PaperlayerView exact revision on arXiv

Source-linked AI summary

SciTS: Scientific Time Series Understanding and Generation with LLMs

Wen Wu, Ziyang Zhang, Liwei Liu, Xuenan Xu, Jimin Zhuang, Ke Fan, Qitan Lv, Junlin Liu, Chen Zhang, Zheqi Yuan, Siyuan Hou, Tianyi Lin, Kai Chen, Bowen Zhou, Chao Zhang

arXiv:2510.03255v2cs.LGcs.AI

TL;DR

Scientific time series are under-explored in LLM research, and existing representations and unified models leave gaps in handling their variety. This paper introduces the SciTS benchmark and TimeOmni, finding stronger domain generalisation from general-purpose LLMs than specialised time-series models, with TimeOmni ranking highest. The authors conclude that dedicated temporal processing is important for handling complex scientific dynamics.

  • Problem

    Scientific time series remain under-explored for LLMs, and existing unified time-series models typically focus on forecasting or analysis while their effectiveness on heterogeneous scientific signals remains unclear.

    Source: §1 · pp. 1–2

  • Method

    The paper introduces SciTS, a multi-domain benchmark, evaluates 17 models, and presents TimeOmni, which combines time-series encoding with an LLM for understanding and generation.

    Source: §1 · p. 2

  • Results

    TimeOmni attains the top rank on SciTS, and general-purpose LLMs generalise better to scientific domains than specialised time-series models.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors conclude that pairing general-purpose LLMs with dedicated temporal processing is crucial for handling complex scientific dynamics.

    Source: §7 · p. 10

    Due to resource constraints, all baselines were evaluated zero-shot without finetuning.

    Source: LIMITATIONS AND ETHICS STATEMENT · p. 11

Abstract

from arXiv · show

The scientific reasoning ability of large language models (LLMs) has recently attracted significant attention. Time series, as a fundamental modality in scientific data, presents unique challenges that are often overlooked in current multimodal LLMs, which either encode numerical sequences as text or convert them into images. Such approaches may be insufficient for comprehensive scientific time series understanding and generation. Existing unified time series models typically specialise in either forecasting or analysis, and their effectiveness on non-periodic, heterogeneous scientific signals remains unclear. To address these gaps, we introduce SciTS, a benchmark spanning 12 scientific domains and 43 tasks, with over 50k+ instances, both univariate and multivariate signals ranging from $10^0$ to $10^7$ in length and up to 10~MHz in frequency. We benchmark 17 models, including text-only LLMs, multimodal LLMs, and unified time series models, and find that general-purpose LLMs exhibit stronger generalisability than specialised time series models, while representing time series as text or images limits their performance due to excessively long sequences and loss of numerical precision, respectively. We then introduce TimeOmni, a framework that equips LLMs with the ability to understand and generate time series while remaining compatible with general-purpose LLM training. This work fills a gap in both dedicated benchmarks and modelling frameworks for scientific time series, paving the way for LLMs to understand and generate complex temporal scientific data.

1 INTRODUCTION

Scientific time series remain under-explored for LLMs, while existing representations and specialised models leave important challenges. SciTS benchmarks these challenges, and TimeOmni explores unified understanding and generation with explicit temporal modelling.

1 INTRODUCTION

  • Text representations create sequences that are too long for LLMs to process, while image representations sacrifice numerical precision; specialised models also tend to target narrow tasks or domains.The effectiveness of existing unified models on non-periodic, heterogeneous scientific signals remains unclear.Source: §1 · pp. 1–2
  • SciTS contains 54,023 instances across 43 tasks and 12 disciplines, covering univariate and multivariate signals with lengths from 100 to 107 and frequencies up to 10 MHz.Its seven task types include anomaly detection, classification, question answering, event localisation, forecasting, imputation, and synthesis.Source: §1 · p. 2
  • TimeOmni attains the top rank on SciTS, while general-purpose LLMs generalise better than specialised time-series models and text or image representations constrain performance.The authors connect TimeOmni’s result to explicitly modelling temporal dynamics within an LLM.Source: §1 · p. 2
  • TimeOmni takes a time-series signal and task prompt and produces text or time-series outputs, using routed patch experts to handle varied lengths, resolutions, and dimensions.The framework is intended for incorporation into general-purpose LLMs and joint training with other modalities and tasks.Source: §1 · p. 2

2 RELATED WORK

Prior benchmarks and modelling approaches leave scientific time series and unified understanding-plus-generation relatively under-studied. SciTS and TimeOmni address these gaps with a broad benchmark and an LLM-compatible framework.

