arxiv: 2605.12550 · v1 · submitted 2026-05-10 · 💻 cs.CV · cs.AI

Recognition: no theorem link

SSDA: Bridging Spectral and Structural Gaps via Dual Adaptation for Vision-Based Time Series Forecasting

Mingrui Zhang , Hanchen Yang , Wengen Li , Xudong Jiang , Yichao Zhang , Jihong Guan , Shuigeng Zhou

Authors on Pith no claims yet

Pith reviewed 2026-05-14 21:36 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords time series forecastingvision-based forecastingspectral adaptationlow-rank adaptationimage renderingdual adaptationtransfer learningpower spectrum

0 comments

The pith

SSDA adapts pre-trained vision models for time series by closing spectral and structural gaps in rendered images.

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

The paper argues that rendering time series as images for large vision models leaves two unaddressed mismatches with the natural images those models were trained on. A shallower power spectrum and fabricated spatial relationships from 1D-to-2D reshaping both limit how well pre-trained knowledge transfers. SSDA counters this with a data-level Spectral Magnitude Aligner that boosts magnitude spectra toward natural-image statistics while keeping phase intact, plus a model-level Structural-Guided LoRA that injects temporal position information into patches and adapts attention layers. These two branches fuse adaptively to produce forecasts. The approach is shown to improve accuracy over strong baselines on seven real-world datasets in both full-data and few-shot regimes.

Core claim

Large vision models can forecast time series when the data is rendered as images, yet spectral and structural gaps still separate those images from the natural images the models were pre-trained on. The spectral gap appears as a markedly shallower power spectrum; the structural gap arises because 1D sequences reshaped into 2D grids create spurious spatial adjacencies and break genuine temporal order. SSDA closes both gaps through a dual-branch architecture: the Spectral Magnitude Aligner applies 2D FFT to selectively enhance magnitude spectra while preserving phase, and Structural-Guided Low-Rank Adaptation injects position-aware temporal encodings into patch embeddings and performs low-rank

What carries the argument

SSDA dual-branch network whose Spectral Magnitude Aligner (SMA) realigns image spectra and whose Structural-Guided Low-Rank Adaptation (SG-LoRA) injects temporal encodings and adapts attention.

If this is right

SSDA outperforms both LVM- and LLM-based baselines on seven real-world time series benchmarks under full-shot and few-shot regimes.
The dual adaptation works at both data level (spectrum alignment) and model level (temporal-aware low-rank updates) without requiring full retraining.
Adaptive fusion of the spectral and structural branches yields the final forecast.
The method remains effective when only limited labeled time series data is available.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar spectral and structural mismatches may appear when other 1D signals are rendered as images for vision models, suggesting the same dual-adaptation pattern could apply to audio or sensor data.
If the phase-preserving property of SMA proves robust, it may allow further spectrum manipulations without destroying temporal ordering information.
Testing whether SG-LoRA's position encodings remain beneficial when the underlying vision backbone changes would clarify how general the structural fix is.

Load-bearing premise

The gaps identified in spectrum and spatial structure are the main factors limiting transfer from natural-image pre-training.

What would settle it

Running the same LVM backbones with and without SMA plus SG-LoRA on the seven benchmarks and finding no consistent accuracy gain would falsify the claim that closing these gaps is what unlocks better forecasts.

Figures

Figures reproduced from arXiv: 2605.12550 by Hanchen Yang, Jihong Guan, Mingrui Zhang, Shuigeng Zhou, Wengen Li, Xudong Jiang, Yichao Zhang.

