arxiv: 2604.27840 · v2 · submitted 2026-04-30 · 💻 cs.LG · cs.AI

Recognition: unknown

CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting

Bokai Pan , Mingyue Cheng , Zhiding Liu , Shuo Yu , Xiaoyu Tao , Yuchong Wu , Qi Liu , Defu Lian

show 1 more author

Enhong Chen

Authors on Pith no claims yet

Pith reviewed 2026-05-07 05:41 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords time series forecastingagentic workflowslarge language modelsrole specializationsupervised fine-tuningreinforcement learningensemble forecastingiterative refinement

0 comments

The pith

CastFlow structures time series forecasting as a role-specialized agentic workflow that pairs a frozen general LLM for reasoning with a fine-tuned domain-specific LLM for numerical prediction to overcome single-pass limitations.

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

The paper argues that most LLM-based forecasting methods remain stuck in a static single-pass mapping from history to future values, which limits how deeply temporal patterns can be examined, how much context can be gathered in one go, and how much refinement or ensemble support is possible. CastFlow instead breaks the task into an agentic loop of planning, action, forecasting, and reflection, backed by a memory module and a multi-view toolkit that supplies diagnostic evidence and an ensemble baseline. A frozen general-purpose LLM handles the reasoning steps while a fine-tuned domain-specific LLM carries out evidence-guided numerical forecasting from that baseline rather than from scratch. The domain-specific model is trained in two stages—first supervised fine-tuning then reinforcement learning with verifiable rewards—so that the workflow can iteratively improve predictions. If this separation of roles works as intended, forecasting systems would gain the ability to adaptively extract patterns and refine outputs across many different datasets without retraining everything from scratch each time.

Core claim

CastFlow is a dynamic agentic forecasting framework that organizes the process into planning, action, forecasting, and reflection phases. The workflow is supported by a memory module that retrieves prior experience and a multi-view toolkit that constructs diagnostic evidence while supplying a reliable ensemble forecast baseline. Under a role-specialized design, a frozen general-purpose LLM preserves broad reasoning capabilities while a fine-tuned domain-specific LLM performs evidence-guided numerical forecasting from the ensemble baseline. Optimization of the domain-specific LLM uses a two-stage workflow-oriented training procedure that combines supervised fine-tuning with reinforcement学习. 受

What carries the argument

the role-specialized agentic workflow of planning, action, forecasting, and reflection phases, supported by memory retrieval and a multi-view toolkit that supplies ensemble baselines for a fine-tuned domain-specific LLM

If this is right

Multi-view temporal pattern extraction becomes feasible through the agentic structure and toolkit.
Multi-round contextual feature acquisition is supported by iterative memory and reflection steps.
Iterative forecast refinement occurs via the reflection phase acting on ensemble evidence.
Forecasting gains direct support from ensemble baselines rather than starting from raw history alone.
Superior overall performance is observed against strong baselines across diverse time series datasets.

Where Pith is reading between the lines

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

The same separation of general reasoning from specialized numerical work could be tested in other sequential prediction settings such as demand planning or sensor data streams.
Two-stage training that first aligns a model to workflow steps and then refines it with verifiable rewards may reduce the data needed for domain adaptation in other agentic systems.
Adding human feedback directly into the reflection phase would be a natural next test of whether the framework can move from fully automated to human-in-the-loop forecasting.
The memory module's retrieval of prior experience suggests the approach could accumulate cross-dataset knowledge over time without full retraining.

Load-bearing premise

The assumption that separating general reasoning in a frozen LLM from numerical forecasting in a fine-tuned LLM, optimized through SFT followed by RLVR, will produce reliably better temporal patterns and forecasts without introducing new failure modes or overfitting on the chosen datasets.

What would settle it

An ablation study or evaluation on fresh held-out datasets in which removing the role specialization or the two-stage training causes CastFlow to lose its performance advantage over strong single-pass LLM baselines or statistical methods would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.27840 by Bokai Pan, Defu Lian, Enhong Chen, Mingyue Cheng, Qi Liu, Shuo Yu, Xiaoyu Tao, Yuchong Wu, Zhiding Liu.

**Figure 1.** Figure 1: Comparison among training-free, fine-tuning, and agentic work view at source ↗

**Figure 2.** Figure 2: Overview of CastFlow. The framework orchestrates a planning-action-forecasting-reflection loop via a multi-view toolkit and a memory module. The view at source ↗

**Figure 3.** Figure 3: Effectiveness of the two-stage workflow-oriented training. We compare view at source ↗

**Figure 5.** Figure 5: Performance comparison of different local base models optimized view at source ↗

**Figure 7.** Figure 7: Qualitative comparison of forecasting trajectories across different optimization stages. The progression from the ensemble baseline to the fine-tuned view at source ↗

**Figure 8.** Figure 8: Evaluation of strategy memory mechanisms. (a) Impact of memory update strategies, highlighting the precision of the append approach over merging. view at source ↗

**Figure 9.** Figure 9: Case study of CastFlow on the BE dataset. CastFlow uses memory to orchestrate the multi-view toolkit and establish an ensemble forecast baseline. view at source ↗

