Pith. sign in

REVIEW 2 major objections 2 minor 39 references

PARA-PV: Physics-Aware Retrieval-Augmented PV Prediction Based on Frozen Foundation Model and Distribution Shift Correction

T0 review · 2 major / 2 minor · reviewed 2026-07-10 · glm-5.2

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

Pith's one-line read Frozen foundation model plus physics-aware retrieval cuts PV forecasting errors across weather shifts

desk verdict Abstract-only review: novel system design for PV forecasting, but no results to evaluate read the letter →

arxiv 2607.08079 v1 pith:LUG7HSJP submitted 2026-07-09 cs.AI

Hang Fan , Weican Liu , Ying Lu , Dunnan Liu , Long Cheng , Wei Wei This is my paper · ORCID
classification cs.AI
keywords PVforecastingretrieval-augmentedlearningfoundationmodelsdistributionshiftcorrectionphysics-awarelossrenewableenergytime-seriesChronos
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 solar power forecasting can be made substantially more accurate and robust by weaving physical knowledge through every stage of prediction rather than treating physics as a post-hoc constraint. The authors build PARA-PV, a pipeline with four linked stages. First, a physics-aware retrieval module finds historical patches of solar generation data that match the current conditions in temporal shape, power level, operating state, and intra-day period, producing a base forecast grounded in analogous past behavior. Second, this local-memory forecast is calibrated against a frozen time-series foundation model (Chronos) via a lightweight residual adapter, injecting broad temporal regularities without overwriting the physics-grounded prediction. Third, because weather and day-night regime changes cause residual distribution shifts that the first two stages cannot fully capture, a gated correction module applies selective mean-shift and scale adjustments using power, weather, timestamp, and day/night signals. Fourth, a physics-constrained loss function partitions samples into peak, ramping, night-time, and regular regimes, adaptively reweighting error contributions so that the numerically dominant regular regime does not crowd out learning of operationally critical but rare states. The central argument is that each of these four components addresses a distinct failure mode of conventional PV forecasting — poor analog matching, missing temporal priors, regime-induced distribution shift, and loss-function imbalance — and that their combination yields forecasts that remain reliable when weather and diurnal conditions change.

What carries the argument

PARA-PV pipeline: (1) patch-level physics-aware retrieval of historical analogs matched on temporal shape, power level, operating state, and intra-day period; (2) frozen Chronos foundation model with lightweight residual adapter for temporal-prior calibration; (3) gated mean-shift and scale correction module conditioned on power, weather, timestamp, and day/night signals; (4) physics-constrained loss partitioning samples into peak, ramping, night-time, and regular regimes with adaptive error reweighting.

What would settle it

Construct a test where the current PV window appears normal (clear-sky power levels, regular temporal shape) but an imminent weather event (cloud front, fog bank) is detectable in upstream meteorological data yet absent from the current PV observation window. If the retrieval module matches clear-sky analogs and the downstream corrections cannot compensate fast enough, PARA-PV would produce large forecast errors on exactly the ramp events its loss function is designed to prioritize. Conversely, if the gated correction module can detect the impending shift from weather features alone and adjust

Watch

Extended reading notes

Core claim

The core discovery is an architectural decomposition of the PV forecasting problem into four sequential, physics-aware stages — retrieval of physically consistent analogs, residual adaptation of a frozen foundation-model prior, gated distribution-shift correction, and regime-aware loss reweighting — each targeting a specific source of forecasting error that conventional methods conflate or ignore. The mechanism that carries the argument is the interplay between the frozen foundation model (providing general temporal regularities) and the physics-aware retrieval system (providing locally grounded, condition-matched analogs), with the gated correction module serving as the bridge that handles残

Load-bearing premise

The load-bearing premise is that the physics-aware retrieval module consistently finds historical analogs whose weather dynamics genuinely predict future PV trajectories. If the matched patches resemble the current window in surface features (power level, temporal shape) but miss the causal weather drivers — an approaching cloud front not yet visible in the current window — then the base forecast starts from an unreliable foundation, and the downstream correction stages, no

