Non-Markovian History Compression Bottleneck
Background: Continuous-time diffusion models frequently employ standard transformer architectures that attend to the entire history, which limits scalability for high-frequency or long-duration temporal processes.
Question / Future Work: Developing efficient architectures, such as state-space models, to selectively compress path history in non-Markovian generative models remains an unresolved challenge for maintaining computational feasibility.
Why It Matters: Addressing the computational overhead of path-dependent history is critical for the application of non-Markovian generative models to long-range forecasting.
Evidence: For a mathematically faithful implementation, we made the transformer network attend to the entire history, but there are efficient architectures (e.g., SSMs) that selectively compress the history…
Metadata & Links
- created_at
- 2026-05-03T06:03:22Z