Towards Interpretable Time Series Foundation Models

Jeremy Pawlus; Matthieu Boileau; Philippe Helluy; Svitlana Vyetrenko

REVIEW 3 major objections 4 minor 22 references

Towards Interpretable Time Series Foundation Models

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Fine-tuning compact language models on 180 synthetic Ornstein-Uhlenbeck series transfers basic time series reasoning, with post-training feature scores of 1.0 for trend, 0.95 for extrema, and 0.8 for noise.

desk verdict The distillation pipeline is plausibly useful, but the feature-based metric is partly circular and without a rule-based baseline the 'interpretive capability' claim is not yet established. read the letter →

arxiv 2507.07439 v1 pith:5A7THPJQ submitted 2025-07-10 cs.CL cs.AI

Matthieu Boileau , Philippe Helluy , Jeremy Pawlus , Svitlana Vyetrenko This is my paper

classification cs.CLcs.AI

keywords timeseriesreasoningknowledgedistillationsmalllanguagemodelsinterpretabilityOrnstein-Uhlenbeckprocessnaturalannotationfeature-basedevaluationNLIscoring

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that distillation can compress rudimentary time series understanding into language models small enough for on-device use. The authors generate 200 mean-reverting Ornstein-Uhlenbeck series with controlled trends and noise, have a large multimodal model write three structured sentences per series about trend, noise intensity, and extrema locations, and fact-check those captions against rule-based feature extraction before using them to fine-tune Qwen2.5 models with 0.5B and 1.5B parameters. On a held-out set of 20 series from the same generator, the post-trained students reach cosine similarity 0.98 with the teacher and feature-based scores of 1.0 for trend, 0.95 for extrema, and 0.8 for noise. If the result holds, compact models could produce natural-language explanations of temporal patterns in privacy-sensitive or resource-constrained settings.

What carries the argument

The load-bearing mechanism is a fact-checked distillation loop built on the Ornstein-Uhlenbeck generator $r_t = r_{t-1} + \kappa(\bar{r} - r_{t-1}) + u_t$ with $u_t \sim N(0, \sigma^2)$. The teacher model sees both an image and numeric values and outputs three JSON sentences; a deterministic feature extractor derives reference sentences from a smoothed trend, variance-based noise bins, and exact extrema positions; NLI comparison flags contradictions, and extrema sentences that contradict the facts are replaced. The cleaned captions are paired with the numerical series, whose values are rescaled and left-padded to two integer digits so each token is a single digit, and used to fine-tune compact Qwen2.5 students. This loop is what transfers reasoning from the teacher to the student.

What would settle it

Compute the post-training feature-based scores on a test set generated from a different stochastic process, such as a random walk or a seasonal autoregressive model, while keeping the same annotation prompt and scoring; if the 0.5B model's trend and extrema scores fall far below 1.0 and 0.95, the claimed interpretive capability does not generalize beyond the exact Ornstein-Uhlenbeck distribution.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a 0.5B-parameter Qwen model, after fine-tuning on 180 synthetic time series whose text annotations were produced by a large multimodal teacher and then corrected by a deterministic fact-checker, can describe whether a series trends up, down, or flat; whether its noise is low, medium, or high; and where its global maximum and minimum lie. The evidence is the jump in evaluation scores from before to after post-training: feature-based trend agreement goes from 0.8/0.6 to 1.0/1.0 for the 1.5B/0.5B models, extrema from 0.475/0.45 to 0.95/0.95, and noise from 0.475/0.45 to 0.8/0.8. The authors read these numbers as showing that meaningful interpretive capability can be distilled, not just memorized phrasing, while also reporting that trend and extrema skills transfer more reliably than noise.

Load-bearing premise

The load-bearing premise is that the rule-based reference sentences, built from a smoothed trend, variance-based noise bins, and computed extrema positions, are the correct ground truth for what trend, noise, and extrema mean, because those same references are used both to score the students and to correct the teacher before training.