Editorial extensions

If this is right

  • Grid operators could deploy regime-aware forecasting that maintains accuracy during operationally critical but rare events — ramps, peaks, night-time transitions — rather than only optimizing for average-case clear-sky performance.
  • The residual-adapter approach to leveraging frozen foundation models could generalize beyond PV to other renewable forecasting domains (wind, load) where physical constraints and distribution shifts co-occur.
  • The regime-partitioned loss function could become a standard component in energy forecasting benchmarks, preventing the systematic underperformance on rare but grid-critical events that average-error metrics mask.
  • If the retrieval module's analog-matching is genuinely capturing causal weather dynamics, the same retrieval-augmented architecture could be applied to other physical systems where historical analogs under similar operating conditions carry predictive signal.

Reading between the lines

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

  • The four-stage decomposition implicitly claims that the failure modes are separable and approximately sequential — that retrieval quality, temporal-prior calibration, distribution-shift correction, and loss balancing can be optimized in stages rather than jointly. If these failure modes are deeply entangled, the sequential architecture may leave residual errors that no single stage can correct.
  • The choice to keep the foundation model frozen and adapt only via a lightweight residual adapter suggests a hypothesis that general temporal regularities are already well-captured by the pre-trained model and need only PV-specific calibration. Testing whether full fine-tuning of the foundation model would outperform the frozen-plus-adapter approach would clarify whether the frozen design is driven
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.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes PARA-PV, a multi-stage pipeline for photovoltaic (PV) power forecasting that combines physics-aware retrieval of historical analog trajectories, a frozen Chronos time-series foundation model prior with a lightweight residual adapter, gated distribution shift correction, and a regime-aware loss function. The design is motivated by the challenges of weather variability, day-night transitions, and regime-dependent dynamics. The authors state that code is available at a public repository. However, only the abstract was available for this review; the full text, including experimental results, ablation studies, and baseline comparisons, could not be examined.

Significance. The architectural design is thoughtful and addresses a genuine problem in PV forecasting. The use of a frozen foundation model as a prior with residual adaptation, rather than fine-tuning, is a reasonable strategy for leveraging general temporal knowledge while preserving PV-specific physics. Public code availability is a positive signal for reproducibility. However, the significance of this work cannot be fully assessed without quantitative evidence: the abstract contains no error metrics, no baseline comparisons, and no ablation results demonstrating that each component contributes to overall performance.

major comments (2)
  1. No quantitative results are available in the abstract. The central claim—that embedding physical knowledge throughout the pipeline yields more accurate and robust PV forecasts—cannot be evaluated without error metrics (e.g., RMSE, MAE), baseline comparisons against standard forecasting methods, and ablation studies isolating the contribution of each module (retrieval, residual adapter, gated correction, regime-aware loss). This is the primary load-bearing concern: the paper's contribution is entirely empirical, and the evidence is absent from what was reviewable.
  2. The abstract describes a pipeline with multiple fitted components (residual adapter parameters, gated mean-shift and scale correction parameters, regime-aware loss reweighting coefficients). Without access to the full text, it is unclear how many free parameters these modules introduce, how they are trained, and whether overfitting is mitigated via cross-validation or held-out test sets. This is a standard but important concern for multi-module empirical ML papers.
minor comments (2)
  1. The abstract would benefit from including at least one headline quantitative result (e.g., percentage improvement over a baseline) to convey the practical significance of the framework.
  2. The claim that retrieval matches patches 'consistent with the current window in temporal shape, power level, PV operating state, and intra-day period' should be substantiated with a retrieval quality metric or qualitative examples in the full paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detected: abstract-only review with no derivation chain to audit