2 RELATED WORK

  • SciTS targets scientific time-series understanding and generation, areas not extensively covered by existing scientific benchmarks focused on reasoning over text and images.The paper points to a comparison with existing benchmarks in Table 1.Source: §2 · p. 2, Table 1 · p. 3
  • Unified time-series approaches face diverse scales, sampling rates, and dynamic patterns, and most prior work focuses on understanding or generation rather than both.UniTS integrates QA and forecasting but uses a separate architecture that is not compatible with general-purpose LLM training; scientific heterogeneous, non-periodic signals remain unexplored.Source: §2 · pp. 2–3

3 SCITS DATASET AND TASKS

SciTS brings together scientific time series from varied domains and signal scales, organizes them into prompt-based tasks, and covers both understanding and generation.

3 SCITS DATASET AND TASKS

  • SciTS contains 54,023 samples across 43 tasks and 12 disciplines, with frequencies from daily observations to megahertz, lengths from fewer than 10 points to millions, and up to 58 EEG channels.Dataset statistics are presented in Table 2.Source: §3 · p. 3, Table 2 · p. 4

3.1 DATA COLLECTION

  • The dataset draws on open-source datasets, scientific websites, and numerical simulations; task inputs and targets derive from accompanying metadata or the time series, with systematic quality checks.All tasks use a prompt-based format intended to support LLM-style training and evaluation.Source: §3.1 · p. 4

3.2 DOMAINS

  • The 12 disciplines include tasks such as gravitational-wave event detection, earthquake arrival-time prediction, animal-call classification, EEG analysis, and multivariate chaotic-system forecasting.The task examples span astronomy, earth science, bioacoustics, neuroscience, and mathematics, among other fields.Source: §3.2 · p. 4

3.3 TYPE OF TASKS

  • SciTS includes four understanding task types—anomaly detection, classification, and multiple-choice question answering among them—and four generation task types, including event localisation, forecasting, imputation, and synthesis.Generation may use auxiliary context, and output sequences range from a few steps to forecasts of up to 720 steps.Source: §3.3 · pp. 4–5

4 TIMEOMNI: A WORKING EXAMPLE

TimeOmni encodes time series for an LLM, combines them with task prompts, and uses task-specific outputs for understanding or generation.

4 TIMEOMNI: A WORKING EXAMPLE

  • TimeOmni encodes the input signal into LLM-dimensional embeddings, concatenates them with prompt embeddings, and uses a pretrained LLM with text or time-series output heads.The encoder selects a patch expert based on signal length and typically produces 100 to 200 embeddings.Source: §4 · p. 5

4.1 TIME SERIES ENCODER

  • The router selects a patch size so the patch-expert output sequence falls between 100 and 200 tokens for a flattened input of length T.The patch expert reshapes the signal into patches and maps them through a one-dimensional convolution before reprogramming.Source: §4.1 · pp. 5–6

4.2 LLM INPUT AND OUTPUT

  • Understanding tasks place the prompt after the encoded series and produce text through a softmax layer; generation tasks place the prompt first and map LLM outputs to the required series length.For generation, the model selects a predefined regression head with the closest output length and truncates if necessary.Source: §4.2 · p. 6

5 EXPERIMENTAL SETUP

The evaluation compares text-only LLMs, multimodal LLMs, and unified time-series models in zero-shot settings, using task-appropriate metrics and accounting for generation failures.