**Figure 10.** Figure 10: Optimization process under GRPO. Rising reward and bounded view at source ↗

**Figure 11.** Figure 11: Tool usage heatmap across diverse datasets, showing the activation view at source ↗

read the original abstract

Recently, large language models (LLMs) have shown great promise in time series forecasting. However, most existing LLM-based forecasting methods still follow a static generative paradigm that directly maps historical observations to future values in a single pass. Under this paradigm, forecasting is constrained by limited temporal pattern extraction, single-round acquisition of contextual features, one-shot forecast generation, and lack of support from ensemble forecasts. To address these limitations, in this work, we propose CastFlow, a dynamic agentic forecasting framework that enables multi-view temporal pattern extraction, multi-round contextual features acquisition, iterative forecast refinement, and forecasting with ensemble forecasts. First, CastFlow organizes the forecasting process into planning, action, forecasting, and reflection, establishing an agentic workflow. Second, this workflow is supported by a memory module that retrieves prior experience and a multi-view toolkit that constructs diagnostic evidence and provides a reliable ensemble forecast baseline. Third, CastFlow adopts a role-specialized design that combines general-purpose reasoning with specialized numerical forecasting. Under this design, a frozen LLM preserves general-purpose reasoning, while a fine-tuned domain-specific LLM performs evidence-guided numerical forecasting based on the ensemble forecast baseline, rather than from scratch. To optimize a fine-tuned domain-specific LLM, we further develop a two-stage workflow-oriented training that combines supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR). To evaluate the effectiveness of CastFlow, we conduct extensive experiments on diverse datasets and show that it achieves superior overall results against strong baselines. We hope that this work can serve as a step toward more adaptive and accurate time series forecasting.

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

CastFlow introduces an agentic workflow with role-specialized LLMs for time series forecasting, but missing ablations on the fine-tuning leave its performance edge unproven.

read the letter

The main thing here is that CastFlow puts together a four-stage agentic workflow for time series forecasting with LLMs, using role specialization and a two-stage training process, but the experiments don't include the ablations that would show whether the fine-tuning actually drives the gains. They organize the process into planning, action, forecasting, and reflection, with a memory module to pull in past experience and a multi-view toolkit that builds diagnostic views and supplies an ensemble baseline. The frozen general LLM handles the reasoning side while the fine-tuned domain-specific one refines the forecast from that baseline. They train it first with supervised fine-tuning and then reinforcement learning with verifiable rewards. This setup aims to fix the limits of single-pass generation like poor temporal pattern handling and no iterative improvement. The paper does a solid job spelling out those limitations in current LLM forecasting and showing how the dynamic workflow and toolkit could help with multi-view extraction and ensemble support. The role split is a reasonable way to keep broad capabilities intact while targeting the numerical forecasting. Where it gets soft is the evidence. The abstract says it beats strong baselines on diverse datasets, but there's no breakdown showing that the SFT plus RLVR step adds something beyond the workflow and toolkit by themselves. An ablation that keeps the agentic structure the same and just swaps the forecasting LLM from frozen to fine-tuned would clarify if specialization helps or risks overfitting. Without that, the load-bearing claim stays untested. The lack of specific metrics or error bars in the summary also makes it tough to gauge how much better it really is. This would interest people working on agentic AI for forecasting or practical time series applications in business and science. A reader could pick up the workflow design and the training approach as ideas to adapt. I would send it for peer review. The core idea is coherent and the engineering looks thoughtful, so referees could help tighten the experimental side.

Referee Report

2 major / 3 minor

Summary. The paper proposes CastFlow, a dynamic agentic framework for time series forecasting that structures the process into planning-action-forecasting-reflection stages. It incorporates a memory module for prior experience retrieval and a multi-view toolkit for diagnostic evidence and ensemble forecast baselines. The framework uses a role-specialized design pairing a frozen general-purpose LLM for reasoning with a fine-tuned domain-specific LLM for evidence-guided numerical forecasting (starting from the ensemble baseline rather than scratch). Optimization employs a two-stage workflow-oriented training combining supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR). Extensive experiments on diverse datasets are reported to demonstrate superior overall performance against strong baselines by enabling multi-view temporal pattern extraction, multi-round contextual acquisition, iterative refinement, and ensemble-supported forecasting.

Significance. If the experimental claims are substantiated with proper controls, CastFlow could advance LLM-based time series forecasting by shifting from static single-pass generation to iterative, role-specialized agentic workflows that integrate ensemble baselines and reflection. The two-stage SFT+RLVR training for domain-specific numerical forecasting and the multi-view toolkit represent potentially reusable engineering contributions for adaptive forecasting systems, provided ablations confirm they add value beyond existing ensemble and prompting techniques.

major comments (2)

