acceptodds
Withdrawn from ICLR 2027

From Temporal Priors to Spatial Reasoning: Graph Adaptation of Frozen Time-Series Foundation Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent spatio-temporal foundation models (STFMs) have demonstrated strong transferability and forecasting capabilities in zero-shot and few-shot settings, but they usually require large-scale spatio-temporal pre-training and substantial computational resources. Moreover, spatial dependencies are often highly domain-specific, making it difficult for a pre-trained spatio-temporal backbone to generalize across heterogeneous spatial structures. To address these limitations, we propose Graft, a framework that integrates lightweight graph adaptation into frozen time-series foundation models (TSFMs), extending them to downstream spatio-temporal forecasting without modifying the foundation backbone. Specifically, Graft extracts latent forecast tokens from the frozen TSFM while constructing temporally aligned relational graph tokens from historical observations, node attributes, and target-domain graph structures. The graph tokens are projected into the forecast-token space and adaptively integrated with the forecast tokens through gated temporal-graph fusion. We further develop a causal-mediated graph adapter (CMGA) that models hierarchical graph dependencies and dynamically injects graph-structured corrections into the temporal representations. Extensive experiments on seven spatio-temporal forecasting benchmarks demonstrate that the proposed method achieves strong few-shot performance and favorable efficiency compared with full TSFM fine-tuning. Further experiments on multivariate time-series anomaly detection validate its generalization across different graph semantics.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-temporal forecasting, Time-series foundation model, Graph adaptation, Cross-task generalization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YWYzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Graft, a lightweight graph adaptation framework. It preserves the temporal priors of the pre-trained TSFM while introducing target-domain relational graph tokens, gated temporal-graph fusion, and a causal-mediated graph adapter to model local, regional, and global dependencies. Experiments show strong performance, good memory efficiency, and cross-task generalization.

Critical strengths

S1: The paper is generally clear and well motivated.

S2: The method is general and is suitable to various TSFMs

Critical weaknesses

W1: The same node-level graph representation is expanded along the output-patch dimension. This may be restrictive, since short-term congestion and long-term trends can involve substantially different dependency graph structures.

W2: The model relies on a fixed target-domain adjacency matrix and does not explicitly model dynamic topology. In traffic systems, dependencies can change with time, incidents, congestion, and other conditions.

W3: The paper represents the entire graph using only the mean of all node tokens, which may discard important spatial heterogeneity.

W4: The paper lacks a formal computational complexity analysis.

W5: The evaluation heavily reply on traffic datasets. Perhaps the paper shall degard to traffic prediction, or use additional domains such as weather, environmental sensing, or energy systems.

W6: Please provide code by platform like anonymous github, as you said in REPRODUCIBILITY STATEMENT.

Questions for authors

Q1: Is there potential functional overlap among the fusion gate, dynamic router, and structural-strength gate?

Reviewer_YgQtOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to efficiently extend pretrained time-series foundation models to graph-structured spatio-temporal forecasting. The authors propose Graft, which freezes the TSFM encoder and native forecasting head, constructs relational graph tokens from historical observations, node attributes, and target-domain graph structure, and injects them into forecast representations through gated temporal–graph fusion and a Causal-Mediated Graph Adapter (CMGA). CMGA models direct local, mediated regional, and global contextual interactions. Experiments on seven forecasting datasets with Chronos-2 and Moirai, plus four multivariate anomaly-detection datasets, suggest that Graft achieves competitive few-shot performance with substantially lower adaptation cost.

Critical strengths

S1. The strongest contribution is the separation of transferable temporal priors from target-specific spatial structure. Freezing the TSFM and learning only a graph adapter directly addresses heterogeneous spatial structures across domains and avoids expensive joint spatio-temporal pretraining, making the framework practically relevant beyond generic parameter-efficient tuning.

S2. The evaluation spans seven forecasting datasets, short- and long-horizon settings, and two TSFM backbones. Graft is competitive with or better than existing TSFM/STFM baselines in many settings. More importantly, compared with full Chronos-2 fine-tuning, it achieves lower forecasting error while reducing training memory by roughly 78%–88% on PEMS08 and PEMS04, directly supporting the efficiency claim.

Critical weaknesses

W1. Table 4 shows that graph adaptation is useful, but it lacks a particularly informative baseline such as frozen TSFM + parameter-matched standard GCN/graph adapter + the same fusion mechanism, without the direct/mediated/global decomposition and routing. Thus, it remains unclear whether CMGA itself provides substantial gains over a simpler graph-conditioned adapter. This directly affects the assessment of methodological novelty.