Editorial extensions

If this is right

Post-trained 0.5B and 1.5B Qwen models can produce the required three-sentence JSON descriptions on held-out series from the same generator, which the paper treats as evidence that the reasoning is usable in practice.
The distillation path avoids sending sensitive time series to a large cloud model at inference time, so the small model can run on-device and still provide structured explanations.
Because fact-checking corrected only a small share of teacher annotations, roughly 5% of extrema cases, the pipeline is inexpensive and reproducible with just 200 synthetic samples.
The large gap between pre-training and post-training scores, for example noise feature agreement rising from 0.475 to 0.8 for the 1.5B model, indicates that the fine-tuning step, not the base model, carries the capability.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A stricter test of the claimed capability would be to hold out series generated by different process families, such as random walks, seasonal, or irregularly sampled series, and see whether trend and extrema scores survive; the paper's test set comes from the same Ornstein-Uhlenbeck generator as training.
The NLI scores after training, which sit between 0.675 and 0.875, are well below the feature-based scores, so a reader should not equate reproducing the heuristic with entailing the intended meaning; a dedicated human-annotation collection would clarify the gap.
One cheap ablation would be to train a student on uncorrected teacher annotations and compare scores; if the numbers do not change, the fact-checking loop is not the active ingredient behind the improvement.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a small, clear workshop paper with a genuinely new combination—distilling time series reasoning from a large multimodal model into 0.5B–1.5B Qwen models—but the headline result is weaker than it looks because the feature-based metric shares its construction with the label-correction step.

What's good: the pipeline is simple and reproducible. Synthetic Ornstein-Uhlenbeck series, a straightforward prompt, and a clean tokenization trick (integer rescaling to single tokens). The before/after tables show large gains, and the paper is honest about the data being synthetic and narrow. The related work seems fair; I found no prior distillation of time-series reasoning into small LMs, so the combination is at least novel.

Where it gets soft. The feature-based scores (1.0 trend, 0.8 noise, 0.95 extrema after training) are computed against reference sentences generated by the same smoothed-trend, variance-bins, and computed-extrema procedure used to fact-check and correct the teacher annotations in Section 2.4. That means the student is being scored on how well it reproduces a heuristic it was trained to match. A rule-based baseline that just emits those fact-based sentences directly would likely get near-perfect feature scores, and the paper doesn't include one. So the feature metric does not independently validate 'interpretive capability.' The NLI scores are against the teacher's annotations, but those annotations were partly corrected by the same heuristic, so there's some contamination there too.

Other concerns are minor or standard: 20 test series from the same generator, no confidence intervals, and the pre/post comparison changes the output format between conditions (JSON vs. separate questions), which could explain part of the gain. These are fixable.

On balance, the narrow claim—small models can be fine-tuned to imitate the teacher on in-distribution synthetic series—probably holds. The stronger claim in the abstract about 'meaningful interpretive capabilities' does not yet hold. That needs a rule-based baseline, a larger held-out test set, and more careful wording.

My recommendation: send it to review with expectation of revision. A serious referee can sort out the metric issue. It's a solid workshop contribution, not a foundation model result.

Referee Report

3 major / 4 minor

Summary. The paper proposes distilling time series interpretive reasoning from the large multimodal model pixtral-large into compact Qwen2.5-Instruct models. It generates 200 synthetic Ornstein-Uhlenbeck series with varied trends and noise, prompts pixtral-large to produce structured JSON annotations, fact-checks and corrects those annotations with algorithmically computed feature-based sentences, and then fine-tunes Qwen2.5-0.5B and Qwen2.5-1.5B on 180 training samples. Evaluation on 20 test samples uses cosine similarity, per-sentence NLI scores, and feature-based scores for trend, noise, and extrema. Tables 1 and 2 report large post-training gains on all metrics. The central claim is that compact models acquire meaningful interpretive capabilities for basic time series features after distillation.

Significance. If established, the result would be a useful step toward interpretable small models that can explain time series in natural language, with clear relevance to on-device and privacy-sensitive deployment. The pipeline is simple and reproducible, and the preprocessing step that rescales and left-pads numeric tokens to preserve digit-level tokenization is thoughtful. The before/after tables make the training effect visible and credit the idea that even very small models can learn structured descriptions of trend, noise, and extrema from a compact synthetic dataset. However, the current evaluation leaves open whether the reported scores reflect genuine feature extraction or reproduction of the heuristic used to construct the reference labels, so the significance depends on the additional controls requested below.

major comments (3)

