TimesJev compresses time series for instant LLM-style reasoning
The arXiv paper proposes a System 1 model that compresses numeric series into a compact representation, cutting inference time while the authors report higher accuracy than long chain-of-thought reasoning models.
In plain words
Reasoning over time series with large language models usually means generating long chains of thought, which is slow and costly. TimesJev instead squeezes a numeric series into a short representation that keeps its key frequency patterns, so a fast, single-step model can reason over it. The authors report better results and much lower latency than slower reasoning models.
What was announced
The paper argues that chain-of-thought reasoning over time series can consume thousands of tokens, creating substantial inference latency, while fast System 1 models decide instantly but cannot directly ingest raw numeric series because the data does not fit their embedding space. TimesJev addresses this by formalizing what the authors call System 1 sufficiency and learning a compressed representation that preserves key frequency patterns. The compression objective has two parts: a sufficiency divergence measuring core information lost during compression, and a realizability divergence quantifying how incompatible the representation is with the model's optimal prediction. A gating mechanism selects the compressed features, which are then aligned with the model's embedding space.
On 13 datasets, the authors report that TimesJev outperforms reasoning models by over 62% while using 62.9% less inference time than their fine-tuned versions on average. These figures come from the paper's own experiments and have not been independently verified. The work is described by its authors as a very early attempt at a time series System 1 model, and the abstract does not detail which reasoning models were compared, how fine-tuning was performed, or which tasks the 13 datasets cover.
What it does not show
- Evidence is limited to the paper's abstract; experimental setup, baselines, and task types are not described, and the reported accuracy and latency gains are the authors' own claims without independent confirmation.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.