PIMSM: Physics-Informed Multi-Scale Mamba for Stable Neural Representations
A new state-space architecture, Physics-Informed Multi-Scale Mamba (PIMSM), addresses distribution shift in scientific time series by preserving multiscale physical structure. The method formalizes temporal kernel mismatch, where models fit in-distribution dynamics without anchoring to physical timescales, causing representation drift. PIMSM maps spectrum-estimated transition points between frequency regimes (knee frequencies) to stabilize neural representations. The paper is published on arXiv (2605.16351) as a cross-type announcement.
Key facts
- PIMSM stands for Physics-Informed Multi-Scale Mamba.
- It targets distribution shift in scientific time series.
- The architecture preserves multiscale physical structure.
- Temporal kernel mismatch is identified as a failure mode.
- Knee frequencies are used to map transition points.
- The paper is on arXiv with ID 2605.16351.
- It is a cross-type announcement.
- The approach aims to improve transfer and robustness.
Entities
Institutions
- arXiv