General Settings

  • All models are assessed zero-shot; TimeOmni is initialized from Qwen3-8B and fine-tuned using DoRA.Text-only LLMs receive digit sequences as text, multimodal LLMs receive time-series images, and unified time-series models are typically specialized for applications such as forecasting.Source: §5 · p. 6

Metrics

  • Understanding tasks use accuracy and F1, while generation tasks report MAE, MAPE, success rate, and swMAPE.MAE and MAPE are computed on the original data scale rather than normalized inputs.Source: §5 · p. 6
  • swMAPE divides MAPE by generation success rate, penalizing models that fail to process samples.Failures include inputs exceeding a model’s context window or failures to follow instructions.Source: §5 · p. 6

6 RESULTS AND DISCUSSION

Across SciTS’s 43 tasks and 12 disciplines, results show broad differences in coverage and performance by model and task type. TimeOmni ranks strongly overall, while the design analyses point to explicit and adaptive temporal modelling as important for handling diverse signals.

6.1 TASK COVERAGE AND SUCCESS RATE

  • TimeOmni handles all benchmark tasks and instances, achieving full coverage and full success, while specialized models support fewer tasks but succeed often on those they support.Open-source text-only LLMs fail on approximately 10% of tasks, and LLM-based methods have relatively low instance-level success rates.Source: §6.1 · p. 7, Figure 4 · p. 7

6.2 UNDERSTANDING TASKS

  • Understanding performance is weak in bioacoustics and radar, with F1 below 10%, while nearly all models perform strongly on economics.Across most domains, TimeOmni has the best performance and Gemini2.5-Flash ranks second; image-input multimodal LLMs outperform text-only LLMs overall.Source: §6.2 · p. 7, Table 3 · p. 8

6.3 GENERATION TASKS

  • For generation, text inputs generally outperform images when precise numerical values matter, but LLMs commonly have success rates below 10% on long-output ENG02 forecasting.TimeOmni has the highest overall generation ranking, closely followed by Gemini2.5-Flash using text; Moirai has the lowest average swMAPE in economics and mathematics among supported tasks.Source: §6.3 · pp. 8–9, Table 4 · p. 8

6.4 MODEL RANKINGS ACROSS DISCIPLINES

  • Across disciplines, the results support treating time series as a distinct modality and combining general-purpose LLMs with dedicated temporal processing.The paper reports stronger generalization by general-purpose LLMs to unseen domains than by specialized time-series models; Figure 5 reports mean task rank within each discipline.Source: §6.4 · p. 9, Figure 5 · p. 9

6.5 DESIGN ANALYSIS: WHAT IS NEEDED FOR SCIENTIFIC TIME SERIES MODELLING?

  • Replacing TimeOmni’s reprogramming module with an MLP degrades performance, while fixed patch sizes struggle across extreme temporal scales.Small patches pose memory and long-range-dependency challenges for very long sequences; large patches can collapse short sequences into one patch. Fine-tuning Qwen2.5VL and TimeMoE does not remove their respective numerical-precision and task-coverage limitations.Source: §6.5 · pp. 9–10

7 CONCLUSIONS

The paper introduces SciTS as a broad benchmark and uses it to evaluate models for scientific time-series understanding and generation. Its findings point to dedicated temporal processing alongside LLMs as a promising direction, with TimeOmni as a preliminary framework.

Conclusions

  • SciTS spans 12 disciplines, 7 task types, 43 tasks, and over 50k instances, and the paper benchmarks 17 text-only, multimodal, and unified time-series models.The benchmark includes signals with varied frequencies, lengths, and multivariate dimensions.Source: §7 · p. 10
  • The authors conclude that general-purpose LLMs generalize better to unseen domains than specialized time-series models, while dedicated temporal processing is important for complex scientific dynamics.Converting signals to text or images incurs information loss and hinders efficiency and numerical fidelity.Source: §7 · p. 10
  • TimeOmni demonstrates an approach that integrates time series into LLMs while remaining compatible with general-purpose LLM training.The paper presents this as a preliminary step toward models that can interrogate dynamical systems and support discovery and experimental design on real-world scientific temporal data.Source: §7 · p. 10

LIMITATIONS AND ETHICS STATEMENT

The authors describe evaluation and scope boundaries alongside background on time-series models and benchmarks. The review contrasts task-specific methods and existing benchmarks with the paper’s broad scientific time-series focus.

