Research map / Probabilistic and scientific modeling / Molecules and proteins
Compositional Monte Carlo Tree Diffusion for Extendable Planning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Compositional Monte Carlo Tree Diffusion (C-MCTD), a training-free inference-time framework that composes whole diffusion-generated plans via tree search (Online, Distributed, and Preplan Composers) to extend planning far beyond training trajectory lengths.
Problem
Diffusion planners and MCTD struggle to generate plans longer than training trajectories, and periodic replanning is myopic because it lacks global context across plan segments.
Approach
Extends MCTD from searching within a single trajectory to searching over compositions of complete plans via stitching-based tree expansion, with guidance sets as meta-actions, fast jumpy-denoising replanning for simulation, parallel multi-origin trees connected at waypoints, and a prebuilt plan graph for fast inference.
Key results
- Preplan Composer achieves 100% success on PointMaze-Giant, versus 68% for the best baseline CompDiffuser.
- Online Composer with Plan Cache reaches 75% on triple-cube and 82% on quadruple-cube manipulation, versus 40% and 24% for MCTD-Replan.
- On visual PointMaze large, Online Composer achieves 54% versus 20% for MCTD-Replan.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compose plans with Preplan Composer rather than the strongest baseline, CompDiffuser | PointMaze giant · PointMaze giant | success rate (%): 68 ± 3 -> 100 ± 0 | Helped |
| compose plans with Preplan Composer rather than CompDiffuser | AntMaze giant · AntMaze giant | success rate (%): 65 ± 3 -> 75 ± 18 | Helped |
| use Online Composer with Plan Cache rather than MCTD-Replan | cube manipulation · Cube manipulation quadruple | success rate (%): 24 ± 8 -> 82 ± 11 | Helped |
| use Online Composer rather than MCTD-Replan | visual maze navigation · Visual PointMaze large | success rate (%): 20 ± 21 -> 54 ± 18 | Helped |
| use a guidance set rather than a fixed guidance level (ablation) | Online Composer · PointMaze medium | success rate (%): 80 ± 0; 90 ± 10; 98 ± 6; 83 ± 8; default 93 ± 9 | Mixed |
| use a guidance set rather than a fixed guidance level (ablation) | Online Composer · PointMaze large | success rate (%): 77 ± 14; 52 ± 13; 27 ± 10; 20 ± 0; default 82 ± 17 | Helped |
| add fast replanning (ablation) | Online Composer · PointMaze medium | success rate (%): 82 ± 6 -> 96 ± 8 (200); 76 ± 12 -> 94 ± 9 (100); 63 ± 15 -> 92 ± 10 (50) | Helped |
| add fast replanning (ablation) | Online Composer · PointMaze large | success rate (%): 54 ± 13 -> 84 ± 15 (200); 49 ± 10 -> 82 ± 14 (100); 32 ± 11 -> 82 ± 19 (50) | Helped |
| vary the stitching threshold (ablation) | Distributed Composer · PointMaze medium | success rate (%): 100 ± 0 at 5.0, 2.0, 1.0, and 0.5; 94 ± 9 at 0.1 | Mixed |
| vary the stitching threshold (ablation) | Distributed Composer · PointMaze large | success rate (%): 65 ± 16 (5.0); 100 ± 0 (2.0, 1.0, 0.5); 73 ± 13 (0.1) | Mixed |
| vary the stitching threshold (ablation) | Distributed Composer · PointMaze giant | success rate (%): 40 ± 24 (5.0); 26 ± 13 (2.0); 26 ± 21 (1.0); 10 ± 10 (0.5); 0 ± 0 (0.1) | Mixed |
Surprising findings
- Distributed and Preplan Composers underperform Online Composer on visual POMDP mazes, attributed to difficulty clustering meaningful waypoints in high-dimensional latent space.
- CompDiffuser degrades sharply on giant mazes (68%) despite achieving 100% on medium and large mazes.
Limitations
- Distributed Composer fails on giant mazes because synthesized plans often exceed the strict 1000-step episode limit (86% when relaxed to 2000).
- Distributed and Preplan Composers underperform in high-dimensional visual latent spaces; real-time inference speed is not achieved.
- Diffusion planners may produce kinematically invalid plans out of distribution and averaged plans under stochastic dynamics.
Experimental setup
Datasets: OGBench PointMaze stitch datasets OGBench AntMaze stitch datasets OGBench cube manipulation play datasets Visual maze datasets from prior work
Benchmarks: OGBench PointMaze (medium, large, giant) OGBench AntMaze (medium, large, giant) OGBench cube manipulation (single to quadruple) Visual PointMaze (medium, large)
Metrics: success rate (%) planning run time (sec) plan length
Baselines: Diffuser-Replan Diffuser-DatasetStitch Diffusion Forcing SSD CompDiffuser MCTD-Replan MCTD-DatasetStitch
Builds on: Monte Carlo Tree Diffusion Diffusion Forcing Diffuser DDIM classifier guidance
Closest papers by meaning
- Monte Carlo Tree Diffusion for System 2 Planning (ICML 2025 · similarity 0.93)
- Fast Monte Carlo Tree Diffusion: 100× Speedup via Parallel and Sparse Planning (NeurIPS 2025 · similarity 0.93)
- Resisting Stochastic Risks in Diffusion Planners with the Trajectory Aggregation Tree (ICML 2024 · similarity 0.89)
- Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning (NeurIPS 2026 · similarity 0.88)
- Tree-Guided Diffusion Planner (NeurIPS 2025 · similarity 0.87)
- Generative Trajectory Stitching through Diffusion Composition (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.