Pith. sign in

REVIEW 4 major objections 5 minor 27 references

Coupling a pre-trained LLM, reconfigured only by attention-topology changes, with an interval-averaged flow-matching decoder yields one-step forecasts that the paper reports consistently outperform eight baselines across ten time-series ben

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 04:53 UTC pith:DK76EOK5

load-bearing objection CoGenCast is a plausible hybrid LLM+flow forecaster with broad experiments, but the 'consistently outperforms' claim overreaches and the pretrained-weight attribution is untested because the random-init control is missing. the 4 major comments →

arxiv 2602.03564 v2 pith:DK76EOK5 submitted 2026-02-03 cs.LG

CoGenCast: A Coupled Autoregressive-Flow Generative Framework for Time Series Forecasting

Mingyue Cheng , Yaguo Liu , Daoyu Wang , Xiaoyu Tao , Qi Liu This is my paper
classification cs.LG
keywords time series forecastinggenerative modelingflow matchinglarge language modelsattention topologyone-step generationprobabilistic forecastingencoder-decoder
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

CoGenCast argues that time series forecasting needs both semantic understanding of context and stochastic modeling of continuous dynamics, and that neither pure LLM nor pure diffusion/flow approaches deliver both. The paper's proposal is a single hybrid: a pre-trained decoder-only LLM whose attention topology is changed to form a bidirectional encoder plus causal decoder with cross-attention, followed by a flow-matching denoising decoder that predicts the average velocity over an integration interval. Because the learned trajectories are near-straight, the model generates each future patch in one function evaluation. On ten public benchmarks the paper reports consistent error reductions over eight baselines — roughly 11% lower MSE than LLM-based methods and over 7% lower than transformer-based methods — plus gains from cross-domain training. The load-bearing premise is that the LLM's pretrained weights survive the attention reconfiguration, a premise the paper does not test against a randomly initialized control.

Core claim

The central claim is that the conditional forecasting distribution is best modeled by jointly learning an autoregressive semantic representation and a continuous stochastic transport, and that pre-trained LLM weights can be repurposed for this by modifying only the attention topology: the encoder uses bidirectional self-attention to fuse look-back patches and textual context, the decoder keeps causal self-attention and adds cross-attention, and a denoising decoder predicts a velocity field averaged over a time interval rather than an instantaneous velocity. A Jacobian-vector-product-corrected MSE loss drives the trajectories straight, so a single Euler step suffices at inference. The paper r

What carries the argument

The load-bearing mechanism is the attention-topology reconfiguration of a pre-trained decoder-only LLM: replacing causal self-attention with bidirectional self-attention in the encoder and adding cross-attention in the causal decoder, initialized entirely from the LLM's pretrained weights. The second mechanism is interval-conditioned average-velocity flow matching: instead of predicting instantaneous velocity, the denoising decoder predicts the average velocity over an interval, with a JVP-based loss term that penalizes curvature; this keeps the transport path nearly straight, enabling one-step (1-NFE) sampling.

Load-bearing premise

The pretrained LLM weights remain useful after swapping causal self-attention for bidirectional attention and inserting cross-attention; if they do not, the reported advantage over a same-scale randomly initialized transformer disappears, and this is not tested against a randomly initialized control.

What would settle it

Train the same encoder–decoder–flow architecture from scratch (random initialization, same 0.6B scale, same data) and compare MSE/MAE on the ten benchmarks; statistical parity with CoGenCast would falsify the claim that pretraining, rather than architecture or training procedure, drives the gains.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the reported gains hold, one-step forecasting with LLM-scale semantic conditioning becomes practical for latency-sensitive applications, since a single function evaluation replaces iterative denoising.
  • Cross-domain training consistently lowers MSE versus in-domain training, implying that a single CoGenCast model can serve multiple domains without per-domain fine-tuning.
  • Removing either the autoregressive LLM backbone or the flow decoder degrades accuracy, supporting the paper's dual-capability thesis.
  • LLM backbones outperforming a vanilla transformer at equal scale suggests pretrained language model weights, not just scale, contribute to forecasting accuracy — provided the transferability assumption holds.
  • Extending the look-back window from 96 to 336 steps yields progressive improvements, so the architecture can exploit longer history without structural changes.