LIMITATIONS AND ETHICS STATEMENT

  • Resource constraints limit baseline evaluation to zero-shot settings without fine-tuning; the authors also disable closed-source LLM thinking mode because it adds substantial computational cost.Their preliminary small-scale experiments found no performance improvement from enabling that mode.Source: LIMITATIONS AND ETHICS STATEMENT · p. 11

B.1 TIME SERIES MODELS

  • The related-work discussion describes traditional time-series models as task- or dataset-specific and reviews unified models and LLM-based approaches for temporal data.It notes prior concerns about LLM forecasting and numerical reasoning, alongside multimodal and alignment-based proposals for connecting time series with LLMs.Source: §B.1 · pp. 16–17

B.2 SCIENCE BENCHMARKS AND TIME SERIES BENCHMARKS

  • The benchmark review contrasts synthetic datasets’ controllability with real-world datasets’ authenticity and describes SciTS as balancing these properties across scientific domains and tasks.It also surveys science-data and time-series benchmarks with differing formats, domains, and question types.Source: §B.2 · p. 17

C IMPLEMENTATION DETAILS OF TIMEOMNI

TimeOmni is trained jointly on understanding and generation tasks, with channel-dependent preprocessing for inputs and outputs. The appendix also reports discipline-level SciTS statistics.

C IMPLEMENTATION DETAILS OF TIMEOMNI

  • TimeOmni is trained for 10 epochs with understanding and generation tasks optimised concurrently, using DoRA, Adam, and a cosine-decay schedule after warmup.The first 5% of training steps use linear warmup; batches contain 6 understanding examples or 1 generation example.Source: C IMPLEMENTATION DETAILS OF TIMEOMNI · p. 17
  • For understanding, multichannel inputs are flattened; generation flattens inputs for single-channel outputs but predicts per channel when input and output channel counts match.The flattening maps an RT′×N input to RNT′×1.Source: C IMPLEMENTATION DETAILS OF TIMEOMNI · p. 18

D MORE DETAILS OF SCITS

  • SciTS discipline-level statistics are presented in Table 5, covering instance counts, signal-length ranges, frequencies, and channel counts.The supplied passage identifies the table’s contents but gives no discipline-level values.Source: D MORE DETAILS OF SCITS · p. 18, Table 5 · p. 18

D.1 TASKS DESCRIPTION

This section catalogs SciTS tasks and supporting datasets by discipline, describing signal types, task formats, construction details, dataset configurations, and assessed model families, their task coverage, and evaluation settings. Evaluations use minimal reasoning effort, based on preliminary experiments suggesting little benefit on SciTS tasks.

D.1 TASKS DESCRIPTION

  • SciTS task descriptions span scientific fields and cover classification, anomaly detection, event localization, forecasting, imputation, and signal generation.Examples include gravitational-wave and earthquake analysis, animal-vocalization classification, temperature forecasting, EEG tasks, energy-signal generation, and chaotic-system forecasting.Source: §D.1 · pp. 18, 19, 21, 23, Table 6 · p. 19