full rationale

This is an abstract-only review. The abstract describes an empirical machine learning framework (PARA-PV) for PV forecasting with multiple components: physics-aware retrieval, a frozen Chronos foundation model prior with residual adapter, gated distribution shift correction, and a regime-aware loss. No equations, derivations, or formal proofs are presented in the available material, so there is no derivation chain to walk for circularity. The frozen Chronos model is externally pre-trained, providing independent grounding rather than self-citation. The distribution shift correction parameters may be fitted to data, but without the full text we cannot exhibit a specific reduction where a prediction equals its input by construction. The reader's concern about fitted correction parameters is a valid correctness/overfitting risk, but it is not demonstrable circularity from the abstract alone. No self-citation chain is visible. This is a standard empirical ML paper description with no identifiable circular reasoning in the available text.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities, particles, forces, or dimensions. It is an engineering framework combining existing components. The free parameters are the trainable components of the architecture (residual adapter, gated correction, loss reweighting). The axioms are domain assumptions about PV predictability and the transferability of foundation model knowledge, plus one ad-hoc design choice for the loss function. Without the full text, the full parameter count and detailed axiom list cannot be exhaustively enumerated.

free parameters (3)
  • Gated mean-shift and scale correction parameters
    The abstract describes 'gated mean-shift and scale corrections' in the distribution shift correction module. These gates and correction terms are almost certainly fitted to data, introducing free parameters that adjust the forecast. Without the full text, the exact number and values are unknown.
  • Regime-aware loss reweighting coefficients
    The physics-constrained loss 'adaptively reweights' error contributions across peak, ramping, night-time, and regular regimes. The reweighting scheme or its adaptation parameters are likely fitted or tuned to the data.
  • Residual adapter parameters
    The 'lightweight residual adapter' that calibrates the base forecast against the frozen Chronos model contains trainable parameters fitted to the PV data.
assumptions (4)
  • domain assumption Historical patches matched on temporal shape, power level, PV operating state, and intra-day period are predictive of future PV trajectories.
    The retrieval-augmented learner assumes that analog trajectories retrieved based on these criteria provide a physically grounded base forecast. This is a domain assumption about the predictability of PV dynamics from matched historical patterns.
  • domain assumption A frozen time-series foundation model (Chronos) captures general temporal regularities that are useful for PV-specific forecasting when adapted with a residual adapter.
    The framework assumes the Chronos prior contains transferable temporal knowledge and that a lightweight residual adapter is sufficient to adapt it to PV dynamics without overriding the physics-grounded prediction.
  • domain assumption Residual conditional distribution shifts after base forecasting and foundation model calibration can be corrected by gated mean-shift and scale corrections conditioned on power, weather, timestamp, and day/night features.
    The distribution shift correction module assumes that the remaining shifts are systematic and can be captured by gated affine transformations of the listed conditioning variables.
  • ad hoc to paper Partitioning samples into peak, ramping, night-time, and regular regimes and adaptively reweighting their error contributions improves forecasting of operationally critical states without degrading overall performance.
    The regime partition and adaptive reweighting in the loss function is a design choice specific to this paper. The assumption that this specific partitioning and reweighting strategy is optimal or even beneficial is not established by prior theory in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PARA-PV: Physics-Aware Retrieval-Augmented PV Prediction Based on Frozen Foundation Model and Distribution Shift Correction." pith.science (2026). https://pith.science/paper/LUG7HSJP

@misc{pith2026260708079,
  author       = {Pith},
  title        = {Pith review of: PARA-PV: Physics-Aware Retrieval-Augmented PV Prediction Based on Frozen Foundation Model and Distribution Shift Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LUG7HSJP}},
  note         = {Machine review of arXiv:2607.08079}
}
read the original abstract