[Experiments (ablation studies)] The central claim—that CastFlow achieves superior results via the role-specialized design and two-stage training—rests on the assumption that fine-tuning the domain-specific LLM (via SFT then RLVR) improves upon the ensemble forecast baseline without introducing new failure modes. However, the experiments do not include ablations that hold the agentic workflow, memory module, and multi-view toolkit fixed while varying only the forecasting LLM (frozen general-purpose vs. fine-tuned domain-specific). Without this isolation, it remains unclear whether reported gains derive from the toolkit/ensemble alone or from the specialization and RLVR stage. This directly affects the load-bearing claim in the abstract and method description.
[Abstract and §4 Experiments] The abstract asserts superior overall results against strong baselines on diverse datasets, yet provides no quantitative metrics, error bars, dataset statistics, or specific baseline comparisons. The full experimental section must report these (including per-dataset MAE/RMSE, statistical significance, and ablation tables) to allow verification of the multi-view extraction, iterative refinement, and ensemble benefits; otherwise the soundness of the headline result cannot be assessed.

minor comments (3)

[Method (training subsection)] The description of the RLVR stage should explicitly define the verifiable reward function (e.g., how forecast accuracy is quantified and verified against ground truth) and the policy optimization details, as these are central to reproducing the two-stage training.
[Method] Clarify the exact composition of the multi-view toolkit (which views are constructed and how they feed into the ensemble baseline) to distinguish its contribution from prior ensemble methods in time series literature.
[Discussion or Experiments] Add a limitations or failure-mode analysis section discussing cases where the reflection stage or fine-tuned LLM may degrade performance relative to the frozen baseline.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the major comments point by point below, indicating planned revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Experiments (ablation studies)] The central claim—that CastFlow achieves superior results via the role-specialized design and two-stage training—rests on the assumption that fine-tuning the domain-specific LLM (via SFT then RLVR) improves upon the ensemble forecast baseline without introducing new failure modes. However, the experiments do not include ablations that hold the agentic workflow, memory module, and multi-view toolkit fixed while varying only the forecasting LLM (frozen general-purpose vs. fine-tuned domain-specific). Without this isolation, it remains unclear whether reported gains derive from the toolkit/ensemble alone or from the specialization and RLVR stage. This directly affects the load-bearing claim in the abstract and method description.

Authors: We agree this specific ablation is necessary to isolate the contribution of the fine-tuned domain-specific LLM. The manuscript includes ablations on workflow stages, memory, and toolkit, but not this exact comparison. We will add it in revision by evaluating a controlled variant that replaces only the forecasting LLM with the frozen general-purpose model while fixing the agentic workflow, memory retrieval, and multi-view toolkit. This will clarify whether gains stem from specialization and RLVR or from the ensemble and toolkit alone. revision: yes
Referee: [Abstract and §4 Experiments] The abstract asserts superior overall results against strong baselines on diverse datasets, yet provides no quantitative metrics, error bars, dataset statistics, or specific baseline comparisons. The full experimental section must report these (including per-dataset MAE/RMSE, statistical significance, and ablation tables) to allow verification of the multi-view extraction, iterative refinement, and ensemble benefits; otherwise the soundness of the headline result cannot be assessed.

Authors: We agree the abstract should include key quantitative highlights. We will revise it to report average MAE/RMSE improvements and dataset diversity. The experimental section already presents per-dataset MAE/RMSE results and baseline comparisons across multiple datasets, along with ablation tables. To fully address the request, we will add error bars, statistical significance tests (e.g., paired t-tests), explicit dataset statistics, and expanded discussion of how results verify multi-view extraction, iterative refinement, and ensemble benefits. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical framework validated by external experiments

full rationale

The paper presents CastFlow as an engineering framework for LLM-based time series forecasting, organized around an agentic workflow (planning/action/forecasting/reflection), memory module, multi-view toolkit, and role-specialized LLMs trained via SFT then RLVR. All central claims of superiority are grounded in reported experiments on diverse datasets against baselines, not in any mathematical derivation, equation, or first-principles result that reduces to its own inputs by construction. No self-definitional steps, fitted inputs relabeled as predictions, or load-bearing self-citations appear in the derivation chain; the contribution is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on domain assumptions about LLM capabilities for reasoning versus numerical tasks and on the empirical effectiveness of the newly introduced workflow components; no explicit free parameters are named in the abstract, and the invented workflow elements lack independent falsifiable handles beyond the reported experiments.

axioms (1)

domain assumption A frozen general-purpose LLM preserves reliable reasoning while a separate fine-tuned LLM can perform evidence-guided numerical forecasting
Invoked in the role-specialized design section of the abstract to justify separating general reasoning from domain-specific forecasting.

invented entities (1)

CastFlow agentic workflow (planning-action-forecasting-reflection stages plus memory and multi-view toolkit) no independent evidence
purpose: To overcome the four listed limitations of static single-pass LLM forecasting
Newly proposed framework whose value is asserted via experiments rather than prior independent evidence.

pith-pipeline@v0.9.0 · 5616 in / 1399 out tokens · 38202 ms · 2026-05-07T05:41:33.300809+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

88 extracted references · 19 canonical work pages · 5 internal anchors

[1]

2025 iflytek renewable power forecasting challenge (wind and solar),

iFLYTEK AI Challenge, “2025 iflytek renewable power forecasting challenge (wind and solar),” https://challenge.xfyun.cn/topic/info?type= renewable-power-forecast&option=ssgy&ch=dwsf259, 2025, accessed: Apr. 30, 2026

2025
[2]

