- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper introduces RouteCast, an adaptive framework complementing a frozen Time Series Foundation Model (TSFM) by integrating two additional views generated from the historical input: a visual view (a plot of the time series input encoded by vision foundation models CLIP/BLIP) and a textual view (a description of the time series input generated by Qwen3-VL-8B and encoded by LLMs like Qwen2/Llama). The paper formulates the problem as “instance-dependent view relevance” when combining all views to improve forecast accuracy. Two components are proposed to address this problem: context distillation to distill instance-conditioned features from the visual and textual views, and context routing to estimate the relevance of each view to the input, leading to improved forecasting performance compared to the backbone.
Critical strengths
RouteCast does not need to update the time series foundation backbone, and both views are derived from the historical input without external data.
Critical weaknesses
-
The novelty is quite weak, and the technical method sounds incremental. The derived multi-view setting (time series -> image + text + frozen VLM) has been studied in the field, such as Time-VLM (ICML 2025) (Table 5), but the paper does not make any empirical comparison with it. The context distillation is similar to conditional prompt learning (CoCoOp [A]), and the context routing (Eqs. 4–7) is standard cross-attention over 2 tokens. Labeling the context usage of TimeCMA as “Fixed” and Time-VLM as having no instance control (Table 5) is not convincing. TimeCMA wraps the values of each input into the prompt and fuses the two views by softmax similarity between the time series and prompt embeddings, which is input-dependent and very similar to the context routing (Eqs. 4–7). Time-VLM also generates the text from the statistics of each input and uses a gate computed from each input.
-
Plot and description inputs could be misleading views, and the routing mechanism does not deal with this. The softmax is applied on 2 views (Eq. 5), so α_vis + α_text = 1. Routing distributes the contribution between the two views, but it cannot turn off both views. This means that if both the plot and the description are misleading, the residual r_t is still added to the hidden state (Eq. 7). In addition, the quality of the description view generated by an LLM is not evaluated in the paper, and there is no robustness test. What if a description is shuffled or does not align with the time series instance, or a plot is too noisy?
-
Lack of thoughtful comparisons. First, the baseline comparison sounds unfair. In Fig. 1 and Table 2, the authors compare RouteCast (already trained on target data) with other zero-shot TSFMs (not trained on target data). The sensible per-backbone comparison (ZS/FT/+RouteCast) is given in Table 3, but it is not sufficient with only 3 datasets. Stronger backbones such as Chronos-2 and Timer-S1 have a better zero-shot average MSE than Chronos-Bolt (1.520 and 1.415 vs. 1.623, Table 2), but they are not used as backbones. Therefore, I wonder whether the gain is still consistent when the backbone is stronger. Second, the FT baselines seem poorly tuned (see Q5).
-
The “lightweight/parameter-efficient” claim is overclaimed. The paper reports that trainable params are 115M–396M (Table 15), approximately 0.56x–1.93x compared to full fine-tuning of Chronos-Bolt (205M). Plus, in Table 16, it seems that the reported inference cost does not include the text view generation phase by Qwen3-VL-8B.
-
The paper lacks important baselines, besides minor gain. The paper claims that the proposed method is parameter-efficient but does not compare with other parameter-efficient fine-tune baselines such as LoRA [B]. Additionally, to see the effectiveness of the two views, the paper does not show how much the method improves compared to noisy or random views under the same modules and backbone. The performance gain is minor compared with other baselines, for example PatchTST (a model using only the time series view) is comparable to RouteCast on ETT (Table 13).
-
Minor error: Table 4 and Table 8 have swapped numbers (BLIP-Qwen and CLIP-Llama rows).
[A] Zhou, Kaiyang, et al. "Conditional prompt learning for vision-language models." 2022 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2022.
[B] Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2022). LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=nZeVKeeFYf9
Questions for authors
-
Please address the above weaknesses
-
In Table 16, does the inference cost include the text generation phase by Qwen3-VL-8B?
-
The authors say that “Because the pretraining corpora of different TSFMs are not necessarily identical, we do not assume that the same exclusion guarantee holds universally for every baseline backbone” (Section 4). Are the datasets excluded from the training corpora of the backbones used in RouteCast, i.e., Chronos-Bolt (the actual checkpoint, Appendix G), Timer, and TimesFM?
-
Appendix C says that standardization is applied “within each input-target window pair” and, for some datasets, “over entire sequences”. Does this imply that the normalization statistics are computed on the target values?
-
Why do the FT baselines use lr 1e-3 (Appendix H), while RouteCast uses 2e-5 (Appendix G)? Please explain why Timer FT is worse than Timer ZS, with std ±0.000 on several datasets? Can the FT baselines be re-run with a tuned lr?