[Sections 2.3 and 2.4, Table 2] The feature-based reference sentences are built from a smoothed-trend rule, variance-based noise bins, and computed extrema positions (Section 2.3), and the same procedure is used to correct 5% of the teacher's extrema annotations before training (Section 2.4). Consequently, the feature scores in Table 2 measure how well the student reproduces the heuristic used to create and correct the training labels, not an independent interpretive capability. I request a rule-based baseline that emits the fact-based sentence directly from the series; without it, the reader cannot tell whether a learned model is needed to achieve the Feature trend 1.0, noise 0.8, and extrema 0.95 scores. An evaluation on a held-out distribution, such as real-world series from TimeSeriesExam or the Fons et al. benchmark, would also separate heuristic reproduction from generalization.
[Section 2.2, Tables 1 and 2] The test set consists of only 20 series generated from the same OU process as the training set, and no confidence intervals, multiple seeds, or significance tests are reported. Each feature score is a proportion over 20 items (e.g., noise 0.8 = 16/20), so the apparent differences between models and metrics are within the sampling noise of such a small test set. Please provide bootstrap confidence intervals, multiple fine-tuning seeds, or a larger test set, and ideally a distribution shift to a different data source, to support the claim that the post-trained models generalize.
[Section 2.6, Tables 1 and 2] The before/after comparison changes the output format: Table 1 evaluates the base models with separate questions and no JSON requirement, while Table 2 evaluates the full JSON three-sentence output after fine-tuning. This confounds the effect of training with the effect of a different task format. The claim of significant improvement after post-training should be supported by evaluating the same models with the same output protocol before and after fine-tuning, for example by parsing JSON from the base models or by applying the feature extraction to free-form answers.

minor comments (4)

[Equation (1)] The process name is misspelled as 'Orstein' rather than 'Ornstein', and the initial condition 'r 0 = 0' is typeset unclearly; it should be r_0 = 0.
[Affiliations] The affiliation contains a typo: 'Univeristy' should be 'University'.
[References] Fons et al. 2024a and 2024b appear to be the same arXiv paper; please cite it once or clearly distinguish the two versions.
[Section 2.6] The text says the generated annotations are compared with 'the initial annotations generated by pixtral-large', but Section 2.3 defines the feature-based metric against fact-based sentences; please state explicitly which reference is used for each metric in Tables 1 and 2.

Circularity Check

1 steps flagged · score 4.0 of 10

Post-training feature-based scores are partly self-referential: for the 5% extrema contradictions, the training label is replaced with the exact fact-based sentence that the feature-based metric later uses as ground truth.

self definitional [Section 2.3 (Feature-based similarity scores) and Section 2.4 (Fact-checking), with results reported in Table 2.]
"Feature-based similarity scores We assess the quality of annotations by comparing them to fact-based reference sentences computed based on explicit derivation of time series features ... These three analyses are used to generate fact-checked sentences ... For the ”extrema” annotation, the NLI comparison identifies 5% of cases as contradictions. Some of these reflect genuine inconsistencies. In such cases, we replace the annotation with the corresponding fact-based sentence to ensure the overall quality of the dataset."

The fact-based reference sentence used as the ground truth for the feature-based metric in Section 2.3 is exactly the sentence inserted into the training set in Section 2.4 when pixtral-large's extrema annotation contradicts it. For those samples, the training target and the evaluation reference coincide by construction, so the post-training feature-based extrema score of 0.95 partly measures the model's ability to reproduce the very label it was trained to emit, rather than independent interpretive capability.

full rationale

The paper's central claim—that post-trained Qwen2.5 models acquire meaningful interpretive capabilities—rests substantially on the feature-based scores in Table 2 (1.0 for trend, 0.8 for noise, 0.95 for extrema). Those scores are computed against fact-based reference sentences that are generated by an explicit heuristic: smoothed-trend determination, variance-based noise bins, and computed extrema positions. The same fact-based sentences are used in Section 2.4 to correct teacher annotations, and in the 5% of extrema contradictions the training label is literally replaced with the corresponding fact-based sentence. Thus for these samples, the evaluation target equals the training target by construction. This is not a case of malicious circularity; the training signal for most samples still comes from pixtral-large, and the cosine similarity and NLI scores (0.98, 0.875, etc.) are calculated against the teacher annotations rather than the fact-based heuristic. However, the feature-based metric—the only one producing near-perfect scores—is not independent of the label-curation procedure, and the paper provides no rule-based baseline to establish that a model is needed to achieve these scores. The 20-series test set is also drawn from the same generator as training, further limiting generalization claims, though this is a data-distribution concern rather than a circularity concern. Overall, the feature-based evaluation has a real, localized circular component, but the central result is not entirely forced by definition, so the score is 4 rather than higher.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a chain of labeling and evaluation choices: synthetic OU data stands in for real time series, pixtral-large annotations are treated as teachable content, and a simple feature-extraction heuristic is used both to correct teacher labels and to score students. The feature-based metric in particular risks double-counting the same heuristic as both ground truth and evaluation.

