Fractional State Space Transition for Long Sequence Modeling

简体中文

L1 finite-mode recurrence and long-tail invariants, not a 1.3B LM result.

Paper information

Field	Details
Paper	Fractional State Space Transition for Long Sequence Modeling
Institution	Huawei Noah's Ark Lab, Montreal Research Center (first author).
First public release	2026-09-28 (arXiv v1)
Author code	anasiri/frac-ssm.
Local adapter or method key	frac-ssm
Local reproduction code	src/auto_research/foundation_latest_20260930_closure.py

Paper method and evidence

Approximate fractional power-law memory with finitely many logarithmically spaced exponential modes, combining long-tailed memory with bounded recurrent state and parallel training.

flowchart LR
  I[Sequence input] --> E[Log-spaced exponential modes]
  E --> R[Bounded recurrence]
  R --> M[Approximate power-law memory]

Original paper figure

Figure 1 shows the state-space mechanism. From the original paper; copyright remains with the authors.

Core equation

The reference preserves recurrent transitions; tests check gradient isolation, type boundaries, and determinism.

Reported results

Paper benchmarks, timing and learning curves are not directly comparable to fixtures.

Local reproduction

Three-seed mechanism results are diagnostic_only=true. Fixed-budget fixtures verify the defining operator and auditable intermediate states; they are not a benchmark, production result, or capability-ranking entry.