acceptodds
Withdrawn from ICLR 2027

RouteCast: Instance-Dependent Multi-View Routing for Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series foundation models (TSFMs) primarily operate on numerical sequences, yet the same observed history can be represented through alternative views that expose different pretrained inductive biases. However, the usefulness of these derived views need not be uniform across forecasting instances, making static multi-view integration potentially suboptimal. We introduce RouteCast, a parameter-efficient framework that formulates multi-view time-series forecasting as an instance-dependent view relevance problem. RouteCast separates view inference from view utilization: Context Distillation constructs instance-conditioned representations from visual and textual views derived solely from the observed history, while Context Routing dynamically regulates their relative influence on a frozen TSFM. All pretrained backbones remain frozen, with training restricted to lightweight adaptation modules. Across 12 forecasting benchmarks, RouteCast consistently improves over its underlying TSFM backbone and remains competitive with strong recent and controlled multi-view baselines. With Chronos as the backbone, RouteCast outperforms full fine-tuning on 11 of 12 datasets, while analogous backbone studies with Timer and TimesFM show similar improvements. Ablations further show complementary benefits from instance-conditioned distillation and adaptive routing. These results demonstrate the value of regulating alternative views according to their instance-dependent relevance for parameter-efficient TSFM adaptation.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Multi-View Learning, Adaptive Routing, Parameter-Efficient Learning

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_cEitOpenReview ↗
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
  1. Please address the above weaknesses

  2. In Table 16, does the inference cost include the text generation phase by Qwen3-VL-8B?

  3. 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?

  4. 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?

  5. 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?

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

RouteCast adapts a frozen time-series foundation model using visual and textual views generated from the numerical history. Context Distillation produces prompts conditioned on each input. Context Routing uses temporal hidden states to select and inject auxiliary features across layers. Experiments evaluate twelve datasets and several backbone and encoder combinations.

Critical strengths
  1. The motivation is clear: different views of the history may help different forecasting samples. The paper also explains that these views are derived from the observed series.
  2. The experiments cover several backbones and auxiliary encoders. The ablations and routing analysis help explain how the model uses the additional views.
Critical weaknesses
  1. The methodological novelty seems limited. Context Distillation mainly uses input-dependent soft prompts, and Context Routing in Eqs. (4)-(7) uses cross-attention with residual connections. Cross-attention already assigns different weights to different inputs. What does the proposed routing add beyond applying these existing components to a frozen forecasting model?
  2. The normalization procedure needs clarification. Appendix C describes standardization within input-target window pairs and, for some datasets, over the entire series. Are future target values included in these statistics? Please specify how the statistics are computed and confirm that they use only information available when each forecast is made.
  3. A simple numerical-only comparison is missing. Could you keep the same backbone frozen and train an extra module using only the numerical history, with a similar number of trainable parameters?
Questions for authors

Please see weaknesses.

Reviewer_VZvaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RouteCast. RouteCast augments a frozen time-series foundation model with visual and textual views derived from the same input history, each processed by frozen pretrained encoders. Context Distillation learns instance-conditioned soft prompts for these views, and Context Routing computes per-layer weights that regulate how strongly each view influences the TSFM's hidden states. On 12 benchmark datasets, the paper reports results about the effectiveness of RouteCast, with ablation analysis.

Critical strengths

S1. The question of when a derived (visual or textual) view of the same history should influence a forecast is a reasonable and underexplored framing, and the proposed decomposition into instance-conditioned distillation and per-layer routing is clearly described.

S2. The experimental coverage is broad: multiple TSFM backbones, vision/text encoder pairs, zero-shot and fine-tuned backbones, controlled multi-view baselines, an ablation with seeds, and additional long-horizon results.

Critical weaknesses

W1. Some important baselines appear to be missing, including newer TSFMs Moirai2.0, TimesFM-3, and Sundial. The experiments do not isolate the effect of the derived views from the effect of the added trainable capacity. Table 15 shows RouteCast trains 115M parameters with Qwen and 396M with Llama, comparable to the 205M-parameter Chronos backbone that full fine-tuning updates. The crucial control, a numerical-only adapter with comparable trainable capacity (learned prompts/adapters on Chronos without any view), is absent from Tables 2 and 4. Calling the method "parameter-efficient" and "lightweight" may not be supported by these numbers.

W2. The evidence that views and routing matter appears to be limited. The average gain over fine-tuned Chronos is about small. The controlled multi-view baselines' results are essentially identical whether encoders are pretrained or trained from scratch and whether an MLP or Transformer is used, which may suggest the auxiliary views are largely ignored by those models and that most of the gain over zero-shot comes from supervised training on the target data. Text-only RouteCast in Table 4 is worse than finetuned Chronos. The CD/CR ablation shows the static-prompt variant in Table 9 is slightly worse than FT, and the full model gains a slight average MSE on three datasets when compared with the static-prompt variant, which may be within the range that an equally sized numerical-only adapter might deliver.

Questions for authors

Q1. It is better to include newer TSFMs Moirai2.0, TimesFM-3, and Sundial in the experiments.

Q2. It is better to add a numerical-only control with the same trainable modules (prompt generator and routing/adapter capacity) but no visual or textual view, on the benchmark datasets, and also a LoRA-style finetuning of Chronos with a comparable parameter budget.

Q3. What are the evaluation details behind Table 2? Were the series normalized per the Chronos protocol? It is better to state the normalization, whether MSE is on scaled values, and the finetuning hyperparameters and early-stopping protocol used for the FT baselines.

Q4. Why are the four controlled multi-view baselines nearly identical across initialization and architecture? Do they use the views at all? What happens if the views are replaced by noise?

Q5. Does the textual description from Qwen3-VL see only the context window? It is better to confirm the prompt and that no future values or dataset-level statistics are included, and report the cost of generating descriptions per instance at inference time.

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

Related papers

Discussion (0)

Sign in to comment.