US MOPEX data set,

J. Schaake, S. Cong, and Q. Duan, “US MOPEX data set,” Lawrence Livermore National Laboratory, Livermore, CA, USA, Tech. Rep., 2006

2006
[3]

A comprehensive survey of time series forecasting: Concepts, challenges, and future directions,

M. Cheng, Z. Liu, X. Tao, Q. Liu, J. Zhang, T. Pan, S. Zhang, P. He, X. Zhang, D. Wanget al., “A comprehensive survey of time series forecasting: Concepts, challenges, and future directions,”TechRxiv,
[4]

Available: https://doi.org/10.36227/techrxiv.174430535

[Online]. Available: https://doi.org/10.36227/techrxiv.174430535. 53879341/v1

work page doi:10.36227/techrxiv.174430535
[5]

TFB: Towards comprehensive and fair benchmarking of time series forecasting methods,

X. Qiu, J. Hu, L. Zhou, X. Wu, J. Du, B. Zhang, C. Guo, A. Zhou, C. S. Jensen, Z. Sheng, and B. Yang, “TFB: Towards comprehensive and fair benchmarking of time series forecasting methods,”Proceedings of the VLDB Endowment, vol. 17, no. 9, pp. 2363–2377, 2024

2024
[6]

Diffusion convolutional recurrent neural network: Data-driven traffic forecasting,

Y . Li, R. Yu, C. Shahabi, and Y . Liu, “Diffusion convolutional recurrent neural network: Data-driven traffic forecasting,” inInternational Conference on Learning Representations, 2018. [Online]. Available: https://openreview.net/forum?id=SJiHXGW AZ

2018
[7]

Multivariate time-series representation learning via hierarchical correlation pooling boosted graph neural network,

Y . Wang, M. Wu, X. Li, L. Xie, and Z. Chen, “Multivariate time-series representation learning via hierarchical correlation pooling boosted graph neural network,”IEEE Transactions on Artificial Intelligence, vol. 5, no. 1, pp. 321–333, 2024

2024
[8]

Deep Time Series Models: A Comprehensive Survey and Benchmark

Y . Wang, H. Wu, J. Dong, Y . Liu, M. Long, and J. Wang, “Deep time series models: A comprehensive survey and benchmark,” arXiv preprint arXiv:2407.13278, 2024. [Online]. Available: https: //arxiv.org/abs/2407.13278

work page internal anchor Pith review Pith/arXiv arXiv 2024
[9]

Deep learning for time series forecasting: Tutorial and literature survey,

K. Benidis, S. S. Rangapuram, V . Flunkert, Y . Wang, D. C. Maddix, C. Turkmen, J. Gasthaus, M. Bohlke-Schneider, D. Salinas, L. Stella et al., “Deep learning for time series forecasting: Tutorial and literature survey,”ACM Computing Surveys, vol. 55, no. 6, pp. 1–36, 2022

2022
[10]

Automatic time series forecasting: The forecast package for R,

R. J. Hyndman and Y . Khandakar, “Automatic time series forecasting: The forecast package for R,”Journal of Statistical Software, vol. 27, no. 3, pp. 1–22, 2008

2008
[11]

Exponential smoothing: The state of the art,

J. Gardner, Everette S., “Exponential smoothing: The state of the art,” Journal of Forecasting, vol. 4, no. 1, pp. 1–28, 1985

1985
[12]

Time series prediction using support vector machines: A survey,

N. I. Sapankevych and R. Sankar, “Time series prediction using support vector machines: A survey,”IEEE Computational Intelligence Magazine, vol. 4, no. 2, pp. 24–38, 2009. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 16

2009
[13]

XGBoost: A scalable tree boosting system,

T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 785–794

2016
[14]

LightGBM: A highly efficient gradient boosting decision tree,

G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y . Liu, “LightGBM: A highly efficient gradient boosting decision tree,” in Advances in Neural Information Processing Systems, vol. 30, 2017

2017
[15]

Machine learning strategies for time series forecasting,

G. Bontempi, S. Ben Taieb, and Y .-A. Le Borgne, “Machine learning strategies for time series forecasting,” inBusiness Intelligence: Second European Summer School, eBISS 2012, Brussels, Belgium, July 15– 21, 2012, Tutorial Lectures, ser. Lecture Notes in Business Information Processing. Springer, 2013, vol. 138, pp. 62–77

2012
[16]

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

H. Wu, J. Xu, J. Wang, and M. Long, “Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting,” in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 22 419–22 430

2021
[17]

Are transformers effective for time series forecasting?

A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transformers effective for time series forecasting?” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 9, 2023, pp. 11 121–11 128

2023
[18]

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

Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” in International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=Jbdc0vTOcol

2023
[19]

Chronos: Learning the language of time series,

A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapuram, S. P. Arango, S. Kapoor, J. Zschiegner, D. C. Maddix, H. Wang, M. W. Mahoney, K. Torkkola, A. G. Wilson, M. Bohlke-Schneider, and Y . Wang, “Chronos: Learning the language of time series,”Transactions on Machine Learning Research, 2024. [Online]. Available: ht...