free parameters (3)

Noise intensity thresholds (low/medium/high)
Used to map variance to categorical noise labels in fact-based sentences; thresholds are not specified or calibrated in the paper (Section 2.3).
Extremum position bins (beginning/middle/end)
Used to bucket global extrema locations in fact-based sentences; bin boundaries are not defined (Section 2.3).
Trend smoothing window
Trend is determined from a smoothed version of the series; smoothing method and window are not specified (Section 2.3).

assumptions (5)

standard math The Ornstein-Uhlenbeck recursion in Eq. (1) is a valid discretization of a mean-reverting process.
The paper's data generation relies on this stochastic process, which is a standard mathematical model in finance and biology.
domain assumption Ornstein-Uhlenbeck series with varied (kappa, r, sigma) are representative of real-world temporal patterns for this task.
All training and test data are generated from Eq. (1); real-world time series are far more diverse, so this grounds the scope of the claim.
domain assumption pixtral-large annotations are sufficiently reliable to serve as training targets after fact-checking.
The teacher outputs are the supervision signal; fact-checking only corrects a subset rather than validating all labels (Section 2.4).
domain assumption The feature-extraction heuristics correctly identify ground-truth trend, noise, and extrema.
Fact-based sentences are treated as ground truth for evaluation and correction (Sections 2.3, 2.4).
domain assumption The NLI model roberta-large-mnli correctly maps domain-specific sentences to contradiction, neutral, or entailment.
NLI scores are used both for fact-checking and evaluation, but the NLI model was not validated on time series phrasing (Section 2.3).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Interpretable Time Series Foundation Models." pith.science (2026). https://pith.science/paper/5A7THPJQ

@misc{pith2026250707439,
  author       = {Pith},
  title        = {Pith review of: Towards Interpretable Time Series Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5A7THPJQ}},
  note         = {Machine review of arXiv:2507.07439}
}

read the original abstract

In this paper, we investigate the distillation of time series reasoning capabilities into small, instruction-tuned language models as a step toward building interpretable time series foundation models. Leveraging a synthetic dataset of mean-reverting time series with systematically varied trends and noise levels, we generate natural language annotations using a large multimodal model and use these to supervise the fine-tuning of compact Qwen models. We introduce evaluation metrics that assess the quality of the distilled reasoning - focusing on trend direction, noise intensity, and extremum localization - and show that the post-trained models acquire meaningful interpretive capabilities. Our results highlight the feasibility of compressing time series understanding into lightweight, language-capable models suitable for on-device or privacy-sensitive deployment. This work contributes a concrete foundation toward developing small, interpretable models that explain temporal patterns in natural language.

Figures

Figures reproduced from arXiv: 2507.07439 by the authors.

**Figure 1.** Sample time series image input to pixtral-large. PROMPT: Describe the time series in three sentences. First sentence: describe trend (increasing/decreasing/flat). Second sentence: noise intensity (low/medium/high). Third sentence: approximate localisation of global maximum (beginning/middle/end) and global minimum (beginning/middle/end). Put the description in a JSON format with the following pattern { ”trend”: <… view at source ↗

**Figure 2.** Prompt used to annotate time series by pixtral-large. It is worth noting that such prompting is requiring the model to simply localize the extrema, without explicitly specifying them. ANNOTATION: The time series exhibits an increasing trend initially, followed by fluctuations and a general decreasing trend towards the end. The noise intensity in this time series is high, with significant fluctuations throughout. Th… view at source ↗

**Figure 3.** Sample time series annotation generated by pixtral-large. 3. Feature-based similarity scores We assess the quality of annotations by comparing them to fact-based reference sentences computed based on explicit derivation of time series features using the following procedure. First, the trend is determined based on the overall increasing or decreasing behavior of a smoothed version of the time series. Second, the … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 3 canonical work pages

[3]

Das, A., Kong, W., Sen, R., and Zhou, Y

URL https://arxiv.org/abs/2410.14752. Das, A., Kong, W., Sen, R., and Zhou, Y . A decoder-only foundation model for time-series forecasting,

arXiv
[4]