**Figure 2.** Figure 2: The Overall architecture of SSDA. patches far better than with discrete tokens, prompting a growing line of LVM adaptations [2, 46, 29]. However, current LVM adaptations rely on constrained strategies such as freezing the backbone, attaching lightweight adapters, or performing costly continual pre-training [31]. These strategies adjust what the model sees but not how it reads patches; neither the spectral … view at source ↗

**Figure 3.** Figure 3: Sensitivity analysis of key hyper-parameters. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Visualization of structural gap mitigation on an ETTm1 sample. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Power spectrum analysis of different modalities. Each subfigure shows the power spectrum [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

**Figure 6.** Figure 6: Power spectral slope analysis for different modalities. Subfigures (a)-(c) show individual [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗

**Figure 7.** Figure 7: Power spectral slope analysis for time series data rendered with different periodicity values. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

**Figure 8.** Figure 8: The bubble chart of performance vs. efficiency. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗

**Figure 9.** Figure 9: Forecasting visualization on a sample from ETTm2. (a) Input image of structural branch. [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗

**Figure 10.** Figure 10: Forecasting visualization on a sample from ETTh1. (a) Input image of structural branch. [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗

**Figure 11.** Figure 11: Forecasting visualization on a sample from ETTh2. (a) Input image of structural branch. [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗

read the original abstract

Large vision models (LVMs) have recently proven to be surprisingly effective time series forecasters, simply by rendering temporal data as images. This success, how ever, rests on a largely unexamined premise: the rendered time series images are sufficiently close to natural images for knowledge in pre-trained models to transfer effectively. We argue that two gaps still remain, i.e., spectral and structural gaps, fundamentally limiting the potential of LVMs for time series forecasting. Spectrally, we systematically reveal that rendered time series images exhibit a markedly shallower power spectrum than the natural images LVMs are pre-trained to recognize. Structurally, reshaping 1D temporal sequences into 2D grids fabricates spurious spatial adjacencies while severing genuine temporal continuities, misleading the spatial inductive biases of pre-trained LVMs. To bridge these gaps, we propose SSDA, a dual-branch network that spectrally and structurally adapts to unlock the full potential of LVMs for time series forecasting. At the data level, a Spectral Magnitude Aligner (SMA) applies 2D FFT to selectively enhance the magnitude spectrum toward natural-image statistics while preserving phase. At the model level, a Structural-Guided Low-Rank Adaptation (SG-LoRA) injects position-aware temporal encodings into patch embeddings and adapts at tention via low-rank updates. The two branches are further adaptively fused to produce the final forecast. Extensive experiments on seven real-world benchmarks demonstrate that SSDA consistently outperforms strong LVM- and LLM-based baselines under both full-shot and few-shot settings. Code is publicly available at https://anonymous.4open.science/r/SSDA-8C5B.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SSDA adds targeted spectral and structural adapters to vision models for time series but the experiments do not isolate whether gains come from closing the claimed gaps or from extra adaptation capacity.

read the letter

The paper's core move is to treat rendered time series images as having two concrete mismatches with natural-image pretraining: a shallower power spectrum and fabricated 2D spatial adjacencies that break real temporal order. SSDA counters the first with a 2D-FFT magnitude aligner that boosts frequencies toward natural-image statistics while keeping phase intact, and the second with a structural-guided LoRA that injects position-aware temporal encodings into patches and applies low-rank updates to attention. The two branches are fused adaptively. This is a clean, explicit engineering response to the gaps the authors identify, and the abstract reports consistent wins over LVM and LLM baselines across seven benchmarks in both full-shot and few-shot regimes. Code release is a plus for anyone who wants to test it directly. The approach is straightforward enough that a practitioner could implement the modules without much trouble. The main weakness is that the results do not yet pin the improvements to the specific mechanisms. No capacity-matched ablations are described that would show whether a generic dual-branch setup with the same number of extra parameters would perform similarly, and there are no reported diagnostics confirming that the SMA actually moves the spectrum closer to natural-image statistics on the training data. Without those controls the attribution to gap-closing remains plausible but unverified. The work is aimed at researchers adapting large vision models to time series forecasting. It gives them a concrete recipe and benchmark numbers worth checking, so it is worth a reading-group discussion even if the causal story needs more evidence. I would send it to peer review; the problem framing is useful and the method is reproducible enough that referees can give targeted feedback on the experimental design.