2024
[20]

A decoder-only foundation model for time-series forecasting,

A. Das, W. Kong, R. Sen, and Y . Zhou, “A decoder-only foundation model for time-series forecasting,” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 235. PMLR, 2024, pp. 10 148–10 167. [Online]. Available: https://proceedings.mlr.press/v235/das24c.html

2024
[21]

Sundial: A family of highly capable time series foundation models,

Y . Liu, G. Qin, Z. Shi, Z. Chen, C. Yang, X. Huang, J. Wang, and M. Long, “Sundial: A family of highly capable time series foundation models,” inProceedings of the 42nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 267. PMLR, 2025, pp. 39 295–39 317. [Online]. Available: https://proceedings.mlr.press/v2...

2025
[22]

PromptCast: A new prompt-based learning paradigm for time series forecasting,

H. Xue and F. D. Salim, “PromptCast: A new prompt-based learning paradigm for time series forecasting,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 11, pp. 6851–6864, 2024

2024
[23]

TEMPO: Prompt-based generative pre-trained transformer for time series forecasting,

D. Cao, F. Jia, S. O. Arik, T. Pfister, Y . Zheng, W. Ye, and Y . Liu, “TEMPO: Prompt-based generative pre-trained transformer for time series forecasting,” inInternational Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/ forum?id=YH5w12OUuU

2024
[24]

Can slow-thinking LLMs reason over time? empirical studies in time series forecasting,

M. Cheng, J. Wang, D. Wang, X. Tao, Q. Liu, and E. Chen, “Can slow-thinking LLMs reason over time? empirical studies in time series forecasting,” inProceedings of the Nineteenth ACM International Con- ference on Web Search and Data Mining, 2026, pp. 99–110

2026
[25]

InstructTime++: Time series classification with multimodal language modeling via implicit feature enhancement,

M. Cheng, X. Tao, H. Zhang, Q. Liu, and E. Chen, “InstructTime++: Time series classification with multimodal language modeling via implicit feature enhancement,”arXiv preprint arXiv:2601.14968, 2026. [Online]. Available: https://arxiv.org/abs/2601.14968

work page arXiv 2026
[26]

GPT4MTS: Prompt- based large language model for multimodal time-series forecasting,

F. Jia, K. Wang, Y . Zheng, D. Cao, and Y . Liu, “GPT4MTS: Prompt- based large language model for multimodal time-series forecasting,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 21, 2024, pp. 23 343–23 351

2024
[27]

TimeSeriesScientist: A general-purpose AI agent for time series analysis,

H. Zhao, X. Zhang, J. Wei, Y . Xu, Y . He, S. Sun, and C. You, “TimeSeriesScientist: A general-purpose AI agent for time series analysis,”arXiv preprint arXiv:2510.01538, 2025. [Online]. Available: https://arxiv.org/abs/2510.01538

work page arXiv 2025
[28]

Position: Beyond model-centric prediction – agentic time series forecasting,

M. Cheng, X. Tao, Q. Liu, Z. Guo, and E. Chen, “Position: Beyond model-centric prediction – agentic time series forecasting,” arXiv preprint arXiv:2602.01776, 2026. [Online]. Available: https: //arxiv.org/abs/2602.01776

work page arXiv 2026
[29]

Time-LLM: Time series forecasting by reprogramming large language models,

M. Jin, S. Wang, L. Ma, Z. Chu, J. Y . Zhang, X. Shi, P.-Y . Chen, Y . Liang, Y .-F. Li, S. Pan, and Q. Wen, “Time-LLM: Time series forecasting by reprogramming large language models,” inInternational Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/forum?id=Unb5CVPtae

2024
[30]

S2IP-LLM: Semantic space informed prompt learning with LLM for time series forecasting,

Z. Pan, Y . Jiang, S. Garg, A. Schneider, Y . Nevmyvaka, and D. Song, “S2IP-LLM: Semantic space informed prompt learning with LLM for time series forecasting,” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 235. PMLR, 2024, pp. 39 135–39 153. [Online]. Available: https://proceedi...

2024
[31]

Time Series Forecasting as Reasoning: A Slow-Thinking Approach with Reinforced LLMs

Y . Zhou, Y . Luo, M. Cheng, Q. Liu, J. Wang, D. Wang, and E. Chen, “Time series forecasting as reasoning: A slow-thinking approach with reinforced LLMs,”arXiv preprint arXiv:2506.10630, 2025. [Online]. Available: https://arxiv.org/abs/2506.10630

work page internal anchor Pith review Pith/arXiv arXiv 2025
[32]

Revisiting catastrophic forgetting in large language model tuning,

H. Li, L. Ding, M. Fang, and D. Tao, “Revisiting catastrophic forgetting in large language model tuning,” inFindings of the Association for Computational Linguistics: EMNLP 2024. Miami, Florida, USA: Association for Computational Linguistics, 11 2024, pp. 4297–4308. [Online]. Available: https://aclanthology.org/2024.findings-emnlp.249/

2024
[33]

ReAct: Synergizing reasoning and acting in language models,

S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y . Cao, “ReAct: Synergizing reasoning and acting in language models,” in International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=WE vluYUL-X

2023
[34]