Accurate photovoltaic (PV) power forecasting is essential for reliable grid dispatch and renewable energy integration, yet it remains challenging because PV generation is jointly shaped by weather variability, day-night transitions, regime-dependent dynamics, and strict physical constraints. We propose PARA-PV, a Physics-Aware Retrieval-Augmented framework that embeds physical knowledge throughout the forecasting process. The framework first encodes multivariate PV observations into patch-level representations and, through a physics-aware retrieval-augmented learner, retrieves historical patches and analog trajectories that are consistent with the current window in temporal shape, power level, PV operating state, and intra-day period; this yields a physically grounded base forecast. To supplement local memory with broader temporal knowledge, the base forecast is then calibrated against a frozen Chronos time-series foundation-model prior through a lightweight residual adapter, so that general temporal regularities are adapted to PV-specific dynamics without overriding the physically grounded prediction. Because residual conditional distribution shifts persist when weather and diurnal regimes change, a physics-aware distribution shift correction module subsequently adjusts the preliminary forecast using power, weather, timestamp, and day/night conditions, applying gated mean-shift and scale corrections selectively. Finally, a physics-constrained loss function partitions the samples into peak, ramping, night-time, and regular regimes and adaptively reweights their error contributions, preventing the dominant regular regime from suppressing learning of operationally critical states. Our code is available at https://github.com/weican1103/PARA-PV.

Figures

Figures reproduced from arXiv: 2607.08079 by the authors.

Figure 1
Figure 1. Architecture of the overall framework. 3.2.1. Patch representation and physics-aware memory bank Patch embedding: Given a normalized multivariate PV input sequence Xt ∈ R B×L×C, where B, L, and C denote the batch size, sequence length, and number of variables, respectively. PA-RAL first divides each channel-wise time series into overlapping patches of length P with stride S. Each patch is linearly projected into a d… view at source ↗
Figure 2
Figure 2. Architecture of the physics-aware retrieval-augmented learner module. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the Physics-aware distribution shift correction. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The results of point forecasting in Data 1 (50 MW). [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: The results of point forecasting in Data 2 (35 MW). [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: MAE variation across prediction horizons in the ablation study. [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: Forecasting performance versus model complexity on Data1 at the 96-Step horizon. [PITH_FULL_IMAGE:figures/full_fig_p029_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages

  1. [1]

    H. Fan, W. Liu, Z. Zhang, Y. Lu, W. Run, D. Liu, IDS-Net: A novel framework for few-shot photovoltaic power prediction with interpretable dynamic selection and feature information fusion, Advanced Engineering Informatics 72 (2026) 104455

    work page 2026

  2. [2]

    Huang, B

    J. Huang, B. Shao, Y. Ke, Y. Gao, S. Mazzoni, PV-MM-diffusion: An end-to-end multi-modal diffusion model for ultra-short-term probabilistic photovoltaic forecasting, Applied Energy 413 (2026) 127816

    work page 2026

  3. [3]

    C. Xia, Y. Xu, N. Tai, H. Sun, A dual-layer feature-selection transformer network for transferable probabilistic forecasting of PV power, Applied Energy 406 (2026) 127331

    work page 2026

  4. [4]

    Z. Tian, Y. Chen, G. Wang, Enhancing PV power forecasting accuracy through nonlinear weather correction based on multi-task learning, Applied Energy 386 (2025) 125525

    work page 2025

  5. [5]

    T.Wang, Y.Xu, Y.Qin, X.Wang, F.Zheng, W.Li, Short-termPVforecastingofmultiplescenarios based on multi-dimensional clustering and hybrid transformer-BiLSTM with ECPO, Energy 334 (2025) 137654

    work page 2025

  6. [6]

    Huang, T

    X. Huang, T. Zhao, B. Huang, Z. Zhang, M. Yue, Advancing energy system optimization via data-centric task-oriented forecasting: An application in PV-battery operation, Applied Energy 378 (2025) 124753

    work page 2025

  7. [7]

    Y.Yang,Y.Liu,Y.Zhang,S.Shu,J.Zheng,DEST-GNN:Adouble-exploredspatio-temporalgraph neural network for multi-site intra-hour PV power forecasting, Applied Energy 378 (2025) 124744

    work page 2025

  8. [8]

    Y. Zhi, T. Sun, X. Yang, A physical model with meteorological forecasting for hourly rooftop photovoltaic power prediction, Journal of Building Engineering 75 (2023) 106997

    work page 2023

Show all 39 references
  1. [9]

