffusion models are widely used in many AI applications, but there is a lack of research into efficient inference time scalability, particularly for inference and planning (known as system 2 capabilities). In response, the researchers developed new technologies that allow for high-performance and efficient inference for planning based on diffusion models.
This technology demonstrated performance by achieving a 100% success rate on huge maze resolution tasks where existing models are not successful. The results are expected to serve as core technology in a variety of areas that require real-time decision-making, such as intelligent robots and real-time-generated AI.
The research team led by Professor Seungjin Anne of the Computing School has developed a new technology that significantly improves the inference time scalability of diffusion-based inference through collaboration with Professor Joshua Bengio of the University of Montreal, a world-renowned scholar of deep learning. This study was conducted through the Prefrontal AI Collaboration Center as part of a collaboration between Kaist and Mila (Quebec AI Institute).
This technology is gaining attention as a core AI technology that allows AI to efficiently utilize more computational resources during inference and solve complex inference and planning problems that cannot be addressed simply by increasing the data or model size. However, current diffusion models used in various applications do not have an effective methodology for implementing such scalability, particularly for inference and planning.
To address this, Ahn’s research team collaborated with Professor Bengio to propose a new diffusion model inference technique based on the Monte Carlo Tree search. This method is designed to investigate the diverse generation paths during the tree structure’s diffusion process and efficiently identify high-quality outputs even with limited computational resources. As a result, I achieved a 100% success rate on the “giant maze resolution” task, where the previous method had a success rate of 0%. This work is published on the ARXIV preprint server.
In a follow-up study, also posted to Arxiv, the team managed to significantly improve the major drawbacks of the proposed method. This means slow speed. By efficiently parallelizing tree searches and optimizing computational costs, we achieved results of up to 100 times faster or better quality than previous versions. This makes a lot of sense as it simultaneously demonstrates the inference capabilities of methods and real-time applicability.
Professor Sungjin Ahn added, “This research fundamentally overcomes the limitations of existing planning methods based on diffusion models that require high computational costs,” adding, “it could serve as a core technology in a variety of fields, such as intelligent robotics, simulation-based decision-making, and real-time generation AI.”
Details: Haesik Yoon et al, Monte Carlo Tree Spreading System 2 Plan, ARXIV (2025). doi:10.48550/arxiv.2502.07202
Jaesik Yoon et al., Fast Monte Carlo Tree Diffusion: 100x Speed Up Through Parallel Sparse Planning, ARXIV (2025). doi:10.48550/arxiv.2506.09498
Journal Information: arxiv
Provided by Korea Institute of Advanced Science and Technology (KAIST)
Quote: The new AI method increases inference and planning efficiency of diffusion models obtained from July 21, 2025 from https://2025-07-07-ai-method-ai-method-efficiency-diffusion.html (July 21, 2025)
This document is subject to copyright. Apart from fair transactions for private research or research purposes, there is no part that is reproduced without written permission. Content is provided with information only.