Referee Report

3 major / 3 minor

Summary. The paper argues that rendering time series as images for pre-trained large vision models (LVMs) is limited by two gaps: a spectral gap (shallower power spectrum than natural images) and a structural gap (fabricated spatial adjacencies that break temporal continuity). It proposes SSDA, a dual-branch architecture with a Spectral Magnitude Aligner (SMA) that uses 2D FFT to enhance magnitude spectra toward natural-image statistics while preserving phase, and Structural-Guided Low-Rank Adaptation (SG-LoRA) that injects position-aware temporal encodings into patch embeddings and performs low-rank attention updates. The branches are fused adaptively, and the method is shown to outperform LVM- and LLM-based baselines on seven real-world benchmarks under both full-shot and few-shot regimes.

Significance. If the gains are causally attributable to closing the identified gaps rather than generic increases in adaptation capacity, the work would provide a principled way to improve transfer from natural-image pre-training to time series forecasting, with particular value in few-shot regimes. The public code release is a positive factor for reproducibility.

major comments (3)

[Experiments] Experiments section: the reported outperformance on seven benchmarks lacks controlled ablations that hold total parameter count fixed when adding SMA and SG-LoRA; without such controls it remains unclear whether gains arise from spectral/structural alignment or from the dual-branch architecture's extra capacity.
[Method] Method section (SMA description): no quantitative diagnostics (e.g., pre- vs. post-SMA power-spectrum slope comparisons or Kolmogorov-Smirnov tests against natural-image statistics) are provided to confirm that the magnitude enhancement actually reduces the claimed spectral gap.
[Introduction] Introduction and §3: the structural-gap claim (that fabricated 2D adjacencies dominate transfer failure) is central but untested; a minimal diagnostic would be to compare a temporally contiguous reshaping baseline against the standard grid reshaping while keeping all other components fixed.

minor comments (3)

[Abstract] Abstract: 'how ever' should be 'however'.
[Abstract] Abstract: 'at tention' should be 'attention'.
[Abstract] Abstract: the anonymous code link should be replaced with a permanent repository URL upon acceptance.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight opportunities to strengthen the empirical support for our claims regarding the spectral and structural gaps. We address each major comment below and will revise the manuscript accordingly to incorporate the suggested controls and diagnostics.

read point-by-point responses

Referee: [Experiments] Experiments section: the reported outperformance on seven benchmarks lacks controlled ablations that hold total parameter count fixed when adding SMA and SG-LoRA; without such controls it remains unclear whether gains arise from spectral/structural alignment or from the dual-branch architecture's extra capacity.

Authors: We agree that controlling for parameter count is necessary to isolate the contribution of the proposed adaptations. In the revised manuscript, we will add new ablation experiments that match the total parameter budget of SSDA. These will include comparisons against a dual-branch baseline using generic (non-spectral, non-structural) adaptations of equivalent capacity, such as standard LoRA without temporal encodings and random magnitude perturbations, to demonstrate that performance gains are attributable to the targeted gap-closing mechanisms rather than added capacity alone. revision: yes
Referee: [Method] Method section (SMA description): no quantitative diagnostics (e.g., pre- vs. post-SMA power-spectrum slope comparisons or Kolmogorov-Smirnov tests against natural-image statistics) are provided to confirm that the magnitude enhancement actually reduces the claimed spectral gap.

Authors: We acknowledge that direct quantitative evidence for the spectral alignment effect would strengthen the method description. In the revision, we will add power-spectrum analysis figures and tables comparing the slope and distribution statistics (including Kolmogorov-Smirnov distances) of rendered time series images before and after SMA application against natural-image references, confirming the reduction in the spectral gap. revision: yes
Referee: [Introduction] Introduction and §3: the structural-gap claim (that fabricated 2D adjacencies dominate transfer failure) is central but untested; a minimal diagnostic would be to compare a temporally contiguous reshaping baseline against the standard grid reshaping while keeping all other components fixed.

