FRAC Architecture Uses Fractional Dynamics for Long-Sequence SSMs
September 27, 2026
FRAC replaces exponential decay in State Space Models with power-law long memory derived from fractional dynamics. It uses a log-spaced sum of exponential modes to enable efficient parallel training and autoregressive decoding.
HOW THIS AFFECTS YOU
●
builderThis offers a path to more efficient long-context modeling with bounded-state decoding.
●
researcherYou can leverage power-law memory to mitigate exponential forgetting in long-context models.