Towards Scalable One-Step Generative Modeling for Autoregressive Dynamical System Forecasting
Towards Scalable One-Step Generative Modeling for Autoregressive Dynamical System Forecasting
Authors: Tianyue Yang, Xiao Xue Date: 2026-05-07 Paper ID: openalex:2605.05540
Summary
MeLISA is a latent-free autoregressive generative surrogate designed for high-dimensional physical dynamics that avoids the computational overhead of iterative denoising or auxiliary latent compression. By leveraging a blockwise stochastic transition kernel optimized with Window-Consistency MeanFlow and Time Increment Consistency losses, the model maintains physical statistical structures over long-horizon rollouts. Evaluated on 2D Kolmogorov and turbulent channel flows, MeLISA achieves superior performance in both short-term accuracy and long-term statistical metrics (e.g., kinetic energy spectra) compared to neural operators, while retaining inference efficiency.
Key Contributions
- Introduces MeLISA, a latent-free autoregressive generative surrogate using pixel-space MeanFlow to enable single-evaluation blockwise forecasting.
- Develops Window-Consistency MeanFlow and Time Increment Consistency objectives to preserve turbulent flow statistics and prevent long-horizon drift.
- Outperforms neural-operator baselines on Kolmogorov and channel-flow benchmarks in both short-term accuracy and long-term energy spectra preservation while maintaining faster inference speeds.
Open Questions & Future Work
Key Concepts
- melisa-framework: A latent-free autoregressive generative surrogate that uses MeanFlow and consistency-based objectives to achieve efficient, physically accurate long-horizon trajectory forecasting.
Archivist Review
I approved the MeLISA framework because it provides a distinct, latent-free approach to high-dimensional autoregressive generative forecasting, addressing the efficiency-vs-accuracy trade-off in physical surrogates. The open question regarding 3D scaling was approved as it represents a significant, well-defined barrier to the deployment of one-step generation techniques in complex physical systems. No datasets were approved as the benchmarks used were standard flow dynamics cases rather than novel, reusable datasets.
Approved Concepts
- MeanFlow Long-term Invariant Spatiotemporal Consistency Autoregressive Models (MeLISA): Introduces a novel latent-free autoregressive generative framework that enables single-pass high-resolution temporal forecasting while maintaining long-term physical consistency.
Approved Open Questions
- Scaling one-step generative models: Scalability to 3D systems is the primary bottleneck for widespread adoption of generative surrogates in real-world fluid dynamics and engineering applications.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.05540
- paper_id
- 2605.05540
- paper_source
- openalex
- domain
- time-series
- tags
- forecastinggenerative-modelingspatio-temporal-forecastingphysical-dynamics
- architectures
-
- datasets
-
- concept_slugs
- melisa-framework
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-10T06:04:26Z