Authors: We appreciate this suggestion for validating the structural gap. In the revised version, we will include a controlled experiment that replaces the standard grid reshaping with a temporally contiguous reshaping baseline (preserving sequential order more explicitly in the 2D layout) while keeping the model architecture, training protocol, and all other components identical. Results from this diagnostic will be reported to quantify the impact of fabricated adjacencies. revision: yes

Circularity Check

0 steps flagged

No significant circularity; new modules and empirical claims are independent of target metrics

full rationale

The derivation introduces SMA (2D-FFT magnitude enhancement preserving phase) and SG-LoRA (position-aware encodings plus low-rank attention updates) as explicit new components whose parameters are optimized from data. The central claim of outperformance on seven benchmarks is presented as an empirical result rather than a quantity defined in terms of itself or a fitted input renamed as prediction. No equations or steps reduce by construction to the target forecasting performance, and no load-bearing self-citations or uniqueness theorems imported from the authors' prior work are invoked to force the result. The paper is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of two new modules whose internal hyperparameters (e.g., adaptation rank, spectrum scaling factors) are not enumerated in the abstract; no new physical entities or unstated mathematical axioms are introduced.

pith-pipeline@v0.9.0 · 5634 in / 1150 out tokens · 30060 ms · 2026-05-14T21:36:04.421783+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages

[1]

Fredn: Spectral disentanglement for time series forecasting via learnable frequency decomposition

Zhongde An, Jinhong You, Jiyanglin Li, Yiming Tang, Wen Li, Heming Du, and Shouguo Du. Fredn: Spectral disentanglement for time series forecasting via learnable frequency decomposition. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 19623–19631, 2026

work page 2026
[2]

Visionts: Visual masked autoencoders are free-lunch zero-shot time series forecasters

Mouxiang Chen, Lefei Shen, Zhuo Li, Xiaoyun Joy Wang, Jianling Sun, and Chenghao Liu. Visionts: Visual masked autoencoders are free-lunch zero-shot time series forecasters. In International Conference on Machine Learning, pages 8979–9007. PMLR, 2025

work page 2025
[3]

Sparse learned kernels for interpretable and efficient medical time series processing.Nature machine intelligence, 6(10):1132–1144, 2024

Sully F Chen, Zhicheng Guo, Cheng Ding, Xiao Hu, and Cynthia Rudin. Sparse learned kernels for interpretable and efficient medical time series processing.Nature machine intelligence, 6(10):1132–1144, 2024

work page 2024
[4]

Imagenet: A large- scale hierarchical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009

work page 2009
[5]

An image is worth 16x16 words: Transformers for image recognition at scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representations, pages 1–21, 2021

work page 2021
[6]

Amplifier: Bringing attention to neglected low-energy components in time series forecasting

Jingru Fei, Kun Yi, Wei Fan, Qi Zhang, and Zhendong Niu. Amplifier: Bringing attention to neglected low-energy components in time series forecasting. InProceedings of the AAAI conference on artificial intelligence, volume 39, pages 11645–11653, 2025

work page 2025
[7]

Masked autoencoders as spatiotemporal learners

Christoph Feichtenhofer, Haoqi Fan, and Yanghao Li Kaiming He. Masked autoencoders as spatiotemporal learners. InProceedings of the 36th International Conference on Neural Information Processing Systems, pages 35946–35958, 2022

work page 2022
[8]

D. J. Field. Relations between the statistics of natural images and the response properties of cortical cells.Journal of the Optical Society of America. A, Optics and Image Science, 4(12):2379–2394, 1987

work page 1987
[9]

Text-to-audio generation using instruction guided latent diffusion model

Deepanway Ghosal, Navonil Majumder, Ambuj Mehrish, and Soujanya Poria. Text-to-audio generation using instruction guided latent diffusion model. InProceedings of the 31st ACM international conference on multimedia, pages 3590–3598, 2023

work page 2023
[10]