AnomaMind: Agentic time series anomaly detection with tool-augmented reasoning,

X. Tao, Y . Wu, M. Cheng, Z. Guo, and T. Gao, “AnomaMind: Agentic time series anomaly detection with tool-augmented reasoning,” arXiv preprint arXiv:2602.13807, 2026. [Online]. Available: https: //arxiv.org/abs/2602.13807

work page arXiv 2026
[35]

Improving time series forecasting via instance-aware post-hoc revision,

Z. Liu, M. Cheng, G. Zhao, J. Yang, Q. Liu, and E. Chen, “Improving time series forecasting via instance-aware post-hoc revision,”arXiv preprint arXiv:2505.23583, 2025. [Online]. Available: https://arxiv.org/abs/2505.23583

work page arXiv 2025
[36]

AlphaCast: A Human Wisdom-LLM Intelligence Co-Reasoning Framework for Interactive Time Series Forecasting

X. Zhang, T. Gao, M. Cheng, B. Pan, Z. Guo, Y . Liu, and X. Tao, “AlphaCast: A human wisdom-LLM intelligence co-reasoning framework for interactive time series forecasting,”arXiv preprint arXiv:2511.08947, 2025. [Online]. Available: https://arxiv.org/abs/2511. 08947

work page internal anchor Pith review Pith/arXiv arXiv 2025
[37]

Cast- R1: Learning tool-augmented sequential decision policies for time series forecasting,

X. Tao, M. Cheng, C. Jiang, T. Gao, H. Zhang, and Y . Liu, “Cast- R1: Learning tool-augmented sequential decision policies for time series forecasting,”arXiv preprint arXiv:2602.13802, 2026. [Online]. Available: https://arxiv.org/abs/2602.13802

work page arXiv 2026
[38]

Time-Series Representation Learning via Temporal and Contextual Contrasting

E. Eldele, M. Ragab, Z. Chen, M. Wu, C. K. Kwoh, X. Li, and C. Guan, “Time-series representation learning via temporal and contextual contrasting,” inProceedings of the Thirtieth International Joint Conference on Artificial Intelligence. International Joint Conferences on Artificial Intelligence Organization, 2021, pp. 2352–2359, main Track. [Online]. Ava...

work page doi:10.24963/ijcai.2021/324 2021
[39]

HyperIMTS: Hypergraph neural network for irregular multivariate time series forecasting,

B. Li, Y . Luo, Z. Liu, J. Zheng, J. Lv, and Q. Ma, “HyperIMTS: Hypergraph neural network for irregular multivariate time series forecasting,” inProceedings of the 42nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 267. PMLR, 2025, pp. 35 502–35 518. [Online]. Available: https://proceedings.mlr.press/v26...

2025
[40]

TimeBase: The power of minimalism in efficient long-term time series forecasting,

Q. Huang, Z. Zhou, K. Yang, Z. Yi, X. Wang, and Y . Wang, “TimeBase: The power of minimalism in efficient long-term time series forecasting,” inProceedings of the 42nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 267. PMLR, 2025, pp. 26 227–26 246. [Online]. Available: https://proceedings.mlr.press/v267...

2025
[41]

TSLANet: Rethinking transformers for time series representation learning,

E. Eldele, M. Ragab, Z. Chen, M. Wu, and X. Li, “TSLANet: Rethinking transformers for time series representation learning,” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol
[42]

12 409–12 428

PMLR, 2024, pp. 12 409–12 428. [Online]. Available: https: //proceedings.mlr.press/v235/eldele24a.html

2024
[43]

A closer look at transformers for time series forecasting: Understanding why they work and where they struggle,

Y . Chen, N. C ´espedes, and P. Barnaghi, “A closer look at transformers for time series forecasting: Understanding why they work and where they struggle,” inProceedings of the 42nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 267. PMLR, 2025, pp. 7763–7780. [Online]. Available: https://proceedings.mlr....

2025
[44]

FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting,

T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting,” inProceedings of the 39th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 162. PMLR, 2022, pp. 27 268–27 286. [Online]. Available: https://proceedings.mlr.press/v162...

2022
[45]

TimeMachine: A time series is worth 4 mambas for long-term forecasting,

M. A. Ahamed and Q. Cheng, “TimeMachine: A time series is worth 4 mambas for long-term forecasting,” inECAI 2024: 27th European Conference on Artificial Intelligence, Including 13th Conference on Prestigious Applications of Intelligent Systems, PAIS 2024, Proceedings, ser. Frontiers in Artificial Intelligence and Applications, vol. 392. IOS Press, 2024, p...

