Elamur Research⌕
☰
home / papers / 2026 / TimeSAF: Towards LLM-Guided Semantic Asynchronous Fusion for Time Series Forecasting

TimeSAF: Towards LLM-Guided Semantic Asynchronous Fusion for Time Series Forecasting

Paper · 2026 · Fan Zhang, Shiming Fan and Hua Wang

Summary

Most LLM-assisted forecasters fuse text and numeric features densely at every network layer. The authors argue this ignores the gap in granularity between the two modalities. Abstract LLM semantics get tangled with fine-grained numerical fluctuations, which they call "semantic perceptual dissonance," and this can even cause negative transfer.

TimeSAF addresses this with hierarchical asynchronous fusion. The time series is normalized with RevIN, split into patches and encoded. In parallel, rule-generated per-variable statistical prompts pass through a frozen GPT-2. A separate fusion trunk with learnable queries collects high-level information from both branches bottom-up into a compact memory. A stage-wise refinement decoder then feeds this memory back top-down into the temporal backbone, at only a few depths and through gated cross-attention blocks.

On standard long-term forecasting benchmarks, TimeSAF beats LLM-based and conventional baselines and generalizes well in few-shot and zero-shot transfer. The work suggests that when and where modalities interact matters as much as whether they interact at all.

Key points

  • Evaluated on seven public LTSF benchmarks, including ETT, Electricity, Weather and Exchange. Comparisons include LLM-based methods such as Time-LLM, UniTime, TimeCMA and CALF, alongside non-LLM models.
  • The semantic backbone is a frozen GPT-2. Prompts come from templates describing each variable's window statistics, following the TimeCMA approach.
  • In the sensitivity analysis, going from 1 to 2 fusion stages reduced error, while 4 stages added little. The default is S = 2, with fusion blocks placed in middle-to-deep layers.
  • Attention maps and t-SNE on Exchange show an "aggregate–refine" loop. After refinement, temporal features move closer to the fusion feature space.
  • Stated limitations:
    • prompts are rule-based templates rather than unstructured text such as news;
    • larger LLMs (e.g., LLaMA-3, GPT-4) were not tested;
    • there is no real-world industrial evaluation.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

Also on the byline: Ludong University.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksDecision Sciences · 249 worksSignal Processing · 493 worksManagement Science and Operations Research · 134 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Fan Zhang, Shiming Fan and Hua Wang. 2026. TimeSAF: Towards LLM-Guided Semantic Asynchronous Fusion for Time Series Forecasting. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1208

@inproceedings{zhang2026timesaf,
  title = {{TimeSAF: Towards LLM-Guided Semantic Asynchronous Fusion for Time Series Forecasting}},
  author = {Fan Zhang, Shiming Fan and Hua Wang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1208},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1208}
}