Where Pith is reading between the lines

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

  • A decisive control the paper does not run: same reconfigured architecture with randomly initialized weights. If that control matches CoGenCast, the advantage would come from the architecture and training recipe rather than from pretrained weights.
  • The interval-averaging trick could transfer to other generative tasks — e.g., time-series imputation or multimodal generation — wherever a nearly straight probability path can be enforced by the JVP objective.
  • The attention-topology reconfiguration recipe suggests a general way to adapt decoder-only pretrained models to encoder–decoder tasks, which could be tested on text summarization or translation with minimal fine-tuning.
  • Because the paper's main table and appendix headers report different look-back lengths (L=96 vs L=336), a reader should confirm which setting produced the headline numbers before generalizing the comparison.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CoGenCast, a generative time-series forecasting framework that (i) reconfigures a pretrained decoder-only LLM (Qwen3-0.6B) into an encoder-decoder by changing causal self-attention to bidirectional self-attention in the encoder and adding cross-attention in the decoder, and (ii) couples this backbone with a flow-matching denoising decoder that predicts an interval-conditioned average velocity field. Training uses an autoregressive patch loss and a JVP-corrected flow-matching objective (Eq. 7); inference generates patches autoregressively and, the authors claim, samples each patch in one function evaluation (Eq. 8). Experiments on ten benchmarks compare against LLM-based, generative, and transformer baselines, with ablations on architecture, AR-flow components, context features, backbone scale, NFE, patch size, and noise schedule.

Significance. If the empirical results are robust, the proposed framework is a meaningful step: it is one of the few systems that combine LLM semantic conditioning with continuous stochastic generation while avoiding iterative sampling. The evaluation spans ten datasets, per-horizon results are provided in the appendix, and the code is released, all of which are strengths. The empirical claim is not circular, since the flow objective is a training loss rather than a quantity fitted to benchmark results. However, the manuscript overstates the consistency of the wins, and the central attribution of gains to pretrained LLM weights is not yet supported by the experimental design. The methodological novelty of the JVP-corrected average-velocity objective also needs a derivation or explicit reference. These issues are fixable but require additional experiments and analysis.