2024
[46]

N-HiTS: Neural hierarchical interpolation for time IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 17 series forecasting,

C. Challu, K. G. Olivares, B. N. Oreshkin, F. Garza, M. M. Canseco, and A. Dubrawski, “N-HiTS: Neural hierarchical interpolation for time IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 17 series forecasting,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 6, 2023, pp. 6989–6997

2023
[47]

ETSformer: Exponential smoothing transformers for time-series forecasting,

G. Woo, C. Liu, D. Sahoo, A. Kumar, and S. Hoi, “ETSformer: Exponential smoothing transformers for time-series forecasting,” in International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=5m 3whfo483

2023
[48]

iTransformer: Inverted transformers are effective for time series forecasting,

Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “iTransformer: Inverted transformers are effective for time series forecasting,” inInternational Conference on Learning Representations,
[49]

Available: https://openreview.net/forum?id=JePfAI8fah

[Online]. Available: https://openreview.net/forum?id=JePfAI8fah
[50]

ConvTimeNet: A deep hierarchical fully convolutional model for multivariate time series analysis,

M. Cheng, J. Yang, T. Pan, Q. Liu, Z. Li, and S. Wang, “ConvTimeNet: A deep hierarchical fully convolutional model for multivariate time series analysis,” inCompanion Proceedings of the ACM on Web Conference 2025, 2025, pp. 171–180

2025
[51]

A unifying perspective on model reuse: From small to large pre-trained models,

D.-W. Zhou and H.-J. Ye, “A unifying perspective on model reuse: From small to large pre-trained models,” inProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25, J. Kwok, Ed. International Joint Conferences on Artificial Intelligence Organization, 8 2025, pp. 10 826–10 835, survey Track. [Online]. Available:...

work page doi:10.24963/ijcai.2025/1201 2025
[52]

Model spider: Learning to rank pre-trained models efficiently,

Y .-K. Zhang, T.-J. Huang, Y .-X. Ding, D.-C. Zhan, and H.-J. Ye, “Model spider: Learning to rank pre-trained models efficiently,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 13 692–13 719. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/2023/hash/ 2c71b14637802ed08eaa3cf50342b2b9-Abstract-Conference.html

2023
[53]

One-embedding-fits-all: Efficient zero-shot time series forecasting by a model zoo,

H.-N. Shi, T.-J. Huang, L. Han, D.-C. Zhan, and H.-J. Ye, “One-embedding-fits-all: Efficient zero-shot time series forecasting by a model zoo,”arXiv preprint arXiv:2509.04208, 2025. [Online]. Available: https://arxiv.org/abs/2509.04208

work page arXiv 2025
[54]

From values to tokens: An LLM-driven framework for context-aware time series forecasting via symbolic discretization,

X. Tao, S. Zhang, M. Cheng, D. Wang, T. Pan, B. Pan, C. Zhang, and S. Wang, “From values to tokens: An LLM-driven framework for context-aware time series forecasting via symbolic discretization,”arXiv preprint arXiv:2508.09191, 2025. [Online]. Available: https://arxiv.org/abs/2508.09191

work page arXiv 2025
[55]

Large language models are zero-shot time series forecasters,

N. Gruver, M. Finzi, S. Qiu, and A. G. Wilson, “Large language models are zero-shot time series forecasters,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 19 622–19 635

2023
[56]

Toolformer: Language models can teach themselves to use tools,

T. Schick, J. Dwivedi-Yu, R. Dess `ı, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom, “Toolformer: Language models can teach themselves to use tools,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 68 539–68 551

2023
[57]

Describe, explain, plan and select: Interactive planning with LLMs enables open-world multi-task agents,

Z. Wang, S. Cai, G. Chen, A. Liu, X. Ma, and Y . Liang, “Describe, explain, plan and select: Interactive planning with LLMs enables open-world multi-task agents,” inAdvances in Neural Information Processing Systems, vol. 36, 2023. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/2023/hash/ 6b8dfb8c0c12e6fafc6c256cb08a5ca7-Abstract-Con...

2023
[58]

Self-refine: Iterative refinement with self-feedback,

A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yang, S. Gupta, B. P. Majumder, K. Hermann, S. Welleck, A. Yazdanbakhsh, and P. Clark, “Self-refine: Iterative refinement with self-feedback,” inAdvances in Neural Infor- mation Processing Systems, vol. 36, 2023, pp. 46 534–46 594

2023
[59]

Reflexion: Language agents with verbal reinforcement learning,

N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao, “Reflexion: Language agents with verbal reinforcement learning,” in Advances in Neural Information Processing Systems, vol. 36, 2023, pp. 8634–8652

2023
[60]

STaR: Bootstrapping rea- soning with reasoning,

E. Zelikman, Y . Wu, J. Mu, and N. Goodman, “STaR: Bootstrapping rea- soning with reasoning,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 15 476–15 488

2022
[61]

DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning,

D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, X. Zhang, X. Yu, Y . Wu, Z. F. Wu, Z. Gou, Z. Shao, Z. Li, Z. Gao, A. Liuet al., “DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning,”Nature, vol. 645, no. 8081, pp. 633– 638, 2025

2025
[62]

Qwen3 Technical Report

A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025. [Online]. Available: https://arxiv.org/abs/2505.09388

work page internal anchor Pith review arXiv 2025
[63]

Integration of Old and New Knowledge for Generalized Intent Discovery: A Consistency-driven Prototype-Prompting Framework

X. Zhang, R. R. Chowdhury, R. K. Gupta, and J. Shang, “Large language models for time series: A survey,” inProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, K. Larson, Ed. International Joint Conferences on Artificial Intelligence Organization, 8 2024, pp. 8335–8343, survey Track. [Online]. Available: htt...

work page doi:10.24963/ijcai.2024/921 2024
[64]