D.2 DATASET DESCRIPTION

  • The GWOSC GW Event dataset uses simulated gravitational-wave templates injected into real detector-noise segments to address the limited number of observed events (Dax et al., 2021; 2023; 2025).Noise segments are 4 seconds long at 16 kHz; the construction also applies data-quality and injection masks, matched-filter SNR scaling, windowing, and whitening.Source: §D.2 · pp. 23–24
  • The described astronomy, seismic, and bioacoustic datasets provide labeled examples for light-curve, earthquake, and vocalization tasks.LEAVES combines survey light curves; STEAD supplies labeled three-channel seismic waveforms; Powdermill, iNatSounds, and MarmAudio provide bird, multispecies, and marmoset recordings, respectively.Source: §D.2 · p. 24
  • Other datasets support time-series question answering and weather- and finance-related tasks using contextual text or aligned modalities.Time MQA contains about 200,000 question-answer pairs across five task types; TimeCAP uses 24 hours of weather data to predict rain in the next 24 hours; MTBench pairs series with text, while FinMultiTime aligns four modalities across 5,105 stocks.Source: §D.2 · pp. 24–25
  • The neuroscience datasets cover depression-related EEG, EEG event classes, and motor-imagery brain-computer-interface recordings.MDD includes resting-state EEG from patients and controls; TUEV labels EEG segments with six classes; WBCIC SHU includes 62 subjects across three sessions.Source: §D.2 · p. 25
  • The assessed time-series datasets span sleep staging, text-conditioned forecasting, radar segmentation, and chaotic-system forecasting, with task-specific signal processing and context settings.SleepEDF uses filtered Fpz-Cz EEG and five AASM stages; textETT pairs captions with ETTh1 fragments of lengths 24, 48, and 96 (Ge et al., 2025); RadSeg provides sample-level annotations across -20 to 20 dB SNR (Huang et al., 2024); and the coarse Chaotic subset retains 116 systems for 100-step contexts and 20-step forecasts (Gilpin, 2023).Source: §D.2 · pp. 25–26
  • Other dataset configurations cover oil-temperature forecasting, news-aligned multi-domain forecasting, and ECG, traffic, bearing, machine-sound, and wireless-signal tasks.ETTh1 uses oil temperature with a 96-step context and 96- or 720-step predictions (Wu et al., 2021); NewsForecast combines four domains with related news (Wang et al., 2024b); PTB XL includes 21,837 12-lead ECGs; MetroTraffic yields about 25k examples across two forecast modes; CWRU covers bearing faults; MIMII DUE covers normal and abnormal machine sounds; and RadarComm provides radar and communication waveforms (Jagannath & Jagannath, 2021).Source: §D.2 · pp. 25–26

E DESCRIPTION AND CONFIGURATION OF THE ASSESSED MODELS

  • Text-only, multimodal, and time-series models are evaluated with minimal reasoning effort, following preliminary experiments that found little benefit from reasoning on SciTS tasks.The setting also reflects resource constraints.Source: E DESCRIPTION AND CONFIGURATION OF THE… · p. 26

E.1 TEXT LLMS

  • The text-only evaluation includes six LLMs, from GPT-4.1-mini and Gemini-2.5-Flash to DeepSeek-V3, Llama3-8B, and Qwen3-4B/8B.Their stated context lengths range from 8k tokens for Llama3-8B to 1M tokens for GPT-4.1-mini and Gemini-2.5-Flash.Source: §E.1 · pp. 26–27

E.2 MULTIMODAL LLMS

  • The multimodal evaluation uses GPT-5-mini, Gemini-2.5-Flash, InternVL3.5-4B/8B, and Qwen2.5-VL-7B.The listed context lengths include 400K tokens for GPT-5-mini, 32K for InternVL3.5-8B, and 128K for Qwen2.5-VL-7B.Source: §E.2 · p. 27

E.3 TIME SERIES MODELS

  • Time-series baselines are restricted to zero-shot-capable models: Moirai, Time-MoE, and Chronos forecast, ChatTS understands but only outputs text, and UniTS supports both understanding and generation.Moirai, Time-MoE, and Chronos are forecasting-only; ChatTS does not generate time series, while UniTS handles forecasting, anomaly detection, classification, and imputation, but requires matching input and output dimensions (Gao et al., 2024).Source: §E.3 · pp. 27–28
  • The forecasting baselines vary in scope: Moirai supports multivariate-to-univariate forecasting, while Time-MoE and Chronos are limited to univariate forecasting.The evaluated Moirai sizes are 14M, 91M, and 311M parameters; Time-MoE sizes are 50M and 200M; Chronos-bolt sizes are 21M and 205M.Source: §E.3 · pp. 27–28

F FULL RESULTS ON SCITS

The full-results section defines task failure codes and evaluation metrics, then provides detailed SciTS results across multiple scientific disciplines.