major comments (4)
  1. [§4.2, Table 2; abstract and contribution 3] The paper claims CoGenCast "consistently outperforms previous compared baselines," but Table 2 shows the opposite on ETTh2: CoGenCast MSE 0.159 vs TimeDART 0.145 and PatchTST 0.142 (MAE 0.246 vs 0.253/0.251). No standard deviations, seeds, or significance tests are provided anywhere, and many between-method gaps in Table 2 are small (e.g., Solar 0.221 vs 0.232; ETTm2 0.109 vs 0.122). Please add multi-seed statistics or revise the claim to "competitive on most benchmarks."
  2. [§3.3 and Table 6] The load-bearing claim that reusing pretrained Qwen3-0.6B weights is beneficial is not tested. Reconfiguring causal self-attention to bidirectional self-attention and adding cross-attention may destroy the value of the pretrained weights, but the only backbone comparisons are against a vanilla Transformer (varying pretraining, architecture, and parameter count together) or against larger Qwen variants that also use pretrained weights. A same-architecture randomly initialized control (same encoder-decoder plus flow decoder, same patching/text embedding scheme) is required to attribute the observed gains to pretraining rather than scale or architecture. Appendix D.1's "full 0.6B > encoder-only 1.7B" comparison does not provide this control.
  3. [§3.4, Eq. (8)] The one-step/low-latency claim is not consistent with the written equation. Eq. (8) defines yout_j as y_j(0) + ∫_0^1 u(zτ,τ,zdec_out)dτ, an integral over a continuum; evaluating this is not a single function evaluation. If the denoising decoder outputs the average velocity û_j over [0,1], the one-step output should be y_j(0) + û_j. Please either rewrite Eq. (8) accordingly or state explicitly that 1-NFE refers to a one-step Euler approximation of the integral, and report measured latency/NFE.
  4. [§3.3, Eq. (7)] The JVP-corrected objective is insufficiently justified. The regression target v_j − (r−t) ∂u_j/∂t depends on the model's own time-derivative, so the loss is not a fixed supervised target; the text does not prove that minimizers learn the average velocity over [t,r]. This is the main methodological novelty, and the cited mean-flow papers (Geng et al. 2025a,b) are not connected to the derivation. Please provide a derivation or reference, and include a small synthetic check of the estimated velocity/straightness.
minor comments (5)
  1. [Appendix I, Table 8] The header states L=336, while Section 4.1 and Table 2 use L=96. Please clarify which look-back setting produced the main results and ensure the appendix tables match the main-text averages.
  2. [§3.3–§3.4, Eqs. (6) and (8)] Notation is inconsistent: Eq. (6) uses hat z_j but Eq. (8) uses z_τ for the noised latent. Please unify.
  3. [Figure 3] The figure contains a typo ("knowdge" for "knowledge") and reports only MSE without error bars or raw values in the main text.
  4. [§4.6, Figure 6] The text calls the predictive intervals "well-calibrated," but no quantitative coverage or calibration metrics are reported. Figure 6 alone is qualitative and does not support that wording.
  5. [§3.3, Related Work] Mean flows and improved mean flows are cited but not discussed. Please articulate the difference between the proposed interval-conditioned objective and those methods.

Circularity Check

0 steps flagged

No significant circularity: benchmark results are self-contained; missing random-init control and borrowed mean-flow mechanism are attribution/novelty concerns, not input-output loops.

full rationale

The paper's central empirical claim is not circular. CoGenCast is trained by minimizing the flow-matching MSE in Eq. (7) on fixed train/test splits, and the 1-NFE inference in Eq. (8) is the corresponding Euler step of the learned velocity field. No parameter is fitted to the reported benchmark numbers, and the comparisons in Tables 2, 5, and 6 are against external or controlled internal baselines. The only self-referential element is the JVP-corrected target in Eq. (7), which includes the model's own time derivative; this is a fixed-point regularizer used during training, not an equivalence between a fitted quantity and a claimed prediction. The 'average velocity' mechanism is closely related to the cited mean-flow literature (Geng et al., 2025a,b), but those works are cited, and the framework's claimed contribution is the coupling of an LLM backbone with a flow-matching decoder. The main weaknesses are attribution gaps rather than circularity: Table 6 lacks a same-architecture randomly initialized control, so the 'pretrained LLM' advantage is not cleanly isolated from scale or architecture, and several self-citations (e.g., Cheng et al., 2025a,b in Sec. 3.1) are background rather than load-bearing. These concerns justify a low non-zero score, but no step in the derivation reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The framework's guarantee rests on unproven modeling assumptions: pretrained weights survive attention reconfiguration, straight-line flow represents the forecast distribution, and baseline comparisons are fair. The listed hyperparameters are manually chosen and affect the reported numbers.

