Video World Models · 2026-11-13
Marginal-Preserving Coupled Diffusion Composition for Long-Horizon Planning
Hengrui Zhang, Qiyu Wang, Yuhu Cheng, Brooks Paige, Xuesong Wang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors propose MCDC, a training-free inference modification for pretrained compositional diffusion planners. Instead of sampling chunk initial noises independently, MCDC samples coupled Gaussian initializations with exact per-chunk standard-Gaussian marginals while inducing positive cross-chunk correlation on neighboring chunks, improving overlap-mode compatibility and long-horizon stitching coherence.
Why it matters
Compositional diffusion planners can fail in multimodal overlap regions: locally feasible chunks may denoise into incompatible overlap modes, yielding incoherent stitched trajectories. MCDC addresses this at sampler initialization by coordinating chunk randomness without retraining, guidance optimization, learned verification, or candidate search.
Method
- MCDC replaces independent initial diffusion noise per chunk with a coupled Gaussian initialization built from global, edge, and private standard-Gaussian components.
- Each chunk’s initial noise marginal is preserved as standard Gaussian, while neighboring chunks receive positively correlated noise via shared components on overlap edges.
- MCDC is inserted only at the initial noise sampling step; the denoiser, reverse update, conditioning mechanism, and stitching rule are kept fixed.
Abstract (from arXiv)
Compositional diffusion planners extend short-horizon trajectory models to long-horizon goal-conditioned planning by generating and stitching multiple local trajectory chunks. However, in multimodal environments, local plausibility is not sufficient for global consistency: independently initialized chunks may denoise toward incompatible overlap modes, producing plans that are individually feasible but incoherent after stitching. Existing remedies typically add search, guidance, pruning, or graph-based structure at inference time, increasing computational cost or relying on task-specific machinery. We instead investigate graph-structured dependence among chunk-wise initial diffusion noises as a lightweight coordination mechanism for compositional trajectory planning. We propose \textbf{M}arginal-Preserving \textbf{C}oupled \textbf{D}iffusion \textbf{C}omposition (MCDC), a training-free inference method for pretrained compositional diffusion planners. Instead of sampling chunk noises independently, MCDC constructs coupled Gaussian initializations whose per-chunk marginals remain exactly standard Gaussian, as required by the pretrained denoiser, while neighboring chunks receive positively correlated noise components. This preserves the standard-Gaussian initialization marginal of each chunk while biasing adjacent chunks toward compatible overlap-mode choices. We prove exact marginal preservation and show, under a simple multimodal overlap model, that positive coupling increases the probability of compatible mode selection. On OGBench long-horizon stitch tasks, MCDC consistently improves over independent compositional diffusion sampling while retaining nearly the same inference time, and achieves a favorable success--cost trade-off compared with search-based composition methods.
Related papers
- Akule: Fast and Scalable Multi-Robot Motion Planning via Sparse Interaction Diffusion
- Training-Free Diffusion Planning with Analytical Local Scores
- DPDP: Diffusion Policy-Steered Diffusion Planner for Offline Reinforcement Learning
- Sparse Planning in Visual World Models via Cost Gradients
- How Long, Not How Close: A Learned Temporal Metric for Planning in Latent World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-13 · Indexed by Pub-AI: 2026-11-13 UTC
Marginal-Preserving Coupled Diffusion Composition for Long-Horizon Planning
First announced by arXiv: 2026-11-13 · Original record ↗
- arxiv
- 2611.15585
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-13 · First detected: 2026-11-13 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.15585v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
Compositional diffusion planners extend short-horizon trajectory models to long-horizon goal-conditioned planning by generating and stitching multiple local trajectory chunks. However, in multimodal environments, local plausibility is not sufficient for global consistency: independently initialized chunks may denoise toward incompatible overlap modes, producing plans that are individually feasible but incoherent after stitching. Existing remedies typically add search, guidance, pruning, or graph-based structure at inference time, increasing computational cost or relying on task-specific machinery. We instead investigate graph-structured dependence among chunk-wise initial diffusion noises as a lightweight coordination mechanism for compositional trajectory planning. We propose \textbf{M}arginal-Preserving \textbf{C}oupled \textbf{D}iffusion \textbf{C}omposition (MCDC), a training-free inference method for pretrained compositional diffusion planners. Instead of sampling chunk noises independently, MCDC constructs coupled Gaussian initializations whose per-chunk marginals remain exactly standard Gaussian, as required by the pretrained denoiser, while neighboring chunks receive positively correlated noise components. This preserves the standard-Gaussian initialization marginal of each chunk while biasing adjacent chunks toward compatible overlap-mode choices. We prove exact marginal preservation and show, under a simple multimodal overlap model, that positive coupling increases the probability of compatible mode selection. On OGBench long-horizon stitch tasks, MCDC consistently improves over independent compositional diffusion sampling while retaining nearly the same inference time, and achieves a favorable success--cost trade-off compared with search-based composition methods.