    Wolff, J

    B. Wolff, J. Kühnert, E. Lorenz, O. Kramer, D. Heinemann, Comparing support vector regression for PV power forecasting to a physical modeling approach using measurement, numerical weather prediction, and cloud motion data, Solar Energy 135 (2016) 197–208

    2016

  2. [10]

    Y. Li, Y. Su, L. Shu, An ARMAX model for forecasting the power output of a grid connected photovoltaic system, Renewable Energy 66 (2014) 78–89

    2014

  3. [11]

    A. A. H. Lateko, H.-T. Yang, C.-M. Huang, Short-term PV power forecasting using a regression- based ensemble method, Energies 15 (11) (2022) 4171

    2022

  4. [12]

    Doubleday, S

    K. Doubleday, S. Jascourt, W. Kleiber, B.-M. Hodge, Probabilistic solar power forecasting using bayesian model averaging, IEEE Transactions on Sustainable Energy 12 (1) (2021) 325–337

    2021

  5. [13]

    Ahmed, V

    R. Ahmed, V. Sreeram, Y. Mishra, M. Arif, A review and evaluation of the state-of-the-art in PV solar power forecasting: Techniques and optimization, Renewable and Sustainable Energy Reviews 124 (2020) 109792

    2020

  6. [14]

    U.K.Das,K.S.Tey,M.Seyedmahmoudian,M.Y.IdnaIdris,S.Mekhilef,B.Horan,A.Stojcevski, SVR-based model to forecast PV power generation under different weather conditions, Energies 10 (7) (2017) 876. 34

    2017

  7. [15]

    G.Xie,Z.Lin,RSMD-RF-BGSkipbasedPVgenerationpredictionmethod,IEEEAccess12(2024) 65799–65809

    2024

  8. [16]

    S. Wang, J. Ma, A novel GBDT-BiLSTM hybrid model on improving day-ahead photovoltaic prediction, Scientific Reports 13 (1) (2023) 15113

    2023

  9. [17]

    A. T. Eseye, J. Zhang, D. Zheng, Short-term photovoltaic solar power forecasting using a hybrid Wavelet-PSO-SVM model based on SCADA and Meteorological information, Renewable Energy 118 (2018) 357–367

    2018

  10. [18]

    Hategan, N

    S.-M. Hategan, N. Stefu, D. Petreus, E. Szilagyi, T. Patarau, M. Paulescu, Short-term forecasting of PV power based on aggregated machine learning and sky imagery approaches, Energy 316 (2025) 134595

    2025

  11. [19]

    W. Liu, M. Gai, PV-MLP: A lightweight patch-based multi-layer perceptron network with time– frequency domain fusion for accurate long-sequence photovoltaic power forecasting, Renewable Energy 251 (2025) 123277

    2025

  12. [20]

    H. K. Ahn, N. Park, Deep RNN-based photovoltaic power short-term forecast using power IoT sensors, Energies 14 (2) (2021) 436

    2021

  13. [21]

    X. Wang, M. Hu, X. Luo, X. Guan, Spatio-temporal photovolatic power forecasting via bayesian- optimizeddynamicgraphconvolutionalnetworkswithtemporalconvolutionalnetworks,Journalof Cleaner Production 542 (2026) 147641

    2026

  14. [22]

    Piantadosi, S

    G. Piantadosi, S. Dutto, A. Galli, S. De Vito, C. Sansone, G. Di Francia, Photovoltaic power forecasting: A Transformer based framework, Energy and AI 18 (2024) 100444

    2024

  15. [23]

    Q. Liu, Y. Li, H. Jiang, Y. Chen, J. Zhang, Short-term photovoltaic power forecasting based on multiplemodedecompositionandparallelbidirectionallongshorttermcombinedwithconvolutional neural networks, Energy 286 (2024) 129580

    2024

  16. [24]

