acceptodds
Withdrawn from ICLR 2027

TimeHSG: Hierarchical Semantic Guidance with Frozen LLM Embeddings for Multivariate Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate multivariate time-series forecasting requires jointly modeling variablespeciffc local dynamics, cross-variable inffuences, and long-range periodic structures. However, existing forecasting methods often fail to capture these complementary factors within a uniffed framework. Recent large language model (LLM)-based approaches typically rely on model ffne-tuning or direct feature fusion without assigning distinct forecasting roles to frozen LLM embeddings. We use frozen LLM-derived embeddings as hierarchical guidance rather than as prediction features for direct fusion. Based on this insight, we propose TimeHSG, a multivariate forecasting framework that combines embedding-guided local and relational modeling with frequency-domain residual learning. Speciffcally, Semantic-Conditioned Adaptive Modulation (SCAM) uses the embeddings for feature-level modulation of local temporal representations, while the Dynamic Semantic Graph (DSG) uses them to construct interaction topologies for dynamic cross-variable dependencies. A frequency-domain residual branch further complements local encoding by capturing long-range periodic and low-frequency patterns. Experiments on multiple real-world benchmarks show that TimeHSG achieves competitive or superior forecasting performance. By keeping the LLM frozen and reusing cached embeddings, TimeHSG also provides a favorable balance between forecasting accuracy and computational efffciency.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, Large Language Models, Hierarchical Semantic Guidance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6UP4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors proposed a hierarchical LLM semantic guidance for MTS forecasting, leveraging LLMs to assist the local feature learning and topology learning. The similar challenges have been addressed by an existing work with similar techniques, so I recommend the authors to provide detailed justifications on how this work is different from the existing work and what is the gap the authors are trying to solve.

Critical strengths

The writing and the presentation are quite clear for readers to understand, and the experiments are comprehensive.

Critical weaknesses
  1. The authors aim to guide the feature and topology learning with LLMs, and they designed two modules for variable-level learning and topology-level learning after using prompts to extract features from LLMs. However, the similar challenge has been addressed by the existing work [1], and this work also leveraged the two modules (node and edge levels) after extracting prompt features from LLMs. Thus, I think the two works adopted the similar methodology to address the similar challenge. I think the authors should justify the difference and what are the challenges the authors are trying to solve.

  2. How the authors ensure the prompts are accurate to guide the feature and topology learning? LLMs always have hallucinations. Meanwhile, I think the prompt designs should have large influences on the guidance and the final performance, so I am wondering whether the authors have detailed analysis on these prompt designs? Additionally, I noticed the authors used GPT-2 to extract features. However, this one is too small to be called as a LLM, so do you have analysis by using different LLMs like the advanced open-sourced LLMs like QWEN or Deepseek?

  3. I appreciated the authors to have efficient analysis. However, my concern is that the authors assumed the cached embeddings. When you faced the new situations, you would not have the cached embedding, which means you have to process the embeddings by using the LLM. Thus, I think it is reasonable to conduct the efficiency analysis based on the raw input instead of the cached embeddings. Meanwhile, I think this problem also challenges the first motivation, i.e., 'introduce substantial memory and inference overhead'. In this method, I dont think this motivation has been addressed, because you assume the cached embeddings instead of designing a mechanism to solve it. This is not fair because in real inference, you still need a LLM to extract the embeddings.

[1] Wang, Y., Jin, R., Wu, M., Li, X., Xie, L., & Chen, Z. (2024). K-link: Knowledge-link graph from llms for enhanced representation learning in multivariate time-series data. arXiv e-prints, arXiv-2403.

Questions for authors

See above

Reviewer_A98TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TimeHSG combines a patch-based Transformer with frozen GPT-2 embeddings for feature-wise modulation and attention-based cross-variable interaction, plus a frequency-domain residual predictor. Its main contribution is using the same cached embeddings to guide both feature processing and variable relationships. Evaluation covers eight forecasting benchmarks, LLM and non-LLM baselines, ablations, three-seed stability, and cached-inference efficiency, showing competitive but dataset-dependent performance.