Moment: A family of open time-series foundation models

Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. Moment: A family of open time-series foundation models. InInternational Conference on Machine Learning, pages 16115–16152. PMLR, 2024

work page 2024
[11]

Textually pretrained speech language models

Michael Hassid, Tal Remez, Tu Anh Nguyen, Itai Gat, Alexis Conneau, Felix Kreuk, Jade Copet, Alexandre Defossez, Gabriel Synnaeve, Emmanuel Dupoux, et al. Textually pretrained speech language models. InAdvances in Neural Information Processing Systems, volume 36, pages 63483–63501, 2023

work page 2023
[12]

Masked autoencoders are scalable vision learners

Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 16000–16009, 2022

work page 2022
[13]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

work page 2016
[14]

Timeapn: Adaptive amplitude-phase non-stationarity normalization for time series forecasting.arXiv preprint arXiv:2603.17436, 2026

Yue Hu, Jialiang Tang, Siwei Yu, Baosheng Yu, Jing Zhang, and Dacheng Tao. Timeapn: Adaptive amplitude-phase non-stationarity normalization for time series forecasting.arXiv preprint arXiv:2603.17436, 2026. 10

work page arXiv 2026
[15]

Adamamba: Adaptive frequency-gated mamba for long-term time series forecasting, 2026

Xudong Jiang, Mingshan Loo, Hanchen Yang, Wengen Li, Mingrui Zhang, Yichao Zhang, Jihong Guan, and Shuigeng Zhou. Adamamba: Adaptive frequency-gated mamba for long-term time series forecasting, 2026

work page 2026
[16]

Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen

Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y . Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. Time-LLM: Time series forecasting by reprogramming large language models. InThe Twelfth International Conference on Learning Representations, 2024

work page 2024
[17]

Vilt: Vision-and-language transformer without convolution or region supervision

Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. InInternational conference on machine learning, pages 5583–5594. PMLR, 2021

work page 2021
[18]

Forecasting with time series imaging.Expert Systems with Applications, 160:113680, 2020

Xixi Li, Yanfei Kang, and Feng Li. Forecasting with time series imaging.Expert Systems with Applications, 160:113680, 2020

work page 2020
[19]

Foundation models for time series analysis: A tutorial and survey

Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. Foundation models for time series analysis: A tutorial and survey. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6555–6565, 2024

work page 2024
[20]

Timecma: Towards llm-empowered multivariate time series forecasting via cross-modality alignment

Chenxi Liu, Qianxiong Xu, Hao Miao, Sun Yang, Lingzheng Zhang, Cheng Long, Ziyue Li, and Rui Zhao. Timecma: Towards llm-empowered multivariate time series forecasting via cross-modality alignment. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18780–18788, 2025

work page 2025
[21]

Text2freq: Learning series patterns from text via frequency domain

Ming-Chih Lo, Ching Chang, and Wen-Chih Peng. Text2freq: Learning series patterns from text via frequency domain. InNeurIPS Workshop on Time Series in the Age of Large Models, 2024

work page 2024
[22]

Frozen pretrained transformers as universal computation engines

Kevin Lu, Aditya Grover, Pieter Abbeel, and Igor Mordatch. Frozen pretrained transformers as universal computation engines. InProceedings of the AAAI conference on artificial intelligence, volume 36, pages 7628–7636, 2022

work page 2022
[23]

Spectral text fu- sion: A frequency-aware approach to multimodal time-series forecasting.arXiv preprint arXiv:2602.01588, 2026

Huu Hiep Nguyen, Minh Hoang Nguyen, Dung Nguyen, and Hung Le. Spectral text fu- sion: A frequency-aware approach to multimodal time-series forecasting.arXiv preprint arXiv:2602.01588, 2026

work page arXiv 2026
[24]