free parameters (4)
  • Patch size P = 2-10; dataset-specific (e.g., 6 for Energy/ETTh1, 2 for Environ/Exchange)
    Chosen per dataset in Table 16; affects tokenization granularity and downstream performance, and the main results depend on this hand search.
  • Learning rate = 5e-4, 1e-4, or 5e-5
    Selected from a set in Appendix B.2; exact value per dataset is not reported.
  • Training epochs = 10
    Fixed at 10 for all datasets; no multiple-seed variation is reported.
  • Look-back window L = 96
    Default L=96; ablations to 192 and 336 show consistent improvements, so the main results may understate the model's capability.
axioms (4)
  • domain assumption Reconfiguring causal attention into bidirectional encoder attention plus cross-attention preserves the usefulness of Qwen3-0.6B pretrained weights.
    Invoked in §3.3; if false, the advantage over a randomly initialized same-scale transformer is not established.
  • domain assumption Straight-line interpolation and the average-velocity objective can faithfully represent the conditional forecast distribution P(Xpred|Xhist,C).
    The method's correctness rests on this modeling assumption; no theoretical guarantee is given.
  • ad hoc to paper The JVP-corrected loss in Eq. 7 yields a valid mean-velocity estimator.
    Introduced to enable one-step generation; no convergence or unbiasedness proof is provided in the paper.
  • domain assumption Official baseline implementations are evaluated fairly and comparably.
    Stated in Appendix B.2; the paper does not report baseline hyperparameters or runtime, so comparability is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 33403 in / 10440 out tokens · 108218 ms · 2026-08-03T04:53:54.731471+00:00 · methodology

0 comments
read the original abstract

Time series forecasting can be viewed as a generative problem that requires both semantic understanding over contextual conditions and stochastic modeling of continuous temporal dynamics. Existing approaches typically rely on either autoregressive large language models (LLMs) for semantic context modeling or diffusion-like models for continuous probabilistic generation. However, neither method alone can adequately model both aspects simultaneously. In this work, we propose CoGenCast, a hybrid generative framework that couples pre-trained LLMs with flow-matching mechanism for effective time series forecasting. Specifically, we reconfigure pre-trained decoder-only LLMs into a native forecasting encoder-decoder backbone by modifying only the attention topology, enabling bidirectional context encoding and causal representation generation. Building on this, a flow-matching mechanism is further integrated to model temporal evolution, capturing continuous stochastic dynamics conditioned on the autoregressively generated representation. Notably, CoGenCast naturally supports multimodal forecasting and cross-domain unified training. Extensive experiments on multiple benchmarks show that CoGenCast achieves competitive performance compared to previous baselines. Code is available at https://github.com/liuyaguo/_CoGenCast.

Figures

Figures reproduced from arXiv: 2602.03564 by Daoyu Wang, Mingyue Cheng, Qi Liu, Xiaoyu Tao, Yaguo Liu.

Figure 1
Figure 1. Figure 1: Overview of our proposed CoGenCast. Left (Training): We reconfigure decoder-only LLMs into an encoder–decoder backbone by attention-only modification, and perform continuous flow-matching mechanism conditioned on the LLM-generated representation. Right (Inference): Future patches are generated autoregressively and sampled via one-step flow-matching generation with low-latency. let C denote the context feat… view at source ↗
Figure 2
Figure 2. Figure 2: Performance comparison between in-domain and cross￾domain training.The bar chart shows the MSE reduction achieved by cross-domain training across four representative benchmarks. based baselines (e.g., PatchTST), indicating that the gains stem from a coupled modeling of both structural depen￾dencies and temporal dynamics rather than architectural scaling alone. In summary, CoGenCast delivers state-of￾the-ar… view at source ↗
Figure 4
Figure 4. Figure 4: Comparative analysis on the number of function evalua￾tions (NFE). We compare the forecasting performance across 1, 2, and 3 sampling steps. 4.4. Comparison with Generative Models Comparison with Generative Models. To further evalu￾ate the architectural advantages of our approach, we conduct a comparative study by replacing our generation module with vanilla diffusion and flow-matching components within th… view at source ↗
Figure 6
Figure 6. Figure 6: Forecasting with uncertainty on four datasets. The plots compare the ground truth trajectories with the model’s mean predictions, along with the 50% and 80% predictive intervals. However, the performance gap between the 0.6B and 4B models is marginal. Considering the comprehensive trade￾off between computational resources, training time, and forecasting performance, we selected Qwen3-0.6B as our default ba… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of forecasting results on Energy, Health, ETTh1, and ETTm1 datasets. We visualize the predicted trajectories of our method against the ground truth and leading baselines (Time-LLM, CDPM, and PatchTST). To provide a qualitative assessment of forecasting fidelity, we visualize the predicted trajectories of our framework alongside the ground truth and competitive baselines, including Time-LLM, CDPM… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 17 linked inside Pith