References

49 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Adebiyi A Ariyo, Adewumi O Adewumi, and Charles K Ayo. 2014. Stock price prediction using the arima model. In 2014 UKSim-AMSS 16th international conference on computer modelling and simulation, pages 106–112. IEEE. find
  2. Ching Chang, Wei-Yao Wang, Wen-Chih Peng, and Tien-Fu Chen. 2025. Llm4ts: Aligning pre-trained llms as data-efficient time-series forecasters. ACM Transactions on Intelligent Systems and Technology, 16(3):1–20. find
  3. Zhiwei Chen, Yupeng Hu, Zhiheng Fu, Zixu Li, Jiale Huang, Qinlei Huang, and Yinwei Wei. 2026. Intent: Invariance and discrimination-aware noise mitigation for robust composed image retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 20463–20471. find
  4. Zhiwei Chen, Yupeng Hu, Zixu Li, Zhiheng Fu, Xuemeng Song, and Liqiang Nie. 2025a. Offset: Segmentation-based focus shift revision for composed image retrieval. In Proceedings of the ACM International Conference on Multimedia, page 6113–6122. find
  5. Zhiwei Chen, Yupeng Hu, Zixu Li, Zhiheng Fu, Haokun Wen, and Weili Guan. 2025b. Hud: Hierarchical uncertainty-aware disambiguation network for composed video retrieval. In Proceedings of the ACM International Conference on Multimedia, page 6143–6152. find
  6. Robert B Cleveland, William S Cleveland, Jean E McRae, Irma Terpenning, and 1 others. 1990. Stl: A seasonal-trend decomposition. J. off. Stat, 6(1):3–73. find
  7. Kuiye Ding, Fanda Fan, Yao Wang, Ruijie Jian, Xiaorui Wang, Luqi Gong, Yishan Jiang, Chunjie Luo, and Jianfeng Zhan. 2025. Dualsg: A dual-stream explicit semantic-guided multivariate time series forecasting framework. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 508–5… find
  8. Shiming Fan, Hua Wang, and Fan Zhang. 2025a. Caw-former: A cross variable attention with discrete wavelet denoising for multivariate time series forecasting. Knowledge-Based Systems, page 113846. find
  9. Shiming Fan, Hua Wang, and Fan Zhang. 2025b. Fs-mamba: A dual-expert architecture with fast global attention and local-enhanced state-space mamba for time series forecasting. Knowledge-Based Systems, page 115233. find
  10. Yunfeng Ge, Ming Jin, Yiji Zhao, Hongyan Li, Bo Du, Chang Xu, and Shirui Pan. 2025a. Eventtsf: Event-aware non-stationary time series forecasting. arXiv preprint arXiv:2508.13434. — [arXiv:2508.13434](https://arxiv.org/abs/2508.13434) find
  11. Yunfeng Ge, Jiawei Li, Yiji Zhao, Haomin Wen, Zhao Li, Meikang Qiu, Hongyan Li, Ming Jin, and Shirui Pan. 2025b. T2s: High-resolution time series generation with text-to-series diffusion models. arXiv preprint arXiv:2505.02417. — [arXiv:2505.02417](https://arxiv.org/abs/2505.02417) find
  12. Nate Gruver, Marc Finzi, Shikai Qiu, and Andrew G Wilson. 2023. Large language models are zero-shot time series forecasters. Advances in Neural Information Processing Systems, 36:19622–19635.
  13. Lu Han, Xu-Yang Chen, Han-Jia Ye, and De-Chuan Zhan. 2024. Softs: Efficient multivariate time series forecasting with series-core fusion. arXiv preprint arXiv:2404.14197. — [arXiv:2404.14197](https://arxiv.org/abs/2404.14197) find
  14. Yupeng Hu, Zixu Li, Zhiwei Chen, Qinlei Huang, Zhiheng Fu, Mingzhu Xu, and Liqiang Nie. 2026. Refine: Composed video retrieval via shared and differential semantics enhancement. ACM Transactions on Multimedia Computing, Communications and Applications. find
  15. Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and 1 others. 2023. Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728.
  16. Duc Kieu, Tung Kieu, Peng Han, Bin Yang, Christian S Jensen, and Bac Le. 2024. Team: Topological evolution-aware framework for traffic forecasting– extended version. arXiv preprint arXiv:2410.19192. — [arXiv:2410.19192](https://arxiv.org/abs/2410.19192) find
  17. Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. 2021. Reversible instance normalization for accurate timeseries forecasting against distribution shift. In International conference on learning representations. find
  18. Zixu Li, Zhiwei Chen, Haokun Wen, Zhiheng Fu, Yupeng Hu, and Weili Guan. 2025. Encoder: Entity mining and modification relation binding for composed image retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 5101–5109. find
  19. Zixu Li, Yupeng Hu, Zhiwei Chen, Qinlei Huang, Guozhi Qiu, Zhiheng Fu, and Meng Liu. 2026a. Retrack: Evidence-driven dual-stream directional anchor calibration network for composed video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 23373–23381. find
  20. Zixu Li, Yupeng Hu, Zhiwei Chen, Shiqi Zhang, Qinlei Huang, Zhiheng Fu, and Yinwei Wei. 2026b. Habit: Chrono-synergia robust progressive learning framework for composed image retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 6762–6770. find
  21. Chenxi Liu, Qianxiong Xu, Hao Miao, Sun Yang, Lingzheng Zhang, Cheng Long, Ziyue Li, and Rui Zhao. 2025a. Timecma: Towards llm-empowered multivariate time series forecasting via cross-modality alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18780–18788. find
  22. Peiyuan Liu, Hang Guo, Tao Dai, Naiqi Li, Jigang Bao, Xudong Ren, Yong Jiang, and Shu-Tao Xia. 2025b. Calf: Aligning llms for time series forecasting via cross-modal fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18915–18923. find
  23. Peiyuan Liu, Beiliang Wu, Yifan Hu, Naiqi Li, Tao Dai, Jigang Bao, and Shu-tao Xia. 2024a. Timebridge: Non-stationarity matters for long-term time series forecasting. arXiv preprint arXiv:2410.04442. — [arXiv:2410.04442](https://arxiv.org/abs/2410.04442) find