    J.Kim,J.Obregon,H.Park,J.-Y.Jung,Multi-stepphotovoltaicpowerforecastingusingtransformer and recurrent neural networks, Renewable and Sustainable Energy Reviews 200 (2024) 114479

    2024

  17. [25]

    Z.Tian,B.Liang,PVMTF:End-to-endlong-sequencetime-seriesforecastingframeworksbasedon patchtechniqueandinformationfusioncodingformid-termphotovoltaicpowerforecasting,Applied Energy 396 (2025) 126263

    2025

  18. [26]

    Z. Liu, Y. Zhang, Y. Ning, Dual-physics-driven anomaly reconstruction and constrained deep learning for robust rooftop PV forecasting with defective data, Measurement 277 (2026) 121439

    2026

  19. [27]

    Y. Gao, Z. Hu, K. Yan, J. Otomo, N. Li, M. Liu, T. Xu, Y. Ruan, W. Gao, Reprogramming language models for energy: Few-shot solar forecasting across urban Japan, Advanced Engineering Informatics 74 (2026) 104765

    2026

  20. [28]

    H. He, R. Hong, D. Luo, S. Huang, X. Dai, Photovoltaic power forecasting based on large language models: A prompt-learning and two-stage fine-tuning approach, Electric Power Systems Research 252 (2026) 112461. 35

    2026

  21. [29]

    H.Dai, Z.Zhen, F.Wang, Y.Lin, F.Xu, N.Duić, Ashort-termPVpowerforecastingmethodbased on weather type credibility prediction and multi-model dynamic combination, Energy Conversion and Management 326 (2025) 119501

    2025

  22. [30]

    A. F. Ansari, O. Shchur, J. Küken, A. Auer, B. Han, P. Mercado, S. S. Rangapuram, H. Shen, L.Stella,X.Zhang,M.Goswami, S.Kapoor, D.C.Maddix,P.Guerron, T.Hu, J.Yin, N.Erickson, P.M.Desai,H.Wang,H.Rangwala, G.Karypis, Y.Wang,M.Bohlke-Schneider,Chronos-2: From univariate to unive...

    2025 arXiv

  23. [31]

    Y. Chen, J. Xu, Solar and wind power data from the chinese state grid renewable energy generation forecasting competition, Scientific Data 9 (2022) 577

    2022

  24. [32]

    Z. Tian, C. Wu, Enhancing long-sequence photovoltaic power forecasting accuracy through multi- modal learning: Integrating satellite cloud images and time–frequency domain fusion, Advanced Engineering Informatics 74 (2026) 104622

    2026

  25. [33]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Computation 9 (8) (1997) 1735– 1780

    1997

  26. [34]

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, W. Zhang, Informer: Beyond efficient transformer for long sequence time-series forecasting, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35, 2021, pp. 11106–11115

    2021

  27. [35]

    A. Zeng, M. Chen, L. Zhang, Q. Xu, Are transformers effective for time series forecasting?, in: Proceedings of the AAAI conference on Artificial Intelligence, Vol. 37, 2023, pp. 11121–11128

    2023

  28. [36]

    Y. Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, M. Long, itransformer: Inverted transformers are effectivefortimeseriesforecasting,in: InternationalConferenceonLearningRepresentations,Vol. 2024, 2024, pp. 11116–11140

    2024

  29. [37]

    H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, M. Long, TimesNet: Temporal 2d-variation modeling for general time series analysis, in: International Conference on Learning Representations, 2023

    2023

  30. [38]

    M. Jin, S. Wang, L. Ma, Z. Chu, J. Zhang, X. Shi, P.-Y. Chen, Y. Liang, Y.-F. Li, S. Pan, Q. Wen, Time-LLM: Time series forecasting by reprogramming large language models, in: International Conference on Learning Representations, 2024

    2024

  31. [39]

    Zhong, W

    S. Zhong, W. Ruan, M. Jin, H. Li, Q. Wen, Y. Liang, Time-VLM: Exploring multimodal vision- language models for augmented time series forecasting, in: International Conference on Machine Learning, 2025. 36

    2025

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.