  1. [1]

    Multi-patch prediction: Adapting llms for time series representation learning.arXiv preprint arXiv:2402.04852,

    Bian, Y ., Ju, X., Li, J., Xu, Z., Cheng, D., and Xu, Q. Multi-patch prediction: Adapting llms for time series representation learning.arXiv preprint arXiv:2402.04852,

    Pith/arXiv arXiv

  2. [6]

    Z., Wang, S., Zhang, J., Li, Z., and Chen, T

    Hu, Y ., Wang, X., Ding, Z., Wu, L., Zhang, H., Li, S. Z., Wang, S., Zhang, J., Li, Z., and Chen, T. Flowts: Time series generation via rectified flow.arXiv preprint arXiv:2411.07506,

    Pith/arXiv arXiv

  3. [7]

    Fstllm: Spatio-temporal llm for few shot time series forecasting

    Jiang, Y ., Chen, Y ., Li, X., Chao, Q., Liu, S., and Cong, G. Fstllm: Spatio-temporal llm for few shot time series forecasting. InForty-second International Conference on Machine Learning. Jin, M., Wang, S., Ma, L., Chu, Z., Zhang, J. Y ., Shi, X., Chen, P.-Y ., Liang, Y ., Li, Y .-F., Pan, S., et al. Time-llm: Time series forecasting by reprogramming la...

    Pith/arXiv arXiv

  4. [8]

    Flow matching with gaussian process priors for probabilistic time series forecasting.arXiv preprint arXiv:2410.03024,

    Kollovieh, M., Lienen, M., L ¨udke, D., Schwinn, L., and G¨unnemann, S. Flow matching with gaussian process priors for probabilistic time series forecasting.arXiv preprint arXiv:2410.03024,

    Pith/arXiv arXiv

  5. [10]

    Gener- ative pretrained hierarchical transformer for time series forecasting

    Liu, Z., Yang, J., Cheng, M., Luo, Y ., and Li, Z. Gener- ative pretrained hierarchical transformer for time series forecasting. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pp. 2003–2013, 2024c. Lu, Y ., Sun, Q., Wang, X., Jiang, Z., Zhao, H., and He, K. Bidirectional normalizing flow: From data to noise and bac...

    arXiv 2003

  6. [11]

    One-step latent-free im- age generation with pixel mean flows.arXiv preprint arXiv:2601.22158,

    Lu, Y ., Lu, S., Sun, Q., Zhao, H., Jiang, Z., Wang, X., Li, T., Geng, Z., and He, K. One-step latent-free im- age generation with pixel mean flows.arXiv preprint arXiv:2601.22158,

    Pith/arXiv arXiv

  7. [13]

    M., Cornelis, B., and Munteanu, A

    Moghadas, S. M., Cornelis, B., and Munteanu, A. Fr `eqflow: Long-term forecasting using lightweight flow matching. arXiv preprint arXiv:2511.16426,

    arXiv

  8. [14]

    H., Sinthong, P., and Kalagnanam, J

    Nie, Y ., Nguyen, N. H., Sinthong, P., and Kalagnanam, J. A time series is worth 64 words: Long-term fore- casting with transformers. arxiv 2022.arXiv preprint arXiv:2211.14730,

    Pith/arXiv arXiv 2022

  9. [15]

    Langtime: A language-guided unified model for time se- ries forecasting with proximal policy optimization.arXiv preprint arXiv:2503.08271,

    Niu, W., Xie, Z., Sun, Y ., He, W., Xu, M., and Hao, C. Langtime: A language-guided unified model for time se- ries forecasting with proximal policy optimization.arXiv preprint arXiv:2503.08271,

    Pith/arXiv arXiv

  10. [16]

    From values to tokens: An llm-driven framework for context-aware time series forecasting via symbolic discretization.arXiv preprint arXiv:2508.09191,

