Fast Monte Carlo Tree Diffusion: 100x Speedup via Parallel Sparse Planning

TL;DR

Fast-MCTD addresses the computational bottlenecks of diffusion-based long-horizon planning by introducing parallelized, redundancy-aware rollouts and trajectory coarsening. By combining Parallel MCTD (P-MCTD) with Sparse MCTD (S-MCTD), the method achieves up to 100× speedups over standard MCTD while maintaining or improving planning performance. The results show notable gains across maze navigation, robotic manipulation, and visual planning, with some tasks even beating diffusion-only baselines in inference speed. This work demonstrates that inference-time scalability and structured, long-horizon reasoning can be jointly achieved in diffusion-based planners, expanding their practical applicability in high-dimensional domains.

Abstract

Diffusion models have recently emerged as a powerful approach for trajectory planning. However, their inherently non-sequential nature limits their effectiveness in long-horizon reasoning tasks at test time. The recently proposed Monte Carlo Tree Diffusion (MCTD) offers a promising solution by combining diffusion with tree-based search, achieving state-of-the-art performance on complex planning problems. Despite its strengths, our analysis shows that MCTD incurs substantial computational overhead due to the sequential nature of tree search and the cost of iterative denoising. To address this, we propose Fast-MCTD, a more efficient variant that preserves the strengths of MCTD while significantly improving its speed and scalability. Fast-MCTD integrates two techniques: Parallel MCTD, which enables parallel rollouts via delayed tree updates and redundancy-aware selection; and Sparse MCTD, which reduces rollout length through trajectory coarsening. Experiments show that Fast-MCTD achieves up to 100x speedup over standard MCTD while maintaining or improving planning performance. Remarkably, it even outperforms Diffuser in inference speed on some tasks, despite Diffuser requiring no search and yielding weaker solutions. These results position Fast-MCTD as a practical and scalable solution for diffusion-based inference-time reasoning.

Figures (7)

• Figure 1: Planning time vs. success rate. As maze size increases, most diffusion-based planners degrade in performance. MCTD maintains high success rates but with long planning times, reflecting the Planning Horizon Dilemma—better performance requires longer planning time. Fast-MCTD breaks this trade-off, achieving strong performance with much faster planning.
• Figure 2: Two key components of Fast-MCTD. (a) Parallel MCTD accelerates planning by performing batched expansion and simulation on a partial denoising tree, followed by delayed tree updates. (b) Sparse MCTD reduces denoising overhead by planning over abstract sub-trajectories, significantly decreasing the number of subplans.
• Figure 3: Planning visualization. Sparse planning (top row) allows for more effective long-horizon trajectories than the dense (bottom row).
• Figure 4: Ablation studies for parallelism degree and interval size. Success rates (%) and planning time (seconds) as the parallelism degree and interval size increase for maze giant tasks.
• Figure 5: Task illustrations: (a) long-horizon maze navigation, (b) multi-object robotic manipulation requiring compositional planning, and (c) visual maze planning from raw RGB observations.