Harnessing vision models for time series analysis: a survey

Jingchao Ni, Ziming Zhao, ChengAo Shen, Hanghang Tong, Dongjin Song, Wei Cheng, Dongsheng Luo, and Haifeng Chen. Harnessing vision models for time series analysis: a survey. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pages 10612–10620, 2025

work page 2025
[25]

A time series is worth 64 words: Long-term forecasting with transformers

Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. InThe Eleventh International Conference on Learning Representations, pages 1–10, 2023

work page 2023
[26]

Understanding the role of textual prompts in llm for time series forecasting: an adapter view.arXiv preprint arXiv:2311.14782, 2023

Peisong Niu, Tian Zhou, Xue Wang, Liang Sun, and Rong Jin. Understanding the role of textual prompts in llm for time series forecasting: an adapter view.arXiv preprint arXiv:2311.14782, 2023

work page arXiv 2023
[27]

Learning transferable visual models from natural language supervision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PMLR, 2021

work page 2021
[28]

Statistics of natural images: scaling in the woods

Daniel L Ruderman and William Bialek. Statistics of natural images: scaling in the woods. In Proceedings of the 7th International Conference on Neural Information Processing Systems, pages 551–558, 1993

work page 1993
[29]

Multi-modal view enhanced large vision models for long-term time series forecast- ing

ChengAo Shen, Wenchao Yu, Ziming Zhao, Dongjin Song, Wei Cheng, Haifeng Chen, and Jingchao Ni. Multi-modal view enhanced large vision models for long-term time series forecast- ing. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, pages 1–28, 2025. 11

work page 2025
[30]

Cross-modal fine-tuning: Align then refine

Junhong Shen, Liam Li, Lucio M Dery, Corey Staten, Mikhail Khodak, Graham Neubig, and Ameet Talwalkar. Cross-modal fine-tuning: Align then refine. InInternational Conference on Machine Learning, pages 31030–31056. PMLR, 2023

work page 2023
[31]

Visionts++: Cross-modal time series foundation model with continual pre-trained vision backbones,

Lefei Shen, Mouxiang Chen, Xu Liu, Han Fu, Xiaoxue Ren, Jianling Sun, Zhuo Li, and Chenghao Liu. Visionts++: Cross-modal time series foundation model with continual pre- trained vision backbones.arXiv preprint arXiv:2508.04379, 2025

work page arXiv 2025
[32]

Very deep convolutional networks for large-scale image recogni- tion

K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recogni- tion. InInternational Conference on Learning Representations, pages 1–14. Computational and Biological Learning Society, 2015

work page 2015
[33]

TEST: Text prototype aligned embedding to activate LLM’s ability for time series

Chenxi Sun, Hongyan Li, Yaliang Li, and Shenda Hong. TEST: Text prototype aligned embedding to activate LLM’s ability for time series. InThe Twelfth International Conference on Learning Representations, pages 1–21, 2024

work page 2024
[34]

ElectricityLoadDiagrams20112014

Artur Trindade. ElectricityLoadDiagrams20112014. UCI Machine Learning Repository, 2015

work page 2015
[35]

Cross-modal projection in multimodal llms doesn’t really project visual attributes to textual space

Gaurav Verma, Minje Choi, Kartik Sharma, Jamelle Watson-Daniels, Sejoon Oh, and Srijan Kumar. Cross-modal projection in multimodal llms doesn’t really project visual attributes to textual space. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 657–664, 2024

work page 2024
[36]

Timesnet: Temporal 2d-variation modeling for general time series analysis

Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. InThe Eleventh International Conference on Learning Representations, pages 1–10, 2023

work page 2023
[37]

Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting

Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. InAdvances in neural information processing systems, volume 34, pages 22419–22430, 2021

work page 2021
[38]

Towards robust and interpretable spatial-temporal graph modeling for traffic prediction.ACM Transactions on Knowledge Discovery from Data, 19(9):1– 20, 2025