    Tao, X., Zhang, S., Cheng, M., Wang, D., Pan, T., Pan, B., Zhang, C., and Wang, S. From values to tokens: An llm-driven framework for context-aware time series forecasting via symbolic discretization.arXiv preprint arXiv:2508.09191,

    Pith/arXiv arXiv

  11. [18]

    and He, K

    Wang, R. and He, K. Diffuse and disperse: Image gener- ation with representation regularization.arXiv preprint arXiv:2506.09027,

    Pith/arXiv arXiv

  12. [19]

    k2 vae: A koopman-kalman enhanced variational autoen- coder for probabilistic time series forecasting.arXiv preprint arXiv:2505.23017,

    Wu, X., Qiu, X., Gao, H., Hu, J., Yang, B., and Guo, C. k2 vae: A koopman-kalman enhanced variational autoen- coder for probabilistic time series forecasting.arXiv preprint arXiv:2505.23017,

    Pith/arXiv arXiv

  13. [20]

    Chatts: Aligning time series with llms via synthetic data for enhanced understanding and reasoning.arXiv preprint arXiv:2412.03104,

    Xie, Z., Li, Z., He, X., Xu, L., Wen, X., Zhang, T., Chen, J., Shi, R., and Pei, D. Chatts: Aligning time series with llms via synthetic data for enhanced understanding and reasoning.arXiv preprint arXiv:2412.03104,

    arXiv

  14. [21]

    Non-stationary diffusion for probabilistic time series forecasting.arXiv preprint arXiv:2505.04278,

    Ye, W., Xu, Z., and Gui, N. Non-stationary diffusion for probabilistic time series forecasting.arXiv preprint arXiv:2505.04278,

    Pith/arXiv arXiv

  15. [22]

    Condi- tional denoising meets polynomial modeling: A flexible decoupled framework for time series forecasting.arXiv preprint arXiv:2410.13253, 2024a

    Zhang, J., Cheng, M., Tao, X., Liu, Z., and Wang, D. Condi- tional denoising meets polynomial modeling: A flexible decoupled framework for time series forecasting.arXiv preprint arXiv:2410.13253, 2024a. Zhang, X. N., Pu, Y ., Kawamura, Y ., Loza, A., Bengio, Y ., Shung, D., and Tong, A. Trajectory flow matching with applications to clinical time series mo...

    arXiv

  16. [23]

    Samples are organized in (Train/Validation/Test)

    Table 7.Full dataset descriptions. Samples are organized in (Train/Validation/Test). DATASETSLOOK-BACKPREDICTEDVARIABLESSAMPLESDOMAINFREQUENCY Energy 96{12,24,36,48}1 928/138/284 Energy 1 Week ETTh1 96{12,24,36,48}7 8293/2869/2869 Energy 1 Hour ETTh2 96{12,24,36,48}7 8293/2869/2869 Energy 1 Hour ETTm1 96{12,24,36,48}7 34417/11473/11473 Energy 15 Mins ETTm...

    1993

  17. [26]

    Solar (Lai et al., 2018)This dataset records the solar power production of 137 PV plants in 2006, which are sampled every 10 minutes. B. Implementation Details B.1. Compared Baselines To more comprehensively evaluate the capabilities of our proposed method, we compare our method with eight representative baselines in our experiments. Specifically, we incl...

    2018

  18. [27]

    This deterioration likely stems from the smoothing effect of large patches, which obscures critical short-term dynamics required for accurate generation

    consistently leads to performance degradation, most notably in the Health benchmark where the error significantly increases compared to the optimal setting. This deterioration likely stems from the smoothing effect of large patches, which obscures critical short-term dynamics required for accurate generation. Similarly, while small patches benefit specifi...

    arXiv 1910

  19. [2016]

    Influenza-Like Illness (ILI) statistics spanning from 1954 to the present, recorded at a weekly frequency

    Health (Liu et al., 2024a)This dataset comprises weekly U.S. Influenza-Like Illness (ILI) statistics spanning from 1954 to the present, recorded at a weekly frequency. The accompanying data is collected from the CDC Weekly U.S. Influenza Surveillance Report, released weekly. Wind (Li et al., 2022)This dataset comprises wind power measurements sampled ever...