Daswani, M., Bellaiche, M

URL https://arxiv.org/abs/2310.10688. Daswani, M., Bellaiche, M. M. J., Wilson, M., Ivanov, D., Papkov, M., Schnider, E., Tang, J., Lamerigts, K., Botea, G., Sanchez, M. A., Patel, Y ., Prabhakara, S., Shetty, S., and Telang, U. Plots unlock time-series understanding in multimodal models,

arXiv
[5]

org/abs/2410.02637

URL https://arxiv. org/abs/2410.02637. 4 Submission and Formatting Instructions for ICML 2025 DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., Zhang, X., Yu, X., Wu, Y ., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., Xue, B., Wang, B., Wu, B., Feng, B., Lu, C., Zhao, C., Deng, C., Zhang, ...

arXiv 2025
[8]

Goswami, M., Szafer, K., Choudhry, A., Cai, Y ., Li, S., and Dubrawski, A

URL https://arxiv.org/abs/ 2403.00131. Goswami, M., Szafer, K., Choudhry, A., Cai, Y ., Li, S., and Dubrawski, A. Moment: A family of open time-series foundation models,

arXiv
[9]

Gruver, N., Finzi, M., Qiu, S., and Wilson, A

URL https: //arxiv.org/abs/2402.03885. Gruver, N., Finzi, M., Qiu, S., and Wilson, A. G. Large lan- guage models are zero-shot time series forecasters,

arXiv
[10]

Gunasekar, S., Zhang, Y ., Aneja, J., Mendes, C

URL https://arxiv.org/abs/2310.07820. Gunasekar, S., Zhang, Y ., Aneja, J., Mendes, C. C. T., Giorno, A. D., Gopi, S., Javaheripi, M., Kauffmann, P., de Rosa, G., Saarikivi, O., Salim, A., Shah, S., Behl, H. S., Wang, X., Bubeck, S., Eldan, R., Kalai, A. T., Lee, Y . T., and Li, Y . Textbooks are all you need,

arXiv
[11]

Hinton, G., Vinyals, O., and Dean, J

URL https://arxiv.org/abs/2306.11644. Hinton, G., Vinyals, O., and Dean, J. Distilling the knowledge in a neural network,

arXiv
[13]

Oksendal, B

URL https://arxiv.org/abs/2402.02713. Oksendal, B. Stochastic differential equations, , an intro- duction with applications

arXiv

Show all 22 references

[14]

Rasul, K., Ashok, A., Williams, A

URL https: //arxiv.org/abs/2412.15115. Rasul, K., Ashok, A., Williams, A. R., Ghonia, H., Bhag- watkar, R., Khorasani, A., Bayazi, M. J. D., Adamopou- los, G., Riachi, R., Hassen, N., Bilo ˇs, M., Garg, S., Schneider, A., Chapados, N., Drouin, A., Zantedeschi, V ., Nevmyvaka, ...

arXiv
[15]

Spathis, D

URL https://arxiv.org/abs/2310.08278. Spathis, D. and Kawsar, F. The first step is the hardest: Pitfalls of representing and tokenizing temporal data for large language models,

arXiv
[16]

org/abs/2309.06236

URL https://arxiv. org/abs/2309.06236. Team, N. Sky-t1: Train your own o1 preview model within $450. https://novasky-ai.github.io/posts/sky-t1,

arXiv
[17]

5 Submission and Formatting Instructions for ICML 2025 Wah, E., Wright, M., and Wellman, M

Ac- cessed: 2025-01-09. 5 Submission and Formatting Instructions for ICML 2025 Wah, E., Wright, M., and Wellman, M. P. Welfare effects of market making in continuous double auctions

2025
[19]

org/abs/2402.02592

URL https://arxiv. org/abs/2402.02592. Xu, X., Li, M., Tao, C., Shen, T., Cheng, R., Li, J., Xu, C., Tao, D., and Zhou, T. A survey on knowl- edge distillation of large language models,

arXiv
[20]

Yuan, Z., Yuan, H., Tan, C., Wang, W., and Huang, S

URL https://arxiv.org/abs/2402.13116. Yuan, Z., Yuan, H., Tan, C., Wang, W., and Huang, S. How well do large language models perform in arith- metic tasks?,

arXiv
[21]

URL https://arxiv.org/ abs/2304.02015. Zhou, Z. and Yu, R. Can llms understand time se- ries anomalies?,

arXiv
[22]