Critical strengths

Compact use of frozen LLM representations. Reusing embeddings for both feature modulation and variable interaction yields lower downstream latency and memory usage than the tested LLM baselines under cached inference, giving the design practical value.

Evidence that the embedding pathway helps. The full model outperforms both MLP and random-embedding replacements on horizon-averaged MSE and MAE across all four ablation datasets, supporting the usefulness of pretrained representations within this architecture.

Critical weaknesses

Limited methodological novelty. SCAM is FiLM-style modulation, DSG is embedding-conditioned attention, and the frequency branch is a linear residual predictor. The contribution largely combines established components, while the ablations do not establish a substantive methodological advance or the necessity of this particular combination.

Inconsistent results weaken the empirical evidence. Table A4 excludes an inconsistent iTransformer–ILI result but retains its four-horizon average. SimpleTM’s ETTh2 MSE values average to approximately 0.389, yet Tables 2/A4 report 0.424. These discrepancies undermine confidence in the aggregate comparisons.

Questions for authors

Clarification: How were the iTransformer–ILI averages computed after excluding the inconsistent H=24H=24 result, and why is SimpleTM’s ETTh2 average MSE reported as 0.424 rather than approximately 0.389? After correcting these discrepancies, do the reported aggregate improvements still hold?

Additional validation: Does TimeHSG retain its advantage when the frequency branch is replaced by a comparable time-domain linear residual predictor? Can you also provide full-model ablations removing DSG and the frequency branch separately? These comparisons would materially clarify whether the gains support the claimed mechanisms.

Reviewer_dCzfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TimeHSG combines frozen LLM-guided feature modulation and cross-variable graph modeling with a frequency-domain prediction branch. Results on eight benchmarks are competitive, but the methodological novelty remains unclear. The experiments do not sufficiently isolate the benefit of language pretraining, and evidence that the frequency branch specifically recovers periodic information lost during patch encoding is limited.

Critical strengths
  1. The paper offers a useful perspective on using frozen LLM representations as guidance for distinct forecasting components. Separating feature-level modulation from cross-variable interaction makes the intended roles of the embeddings explicit.
  2. The paper presents a comprehensive empirical study across multiple datasets and forecasting horizons, comparing against both LLM-based and conventional baselines. Extensive experiments provide a broad assessment of the proposed framework.
Critical weaknesses
  1. The motivation is not sufficiently developed. The abstract and introduction identify local dynamics, cross-variable influences, and long-range periodic structures, then motivate a separate component for each. However, the importance of these factors does not by itself establish a research gap. Although the introduction discusses the cost of LLM adaptation and limitations of direct fusion, it does not convincingly explain why these limitations require the proposed hierarchical design. The architecture consequently appears motivated more by combining desirable functions than by addressing a clearly demonstrated failure of existing approaches.
  2. The MLP embedding-replacement ablation in Section 4.3 (Table 3; detailed results in Table A5) lacks implementation details, including the MLP’s inputs, hidden dimensions, and tuning procedure. Without these details, it is difficult to assess whether the reported gains reflect an advantage of LLM embeddings over a comparably tuned numerical encoder.
Questions for authors
  1. In Tables A3/A4, MSE decreases from H=336 to H=720 on ETTh1 and Solar. Is drop_last=True used during evaluation, and are all test windows included? Please also compare TimeHSG with representative baselines using input lengths of 192, 336, and 720 under consistent validation-based model selection. Do its relative advantage and the frequency branch’s benefit persist?
  2. Please provide the inputs, architecture, and tuning details of the MLP replacement experiment in Section 4.3 (Table 3). Please also explain the large degradation of CA+F on ETTm1 and why its performance varies so substantially across datasets.
  3. Page 5 motivates the frequency branch as capturing spectral patterns potentially lost during patch encoding. Figure A4 illustrates its complementary contribution. Could the authors clarify what evidence supports this specific explanation, beyond the benefit of adding a second prediction branch?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.