  24. Qingxiang Liu, Xu Liu, Chenghao Liu, Qingsong Wen, and Yuxuan Liang. 2024b. Time-ffm: Towards lm-empowered federated foundation model for time series forecasting. Advances in Neural Information Processing Systems, 37:94512–94538. find
  25. Xu Liu, Junfeng Hu, Yuan Li, Shizhe Diao, Yuxuan Liang, Bryan Hooi, and Roger Zimmermann. 2024c. Unitime: A language-empowered unified model for cross-domain time series forecasting. In Proceedings of the ACM Web Conference 2024, pages 4095–4106. find
  26. Xvyuan Liu, Xiangfei Qiu, Xingjian Wu, Zhengyu Li, Chenjuan Guo, Jilin Hu, and Bin Yang. 2026. Rethinking irregular time series forecasting: A simple yet effective baseline. In AAAI. find
  27. Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024d. itransformer: Inverted transformers are effective for time series forecasting. In The Twelfth International Conference on Learning Representations. find
  28. Jiaming Ma, Binwu Wang, Qihe Huang, Guanjun Wang, Pengkun Wang, Zhengyang Zhou, and Yang Wang. 2025. Mofo: Empowering long-term time series forecasting with periodic pattern modeling. Proc. Adv. Neural Inf. Process. Syst. find
  29. Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730.
  30. Xiangfei Qiu, Xingjian Wu, Hanyin Cheng, Xvyuan Liu, Chenjuan Guo, Jilin Hu, and Bin Yang. 2025a. Dbloss: Decomposition-based loss function for time series forecasting. In NeurIPS. find
  31. Xiangfei Qiu, Xingjian Wu, Yan Lin, Chenjuan Guo, Jilin Hu, and Bin Yang. 2025b. DUET: Dual clustering enhanced multivariate time series forecasting. In SIGKDD, pages 1185–1196. find
  32. Xiangfei Qiu, Yuhan Zhu, Zhengyu Li, Xingjian Wu, Bin Yang, and Jilin Hu. 2025c. Dag: A dual correlation network for time series forecasting with exogenous variables. arXiv preprint arXiv:2509.14933. — [arXiv:2509.14933](https://arxiv.org/abs/2509.14933) find
  33. Frank Schorfheide. 2005. Var forecasting under misspecification. Journal of Econometrics, 128(1):99– 136. find
  34. Qiannan Shen and Jing Zhang. 2026. Mftformer: Meteorological-frequency-temporal transformer with block-aligned fusion for traffic flow prediction. Research Square. Preprint, doi:10.21203/rs.3.rs-8770196/v1. — [doi:10.21203/rs.3.rs-8770196/v1](https://doi.org/10.21203/rs.3.rs-8770196/v1) find
  35. Alex Sherstinsky. 2020. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. Physica D: Nonlinear Phenomena, 404:132306. find
  36. Sima Siami-Namini, Neda Tavakoli, and Akbar Siami Namin. 2018. A comparison of arima and lstm in forecasting time series. In 2018 17th IEEE international conference on machine learning and applications (ICMLA), pages 1394–1401. Ieee. find
  37. Hua Wang, Jinghao Lu, and Fan Zhang. 2026a. Eeo-tfv: Escape-explore optimizer for web-scale timeseries forecasting and vision analysis. arXiv preprint arXiv:2602.02551. — [arXiv:2602.02551](https://arxiv.org/abs/2602.02551) find
  38. Hua Wang, Jinghao Lu, and Fan Zhang. 2026b. Ide-altsf: Can non-ideal data contribute to enhancing the performance of time series forecasting models? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 26224–26232. find
  39. Huiqiang Wang, Jian Peng, Feihu Huang, Jince Wang, Junhui Chen, and Yifei Xiao. 2023. Micn: Multi-scale local and global context modeling for long-term series forecasting. In The Eleventh International Conference on Learning Representations. find
  40. Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y Zhang, and Jun Zhou. 2024. Timemixer: Decomposable multiscale mixing for time series forecasting. arXiv preprint arXiv:2405.14616. — [arXiv:2405.14616](https://arxiv.org/abs/2405.14616) find
  41. Shunnan Wang, Min Gao, Zongwei Wang, Yibing Bai, Feng Jiang, and Guansong Pang. 2025. Freqllm: frequency-aware large language models for time series forecasting. In Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pages 3389–3397. find
  42. Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2023. Timesnet: Temporal 2d-variation modeling for general time series analysis. find
  43. Canran Xiao, Jiabao Dou, Zhiming Lin, Zong Ke, and Liwei Hou. 2026a. From points to coalitions: Hierarchical contrastive shapley values for prioritizing data samples. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 15995– 16003. find
  44. Canran Xiao, Tianxiang Xu, Yiyang Jiang, Haoyu Gao, Yuhan Wu, and 1 others. 2026b. Reversible primitive– composition alignment for continual vision–language learning. In The Fourteenth International Conference on Learning Representations. find
  45. Zhijian Xu, Ailing Zeng, and Qiang Xu. 2023. Fits: Modeling time series with 10k parameters. arXiv preprint arXiv:2307.03756. — [arXiv:2307.03756](https://arxiv.org/abs/2307.03756) find
  46. Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121– 11128. find
  47. Fan Zhang, Shiming Fan, and Hua Wang. 2026a. Time-tk: A multi-offset temporal interaction framework combining transformer and kolmogorov-arnold networks for time series forecasting. arXiv preprint arXiv:2602.11190. — [arXiv:2602.11190](https://arxiv.org/abs/2602.11190) find
  48. Fan Zhang, Zhiwei Gu, and Hua Wang. 2026b. Decoding with structured awareness: integrating directional, frequency-spatial, and structural attention for medical image segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 12421–12429. find
  49. Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, and 1 others. 2023. One fits all: Power general time series analysis by pretrained lm. Advances in neural information processing systems, 36:43322–43355. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-timesaf-llm-guided-semantic-asynchronous/