URL https://arxiv.org/ abs/2410.05440. 6

arXiv
[2015]

Jin, M., Zhang, Y ., Chen, W., Zhang, K., Liang, Y ., Yang, B., Wang, J., Pan, S., and Wen, Q

URL https: //arxiv.org/abs/1503.02531. Jin, M., Zhang, Y ., Chen, W., Zhang, K., Liang, Y ., Yang, B., Wang, J., Pan, S., and Wen, Q. Position: What can large language models tell us about time series analysis,

arXiv
[2018]

org/abs/1704.05426

URL https://arxiv. org/abs/1704.05426. Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., and Sahoo, D. Unified training of universal time series fore- casting transformers,

arXiv
[2019]

URL https://arxiv.org/abs/1909. 11650. Cai, Y ., Choudhry, A., Goswami, M., and Dubrawski, A. Timeseriesexam: A time series understanding exam,

1909
[2023]

Fons, E., Kaur, R., Palande, S., Zeng, Z., Balch, T., Veloso, M., and Vyetrenko, S

URL https://arxiv.org/abs/2305.07759. Fons, E., Kaur, R., Palande, S., Zeng, Z., Balch, T., Veloso, M., and Vyetrenko, S. Evaluating large language models on time series feature understanding: A com- prehensive taxonomy and benchmark. arXiv preprint arXiv:2404.16563, 2024a. Fo...

arXiv
[2024]

URL https://arxiv.org/abs/ 2309.16741. Byrd, D. Explaining agent-based financial market simula- tion,

arXiv
[2025]

Eldan, R

URL https://arxiv.org/abs/2501.12948. Eldan, R. and Li, Y . Tinystories: How small can language models be and still speak coherent english?,

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [3]

Das, A., Kong, W., Sen, R., and Zhou, Y

URL https://arxiv.org/abs/2410.14752. Das, A., Kong, W., Sen, R., and Zhou, Y . A decoder-only foundation model for time-series forecasting,

arXiv

[2] [4]

Daswani, M., Bellaiche, M

URL https://arxiv.org/abs/2310.10688. Daswani, M., Bellaiche, M. M. J., Wilson, M., Ivanov, D., Papkov, M., Schnider, E., Tang, J., Lamerigts, K., Botea, G., Sanchez, M. A., Patel, Y ., Prabhakara, S., Shetty, S., and Telang, U. Plots unlock time-series understanding in multimodal models,

arXiv

[3] [5]

org/abs/2410.02637

URL https://arxiv. org/abs/2410.02637. 4 Submission and Formatting Instructions for ICML 2025 DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., Zhang, X., Yu, X., Wu, Y ., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., Xue, B., Wang, B., Wu, B., Feng, B., Lu, C., Zhao, C., Deng, C., Zhang, ...

arXiv 2025

[4] [8]

Goswami, M., Szafer, K., Choudhry, A., Cai, Y ., Li, S., and Dubrawski, A

URL https://arxiv.org/abs/ 2403.00131. Goswami, M., Szafer, K., Choudhry, A., Cai, Y ., Li, S., and Dubrawski, A. Moment: A family of open time-series foundation models,

arXiv

[5] [9]

Gruver, N., Finzi, M., Qiu, S., and Wilson, A

URL https: //arxiv.org/abs/2402.03885. Gruver, N., Finzi, M., Qiu, S., and Wilson, A. G. Large lan- guage models are zero-shot time series forecasters,

arXiv

[6] [10]

Gunasekar, S., Zhang, Y ., Aneja, J., Mendes, C

URL https://arxiv.org/abs/2310.07820. Gunasekar, S., Zhang, Y ., Aneja, J., Mendes, C. C. T., Giorno, A. D., Gopi, S., Javaheripi, M., Kauffmann, P., de Rosa, G., Saarikivi, O., Salim, A., Shah, S., Behl, H. S., Wang, X., Bubeck, S., Eldan, R., Kalai, A. T., Lee, Y . T., and Li, Y . Textbooks are all you need,

arXiv

[7] [11]

Hinton, G., Vinyals, O., and Dean, J

URL https://arxiv.org/abs/2306.11644. Hinton, G., Vinyals, O., and Dean, J. Distilling the knowledge in a neural network,

arXiv

[8] [13]

Oksendal, B

URL https://arxiv.org/abs/2402.02713. Oksendal, B. Stochastic differential equations, , an intro- duction with applications

arXiv