W2. The paper states that all methods use the same splits and 10% labeled subset, but it does not clearly specify which parameters are updated for TimesFM, Moirai, Rose, UniST, OpenCity, FactoST-v2, etc., what adaptation strategy is used, or how many trainable parameters are involved. Without these details, it is difficult to judge whether the main comparisons are fully matched in adaptation budget.

Questions for authors

Q1. Could the authors add a parameter-matched simple graph-adapter baseline, e.g., frozen TSFM + two-layer GCN + the same gated fusion, but without CMGA’s three branches and dynamic routing?

Q2. For the TSFM/STFM baselines in Table 1, which parameters are updated under the 10% few-shot setting? Are official/recommended adaptation protocols used, and can the authors report the number of trainable parameters?

Reviewer_iiRiOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a method to adapt existing time series foundation models (TSFMs) to spatiotemporal data where dependencies among time series are represented as a graph. The core idea is to 1) take raw inputs and representations extracted by a TSFM, 2) process them with a module trained in a few-shot fashion to integrate information from other nodes, and 3) use the TSFM's decoder to output predictions. The idea is quite interesting and the research timely. However, some of the designs are not fully justified, and there are several issues with the empirical evaluation.

Critical strengths
  • The idea of extending the capabilities of current foundation models to account for spatiotemporal data and priors on their relational structure is definitely relevant and timely.
  • Overall, the proposed graph-based adapter module is a reasonable approach to tackle the problem.
Critical weaknesses

Main weaknesses:

  • The empirical analysis has many limitations which put the paper clearly below the bar for ICLR.
    • There are no reported standard deviations (except for a table in the appendix). As most of the experiments are in a few-shot setting, they should not be too expensive to run, and I don't see any good reason not to run multiple independent runs of each experiment here.
    • The empirical analysis is limited to evaluating the full architecture and some simple ablations. However, these ablations are not enough to justify the complexity of some of the designs. Would the performance change if the proposed components are replaced with simpler processing? In particular, the main important analysis is completely missing: what is the actual impact of wrapping a TSFM? It would be important to at least check how performance changes when: 1) tokens extracted by the TSFM are removed from the input and 2) the FM is replaced by simple linear encoders/decoders. Please see the similar analyses in [1].
    • Related to the previous point, the window length used in all the 12->12 experiments is smaller than the path size typically used by the TSFMs used as backbone. Even in the 96->96 experiment, the context is much lower than that typically used by those models to obtain good results. This again raises some doubts on the impact of the backbone on the observed results.
    • Missing baselines: this is connected to the previous points, using simpler baselines (e.g., historical averages, simple linear AR models, etc.) would provide a better understanding of the current performance of this approach. Moreover, it would be important to share more details on how the baselines were tuned for each scenario. I encourage the authors to share anonymized code for reproducing the results in the experiments to improve reproducibility.
    • Several results seem to be directly copied and pasted from the FactoST-v2 paper. This should be clearly stated.
    • The obtained results are often poor compared to simple approaches in the literature. For example, performance on METR-LA and PEMS-BAY is similar to that of simple baselines. It would help to include baselines such as historical averages and seasonal naïve baselines. Analogous spatiotemporal FMs report similar results, but this does not change that the current setup seems to result in weak forecasting models in some cases.
    • Results in Appendix E.1 give some context on how performance changes with more data available: it would help to have similar curves for the baselines for comparison.
    • My understanding is that Morai includes several of the benchmark datasets used here for pre-training (Chronos 2 might also include those). This would make several of the results in the paper biased.
    • In the anomaly detection experiments, the thresholds are tuned on the test set (lines 355–357)). It does not matter that this is done for all the baselines in the same way. This is a flawed procedure which implies that the results are compeltely biased.

Given the above, while the idea might have potential, its execution falls well below what one would expect from a strong ICLR submission.

Minor issues:

  • As the model cannot be used in a zero-shot setting, the contribution would be much more relevant if the proposed approach yielded performance competitive with state-of-the-art STGNNs, at least in full training scenarios. The results suggest that it is not the case.
  • The motivations for the designs of the different modules are not fully discussed. What's the advantage of, e.g., the CMGA modules over simpler alternatives?
Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.