- Rating
- 3
- Confidence
- 2(−1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper proposes a time series foundation model UniTFU, that incorporates unit-level Time Series Foundation Units (TFUs) and their aggregated scene context into the attention mechanism to guide temporal dynamics modeling. The contributions include (1) introduce TFUs as reusable and composable prototypes of recurring local dynamics discovered from multi-scale temporal segments (2) UniTFU, a TSFM that explicitly integrates the discovery, reuse, and composition of TFUs into forecasting, enabling shared local dynamics to be leveraged across domains (3) systematically evaluate UniTFU in cross-domain settings and on standard forecasting benchmarks, demonstrating robust cross-domain generalization and strong forecasting performance.
Critical strengths
- The authors proposed a novel Time Series Foundation Units (TFUs), where the motivation is clear and the computation is straightforward.
- Based on TFUs, the authors introduced TFU attention that incorporates the learned prototype and contextualized information for time series forecasting.
- The proposed method is shown to be effective in cross-domain forecasting and performs better than other foundation models.
Critical weaknesses
- The discussion of Adaptive TFU Prototype Bank can be improved. Maybe consider the case with and without new dynamics. The authors introduced a few new terms in the paragraph after Eq. 6, such as anchor and , and it would be better to make it clearer.
- According to Table 3, the gap between UniTFU and other forecasting modules is small (ETTh2, ETTm2, ECL, Weather and Solar). Given the efficiency analysis in Table 6, it is not quite easy to tell why the seemingly more heavily loaded model UniTFU is better than the standard Forecasting Model.
Questions for authors
- Is each prototype a learnable vector, randomly initialised?
- For line 232, the prototype-conditioned TFU sequence U is written as a sequence of . If only one prototype matches one position, does it indicate ?
- The merge buffer and novelty buffer part is a bit unclear. Does this process (Eq. 5) change the total number of prototypes?
6 comments
Summary of Response to Reviewer 5zWt
Thank you again for your thoughtful and constructive comments, which have helped us improve our paper. We clarified the adaptive TFU mechanism and extended the efficiency analysis to examine the accuracy–efficiency trade-off. We kindly invite you to reconsider the contributions and insights of our work. We welcome any further questions or discussion.
Eagerly Await Your Response
Dear Reviewer 5zWt,
Thank you again for your valuable and constructive review. We have clarified the adaptive TFU mechanism and provided additional accuracy–efficiency comparisons, explicitly acknowledging the computational trade-offs in our revised manuscript.
We would greatly appreciate it if you could review our responses and share any remaining concerns. We would be happy to provide further clarifications before the rebuttal period ends. If our responses satisfactorily address your concerns, we kindly hope you could reconsider your rating.
Thank you again for your support and constructive feedback.
Best regards,
Authors
Response to Reviewer 5zWt
Questions 1–3 and Weakness 1: Adaptive TFU Prototype Bank.
Thank you for these constructive comments. We clarify prototype initialization, TFU composition, and adaptive dictionary maintenance as follows:
Prototype definition and initialization. The TFU bank stores a collection of D-dimensional vectors representing recurring local temporal patterns, rather than raw waveforms. Each segment is encoded from the mean, peak, and temporal change of its token representations. The bank starts empty, with the first retained segment embedding initializing the first TFU. These prototypes are refined through grouped EMA updates (Eq. 6), rather than gradient descent. During cross-domain adaptation, the bank is initialized from the source checkpoint.
Single-prototype match. When only one TFU pk is active, all retained candidates match it, yielding ut=pk at covered positions, even with overlapping segments. More generally, a single candidate covering position t also yields ut=pk, while overlapping candidates assigned to different prototypes produce a confidence-weighted combination (Eq. 4).
Dictionary maintenance with or without new dynamics. Eq. (5) only routes candidates, without changing the number of active TFUs. Existing dynamics enter the merge buffer for EMA refinement. Novel candidates enter the novelty buffer, where an anchor candidate defines a group of similar embeddings. A new TFU is created only when the group contains sufficient candidates and capacity permits. Thus, prototype birth increases the active count, while consolidation of redundant prototypes decreases it; the maximum capacity remains fixed.
These mechanisms are further clarified in Section 3.1 and Appendix A.3 of the revised version.
Weakness 2. Forecasting Accuracy and Inference Cost.
Thank you for raising this concern. We additionally evaluated DLinear, PatchTST, iTransformer, and PhaseFormer under the same inference setting (672→336, batch size 1). The results are reported in Table 7 of the revised version, with forecasting errors in Appendix Table 12. We acknowledge that UniTFU incurs substantially higher inference costs for modest supervised forecasting gains (e.g., on Weather, MSE: 0.256 vs. 0.251; latency: 402.177 vs. 8.964 ms, compared with PatchTST).
However, UniTFU's primary benefit lies in cross-domain transfer rather than inference efficiency. In direct transfer, TFU learning reduces average MSE by 13.4% relative to w/o TFU and 4.3% relative to a fixed VQ dictionary (Section 4.2, Table 2), demonstrating the value of reusable local dynamics. We have clarified this trade-off in the revised manuscript.