Glonce

TimeOmni-o1 couples a time series model with an LLM for reasoning and forecasting

researchThe Glonce editorsarXiv🔥58Glonce rating58Impact45Novelty60Relevance95Surprise30Players10Credibility70

The paper starts from a common situation: time series from real-world systems often come with text, such as a report on an event that affects them. Two usual tasks on such data are understanding (for example, answering questions about a series) and forecasting its future values. Both need reasoning over the series and the text together. They also differ in output: understanding produces text, forecasting produces a time series. Serving both therefore calls for a model that can read and generate both modalities.

The authors say existing approaches each fall short. Large language models read time series as digit tokens and handle long series poorly. Time series foundation models (TSFMs) forecast well thanks to large-scale pretraining, but typically neither read nor generate text. Time series language models add a time series modality to the LLM, but they forecast on the LLM side, which leaves the TSFM's pretrained forecasting ability unused.

TimeOmni-o1 is their answer: it couples a pretrained TSFM and an LLM bidirectionally. In one direction, the TSFM encodes the series in compressed form for the LLM, which reasons over it together with the text. In the other direction, the LLM's reasoning is fed back into the TSFM, which produces the forecast conditioned on it. The authors summarise the result as one model architecture that reads and generates both series and text, reasons in text, and draws on the TSFM's pretrained forecasting ability.

To train the model they built MMTR, a multimodal time series reasoning dataset with annotated reasoning traces. It covers both tasks, understanding and forecasting, across 12 domains and includes 500k+ samples.

On MMTR, the authors report that TimeOmni-o1 outperforms existing models on both understanding and forecasting across most domains. It also outperforms the LLM and the TSFM inside it when each is used individually. The gains carry over to three open benchmarks, where it outperforms the dedicated time series models. The text gives these results only as qualitative claims of outperforming.

Key facts

  • TimeOmni-o1 couples a pretrained time series foundation model (TSFM) and an LLM bidirectionally, so one architecture reads and generates both series and text.
  • Direction one: the TSFM encodes the series in compressed form and the LLM reasons over it with the text. Direction two: the LLM's reasoning is fed back into the TSFM, which produces the forecast conditioned on it.
  • The authors built MMTR, a multimodal time series reasoning dataset with annotated reasoning traces, covering understanding and forecasting across 12 domains with 500k+ samples.
  • On MMTR the model outperforms existing models across most domains, plus the LLM and the TSFM inside it used individually; the gains carry over to three open benchmarks, where it beats dedicated time series models.

Why it matters

Time series often arrive with text, and the useful questions need both: what a series means given a report, and what it will do next given that report. The authors argue current model families each cover only part of this. LLMs read series as digit tokens and handle long ones poorly; TSFMs forecast well but typically neither read nor generate text; time series language models forecast on the LLM side and so leave the TSFM's pretrained forecasting ability unused. TimeOmni-o1 is a proposal to keep the strengths of both by letting the LLM reason in text and the TSFM produce the forecast.

Who it affects

The paper is aimed at work on time series that come with accompanying text, such as a report on an event that affects a series. That means researchers building time series understanding and forecasting models, and anyone who wants one system to both answer questions about a series and forecast it. The text does not name specific industries or users.

How to use it

The text gives no practical usage path. No code, weights or dataset release is mentioned. What it does describe is the design pattern: encode the series with a pretrained TSFM, let an LLM reason over that encoding and the text, then pass the reasoning back to the TSFM to condition the forecast. The MMTR dataset, with 12 domains and 500k+ samples, is described as the training set for this approach.

How solid is it

This is a preprint abstract, and every result is the authors' own claim. The text reports that TimeOmni-o1 outperforms existing models on MMTR across most domains, beats the LLM and the TSFM inside it when used individually, and outperforms dedicated time series models on three open benchmarks. No numerical results (accuracy, error, margin of improvement) are given; only qualitative 'outperforms'. The three open benchmarks are not named, nor are the 12 MMTR domains or the existing models compared against.

Risks and caveats

The headline result holds across most MMTR domains, not all, and the domains where it does not outperform existing models are not identified. The comparison dataset, MMTR, was built by the same authors to train the model. The base LLM and the base TSFM used inside TimeOmni-o1 are not named, and no model size, parameter count or training compute is given. It is not stated whether the reasoning traces are human-written or model-generated; they are only called 'annotated'. Without figures, the size of any gain cannot be judged.

“One model architecture thus reads and generates both series and text, reasons in text, and draws on the TSFM's pretrained forecasting ability.”

— TimeOmni-o1 paper abstract