    1954

  20. [2018]

    Environment (Liu et al., 2024a)This dataset comprises daily air quality time series collected from the U.S

    It is divided into four subsets, each with different recording intervals: ETTh1 and ETTh2 have hourly recordings, while ETTm1 and ETTm2 are recorded every 15 minutes. Environment (Liu et al., 2024a)This dataset comprises daily air quality time series collected from the U.S. EPA spanning the full available period, recorded at a daily frequency. Exchange (W...

    2021

  21. [2019]

    Z., and He, K

    Geng, Z., Deng, M., Bai, X., Kolter, J. Z., and He, K. Mean flows for one-step generative modeling.arXiv preprint arXiv:2505.13447, 2025a. Geng, Z., Lu, Y ., Wu, Z., Shechtman, E., Kolter, J. Z., and He, K. Improved mean flows: On the chal- lenges of fastforward generative models.arXiv preprint arXiv:2512.02012, 2025b. Gruver, N., Finzi, M., Qiu, S., and ...

    Pith/arXiv arXiv

  22. [2021]

    Timedart: A dif- fusion autoregressive transformer for self-supervised time series representation.arXiv preprint arXiv:2410.05711,

    Wang, D., Cheng, M., Liu, Z., and Liu, Q. Timedart: A dif- fusion autoregressive transformer for self-supervised time series representation.arXiv preprint arXiv:2410.05711,

    Pith/arXiv arXiv

  23. [2022]

    Time-mmd: Multi-domain multimodal dataset for time series analysis.Advances in Neural Information Processing Systems, 37:77888–77933, 2024a

    Liu, H., Xu, S., Zhao, Z., Kong, L., Prabhakar Kamarthi, H., Sasanur, A., Sharma, M., Cui, J., Wen, Q., Zhang, 9 CoGenCast: A Coupled Autoregressive–Flow Generative Framework for Time Series Forecasting C., et al. Time-mmd: Multi-domain multimodal dataset for time series analysis.Advances in Neural Information Processing Systems, 37:77888–77933, 2024a. Li...

    Pith/arXiv arXiv

  24. [2023]

    Dynamical diffusion: Learning temporal dynamics with diffusion models.arXiv preprint arXiv:2503.00951,

    Guo, X., Zhang, Y ., Chen, B., Xu, H., Wang, J., and Long, M. Dynamical diffusion: Learning temporal dynamics with diffusion models.arXiv preprint arXiv:2503.00951,

    Pith/arXiv arXiv

  25. [2024]

    A com- prehensive survey of time series forecasting: Concepts, challenges, and future directions.Authorea Preprints, 2025a

    Cheng, M., Liu, Z., Tao, X., Liu, Q., Zhang, J., Pan, T., Zhang, S., He, P., Zhang, X., Wang, D., et al. A com- prehensive survey of time series forecasting: Concepts, challenges, and future directions.Authorea Preprints, 2025a. Cheng, M., Tao, X., Liu, Q., Zhang, H., Chen, Y ., and Lian, D. Cross-domain pre-training with language models for transferable ...

    arXiv

  26. [2025]

    Deconstructing denoising diffusion models for self-supervised learning

    Chen, X., Liu, Z., Xie, S., and He, K. Deconstructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404,

    Pith/arXiv arXiv

  27. [2026]

    Time series forecasting as reasoning: A slow-thinking approach with reinforced llms.arXiv preprint arXiv:2506.10630,

    Luo, Y ., Zhou, Y ., Cheng, M., Wang, J., Wang, D., Pan, T., and Zhang, J. Time series forecasting as reasoning: A slow-thinking approach with reinforced llms.arXiv preprint arXiv:2506.10630,

    Pith/arXiv arXiv