Hanchen Yang, Jiannong Cao, Wengen Li, Yu Yang, Xiaoyi Li, Lingbai Kong, Yichao Zhang, Jihong Guan, and Shuigeng Zhou. Towards robust and interpretable spatial-temporal graph modeling for traffic prediction.ACM Transactions on Knowledge Discovery from Data, 19(9):1– 20, 2025

work page 2025
[39]

Vitime: Foundation model for time series forecasting powered by vision intelligence.Transactions on Machine Learning Research, 2025

Luoxiao Yang, Yun Wang, Xinqi Fan, Israel Cohen, jingdong chen, and Zijun Zhang. Vitime: Foundation model for time series forecasting powered by vision intelligence.Transactions on Machine Learning Research, 2025

work page 2025
[40]

Loft-llm: Low-frequency time-series forecast- ing with large language models

Jiacheng You, Jingcheng Yang, Yuhang Xie, Zhongxuan Wu, Xiucheng Li, Feng Li, Pengjie Wang, Jian Xu, Bo Zheng, and Xinyang Chen. Loft-llm: Low-frequency time-series forecast- ing with large language models. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1, pages 1809–1820, 2026

work page 2026
[41]

Harnessing llms for temporal data-a study on explainable financial time series forecasting

Xinli Yu, Zheng Chen, and Yanbin Lu. Harnessing llms for temporal data-a study on explainable financial time series forecasting. InProceedings of the 2023 conference on empirical methods in natural language processing: industry track, pages 739–753, 2023

work page 2023
[42]

Zhenxiang Zang, Yang Qiao, Shaozhen Yan, and Jie Lu. Reliability and validity of power spectrum slope (pss): A metric for measuring resting-state functional magnetic resonance imaging activity of single voxels.Frontiers in Neuroscience, 16:871609, 2022

work page 2022
[43]

Are transformers effective for time series forecasting?Proceedings of the AAAI Conference on Artificial Intelligence, 37(9):11121–11128, Jun

Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting?Proceedings of the AAAI Conference on Artificial Intelligence, 37(9):11121–11128, Jun. 2023

work page 2023
[44]

From pixels to predictions: Spectrogram and vision transformer for better time series forecasting

Zhen Zeng, Rachneet Kaur, Suchetha Siddagangappa, Tucker Balch, and Manuela Veloso. From pixels to predictions: Spectrogram and vision transformer for better time series forecasting. In Proceedings of the Fourth ACM International Conference on AI in Finance, pages 82–90, 2023

work page 2023
[45]

Piformer: Towards subseasonal sst prediction with spatial-patched inverted transformer.Expert Systems with Applications, page 131618, 2026

Mingrui Zhang, Hanchen Yang, Wengen Li, Xudong Jiang, Jihong Guan, Yichao Zhang, and Shuigeng Zhou. Piformer: Towards subseasonal sst prediction with spatial-patched inverted transformer.Expert Systems with Applications, page 131618, 2026. 12

work page 2026
[46]

Time- vlm: Exploring multimodal vision-language models for augmented time series forecasting

Siru Zhong, Weilin Ruan, Ming Jin, Huan Li, Qingsong Wen, and Yuxuan Liang. Time- vlm: Exploring multimodal vision-language models for augmented time series forecasting. In International Conference on Machine Learning, pages 78478–78497. PMLR, 2025

work page 2025
[47]

Informer: Beyond efficient transformer for long sequence time-series forecasting

Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11106–11115, 2021

work page 2021
[48]

Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting

Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. InInternational conference on machine learning, pages 27268–27286. PMLR, 2022

work page 2022
[49]

Limitations

Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al. One fits all: Power general time series analysis by pretrained lm. InAdvances in neural information processing systems, volume 36, pages 43322–43355, 2023. 13 A Modality Gap Analysis A.1 Experimental Principle Frequency Distribution and Its Physical SignificanceAny signal in the spatial (or temporal) dom...

work page 2023
[50]

Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

work page