图灵奖得主加持，蒙特卡洛树搜索×扩散模型杀回规划赛道｜ICML 2025 Spotlight

Core Insights - The article discusses the introduction of a new model called Monte Carlo Tree Diffusion (MCTD), which combines Monte Carlo Tree Search (MCTS) with diffusion models, achieving a 100% success rate in maze navigation tasks [4][3]. Group 1: MCTD Overview - MCTD addresses the limitations of traditional diffusion models in long-range reasoning by integrating MCTS's exploration capabilities with the global consistency of diffusion models [8][4]. - The model balances exploration and exploitation by dividing trajectories into sub-plans, allowing for differentiated denoising scheduling [8][12]. Group 2: Experimental Results - MCTD demonstrated near 100% success rates across various maze sizes, significantly outperforming baseline methods [17]. - In robotic arm tasks, MCTD-Replanning improved success rates from 22% to 50% in multi-block scenarios [19]. - The model's performance in visual mazes indicates robustness in high-dimensional perceptual spaces [20]. Group 3: Efficiency Improvements with Fast-MCTD - Fast-MCTD was introduced to address the high computational costs of MCTD, achieving up to 100 times faster inference in specific tasks [25][40]. - The model incorporates parallel processing and trajectory coarsening to enhance efficiency while maintaining performance [29][35]. - In maze navigation tests, Fast-MCTD achieved significant speed improvements of 80-110 times with minimal performance loss [36]. Group 4: Authors and Research Background - The primary authors of the papers are Jaesik Yoon and Sungjin Ahn from KAIST, with Ahn also affiliated with New York University [41][43].