F FULL RESULTS ON SCITS

  • Understanding tasks use accuracy and F1, while generation tasks use MAE, MAPE, and success rate; failure codes cover sequence limits, channel support, and instruction following.Unsupported tasks are marked “-”, and results are presented as percentages.Source: F FULL RESULTS ON SCITS · p. 28
  • Detailed results are tabulated for Astronomy and Earth Science, Bioacoustics and Meteorology, and Meteorology and Economics.The supplied table captions identify these scopes but do not provide cell values.Source: Table 8 · p. 28, Table 9 · p. 29, Table 10 · p. 29
  • Additional detailed-results tables cover Economics and Neuroscience, Neuroscience and Energy, Energy and Physiology, and Physiology.These captions establish table scope, not model rankings or metric values.Source: Table 11 · p. 30, Table 12 · p. 30, Table 13 · p. 31
  • Further detailed-results tables cover Physiology and Urbanism, Urbanism and Manufacturing, and Radar and Math.The supplied captions do not state comparative outcomes.Source: Table 15 · p. 32, Table 16 · p. 32, Table 17 · p. 33

G ABLATION STUDY ON TIMEOMNI

Ablations indicate that both the reprogramming module and adaptive patch-expert routing contribute to TimeOmni’s performance. Routing is particularly important across the benchmark’s range of signal lengths.

Ablation on the reprogramming module

  • Replacing reprogramming with a two-layer MLP degrades performance across most tasks.The authors attribute the difference to stronger alignment between time-series inputs and the text LLM, improving robustness and generalization across diverse tasks.Source: G ABLATION STUDY ON TIMEOMNI · p. 34, Table 18 · p. 34

Ablation on the patch-expert routing module

  • A fixed patch size of 1024 has little effect on medium-length tasks and can slightly improve them, but hurts very long and very short sequences.Short sequences collapse into a single patch, while very long sequences lose long-range dependencies; smaller fixed patches were infeasible because of GPU memory limits.Source: G ABLATION STUDY ON TIMEOMNI · p. 34, Table 19 · p. 35
  • Patch-expert routing is essential for handling scientific time series spanning very short to extremely long temporal scales.The fixed-size ablation shows why a single patch size is not suitable across these scales.Source: G ABLATION STUDY ON TIMEOMNI · p. 34, Table 19 · p. 35

H DISCUSSION OF THE EVALUATION SETUP

The discussion frames benchmark failures as architectural limits that fine-tuning alone does not resolve, while noting TimeOmni’s proof-of-concept scope. It also presents examples of SciTS understanding and generation tasks across scientific domains.

H DISCUSSION OF THE EVALUATION SETUP

  • TimeOmni is a proof-of-concept, not an all-purpose model; it handles imputation and QA that forecasting specialists cannot, but may not outperform them on every forecasting task.Its jointly trained data cannot cover all SciTS domains and tasks; ENG04–05 and URG02–04 are evaluated out of domain.Source: H DISCUSSION OF THE EVALUATION SETUP · p. 36
  • For fair comparison, the authors fine-tuned an open-source image–text LLM and a time-series foundation model using TimeOmni’s training data; fine-tuning a text-only LLM was impractical because converting signals to text produced extremely long token sequences.Source: H DISCUSSION OF THE EVALUATION SETUP · p. 36
  • Fine-tuning Qwen2.5VL-7B yields limited gains on some tasks, and more than half of ENG03 samples still fail to produce the required output length.The discussion notes instruction-following improvement on ENG03, alongside instability and difficulty maintaining balanced performance across tasks.Source: H DISCUSSION OF THE EVALUATION SETUP · p. 36
  • Fine-tuning TimeMoE improves some forecasting tasks but does not enable imputation or QA, which it could not previously address.The comparison uses a time-series model specialized for forecasting.Source: H DISCUSSION OF THE EVALUATION SETUP · p. 36
  • Existing models’ inability to support SciTS’s diverse signals and task types is described as structural, not something fine-tuning alone resolves.The authors cite success rates and failure analyses in Fig. 4b and results from the fine-tuning comparisons.Source: H DISCUSSION OF THE EVALUATION SETUP · p. 38

I CASE EXAMPLE

Related papers

Loading 2510.03255v2…