TimeOmni-1: Incentivizing complex reasoning with time series in large language models,

T. Guan, Z. Meng, D. Li, S. Wang, C.-H. H. Yang, Q. Wen, Z. Liu, S. M. Siniscalchi, M. Jin, and S. Pan, “TimeOmni-1: Incentivizing complex reasoning with time series in large language models,” in International Conference on Learning Representations, 2026. [Online]. Available: https://openreview.net/forum?id=kOIclg7muL

2026
[65]

AlphaAgentEvo: Evolution-oriented alpha mining via self-evolving agentic reinforcement learning,

Z. Tang, X. Yin, W. Chen, Z. Chen, Y . Zheng, W. Ye, K. Wang, and L. Lin, “AlphaAgentEvo: Evolution-oriented alpha mining via self-evolving agentic reinforcement learning,” inInternational Conference on Learning Representations, 2026. [Online]. Available: https://openreview.net/forum?id=lNmZrawUMu

2026
[66]

Empowering time series analysis with large language models: A survey,

Y . Jiang, Z. Pan, X. Zhang, S. Garg, A. Schneider, Y . Nevmyvaka, and D. Song, “Empowering time series analysis with large language models: A survey,” inProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24. International Joint Conferences on Artificial Intelligence Organization, Aug. 2024, pp. 8095–8103, sur...

work page doi:10.24963/ijcai.2024/895 2024
[67]

TimeXer: Empowering transformers for time series fore- casting with exogenous variables,

Y . Wang, H. Wu, J. Dong, G. Qin, H. Zhang, Y . Liu, Y . Qiu, J. Wang, and M. Long, “TimeXer: Empowering transformers for time series fore- casting with exogenous variables,” inAdvances in Neural Information Processing Systems, vol. 37, 2024, pp. 469–498

2024
[68]

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

H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

2021
[69]

Forecasting at scale,

S. J. Taylor and B. Letham, “Forecasting at scale,”The American Statistician, vol. 72, no. 1, pp. 37–45, 2018

2018
[70]

Grok 4 Model Card,

xAI, “Grok 4 Model Card,” xAI, Model Card, aug 2025. [Online]. Available: https://data.x.ai/2025-08-20-grok-4-model-card.pdf

2025
[71]

Transformers: State-of-the-art natural language processing,

T. Wolf, L. Debut, V . Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen, C. Ma, Y . Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush, “Transformers: State-of-the-art natural language processing,” inProceedings of the 2020 Conference on Empirical Met...

2020
[72]

Agent lightning: Train any ai agents with reinforcement learning,

X. Luo, Y . Zhang, Z. He, Z. Wang, S. Zhao, D. Li, L. K. Qiu, and Y . Yang, “Agent lightning: Train ANY AI agents with reinforcement learning,”arXiv preprint arXiv:2508.03680, 2025. [Online]. Available: https://arxiv.org/abs/2508.03680

work page arXiv 2025
[73]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . K. Li, Y . Wu, and D. Guo, “DeepSeekMath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024. [Online]. Available: https://arxiv.org/abs/2402.03300

work page internal anchor Pith review arXiv 2024
[74]

Reversible instance normalization for accurate time-series forecasting against distribution shift,

T. Kim, J. Kim, Y . Tae, C. Park, J.-H. Choi, and J. Choo, “Reversible instance normalization for accurate time-series forecasting against distribution shift,” inInternational Conference on Learning Representations, 2022. [Online]. Available: https://openreview.net/ forum?id=cGDAkQo1C0p Bokai Panis a senior undergraduate student at the University of Scien...

2022
[75]

Exponential and polynomial decay for a quasilinear viscoelastic equation,

S. A. Messaoudi and N.-e. Tatar, “Exponential and polynomial decay for a quasilinear viscoelastic equation,” Nonlinear Analysis: Theory, Methods & Applications, vol. 68, no. 4, pp. 785–793, 2008

2008
[76]

A tutorial on energy-based learning,

Y. Lecun, S. Chopra, and et al., “A tutorial on energy-based learning,” in Predicting Structured Data. MIT Press, 2006

2006
[77]

Adam: a method for stochastic optimiza- tion,

D. P . Kingma and J. Ba, “Adam: a method for stochastic optimiza- tion,” in Proceedings of the 3rd International Conference on Learning Representations, 2015, pp. 1–15

2015
[78]

Modeling dynamic missingness of implicit feedback for sequential recommendation,

X. Zheng, M. Wang, and et al., “Modeling dynamic missingness of implicit feedback for sequential recommendation,” IEEE Transac- tions on Knowledge & Data Engineering, no. 34, pp. 405–418, 2022

2022
[79]

Sliding window-based frequent pattern mining over data streams,

S. K. Tanbeer, C. F. Ahmed, and et al., “Sliding window-based frequent pattern mining over data streams,” Information Sciences, vol. 179, no. 22, pp. 3843–3865, 2009

2009
[80]

Incorporating rna-seq data into the zebrafish ensembl genebuild,

J. E. Collins and et al., “Incorporating rna-seq data into the zebrafish ensembl genebuild,” Genome Research, vol. 22, no. 10, pp. 2067–2078, 2012

2067

Showing first 80 references.