acceptodds
Under review as a conference paper at ICLR 2027

GUST: Graph-Guided Continual Adaptation of Weather and Climate Foundation Models

Abstract

Weather and climate foundation models must accommodate evolving prediction needs, yet existing adaptation typically assumes a static lifecycle in which each task is learned independently from the same pretrained checkpoint. We systematically investigate continual learning (CL) under sequential access to task data across three settings: variable CL expands atmospheric targets, regional CL extends geographical coverage, and climate-projection CL introduces new response objectives. We find that forgetting is strongly scenario-dependent. Direct sequential adaptation remains effective under compatible task shifts but severely degrades earlier capabilities when prediction objectives change. Cross-task transfer is also uneven and directional. Motivated by these findings, we propose Graph-guided Update Selection and Transfer (GUST), which stores acquired knowledge in frozen low-rank adapters and selectively composes their updates through a causal task graph guided by structured relations, learned task geometry, and online transfer feedback. GUST achieves the strongest overall acquisition-retention trade-off among CL methods. In variable CL, it obtains an average relative wRMSE of 1.102 and reduces relative forgetting from 39.55% to 9.67%. In regional CL, it achieves a relative wRMSE of 1.044 with zero measured forgetting. In climate-projection CL, it achieves the lowest oracle-normalized error on three of four metrics. These results identify when dedicated CL is necessary and provide an effective avenue for long-term evolution of weather and climate foundation models.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Weather Forecast, Climate Projection, Continual Learning, Foundation Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FTP2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies continual learning (CL) for weather and climate foundation models across three task-expansion scenarios: variable CL (new atmospheric targets), regional CL (new geographical domains), and climate-projection CL (new response objectives). Through an empirical investigation, the authors find that forgetting is strongly scenario-dependent and that cross-task transfer is uneven and directional. They propose GUST, which stores task knowledge in frozen low-rank adapters and selectively composes historical updates via a directed task graph integrating structured domain relations, learned task geometry, and online transfer feedback. Experiments on WeatherBench (ClimaX backbone) and ClimateBench report improvements in the acquisition–retention trade-off relative to compared CL baselines.

Critical strengths

S1. Well-motivated problem formulation. The paper addresses a genuine gap: existing weather/climate foundation models operate under a static lifecycle, yet deployment demands sequential capability expansion. Formulating CL across three complementary axes (variables, regions, climate responses) provides representative scenarios that are broadly relevant to the AI4Science community.

S2. Thorough empirical investigation before method design. Section 3.2 is a standout contribution. The finding that forgetting severity is associated with task-structural compatibility (rather than sequential optimization per se), and that parameter isolation alone leaves a substantial gap to joint training, provides clear and actionable guidance. The stability–plasticity analysis in Fig. 1(b) is informative.

S3. Reasonable method design with interpretable structure. GUST cleanly decouples stable knowledge storage (frozen LoRA adapters) from adaptive knowledge reuse (graph-guided composition). The integration of structured domain priors, learned task geometry, and online utility/interference signals into a unified routing score (Eq. 4) is principled. The temporal DAG structure enforces causal ordering and prevents access to future tasks.

Critical weaknesses

W1. Task-order robustness is not evaluated. All experiments use fixed task sequences. Since GUST freezes routing rows after task acquisition, changing the order affects available historical adapters and graph propagation paths. Results under at least one alternative ordering would help establish whether the improvements generalize beyond the chosen sequences.

W2. Scalability characterization is insufficient. GUST stores one adapter per task and maintains a K×K routing matrix. The paper lacks measurements of memory growth, training overhead, and inference latency. Although the current propagated routing row can be computed in O(K²) per routing update, the practical costs of adapter composition and deployment remain unclear, particularly for longer task sequences.

W3. Baseline adaptation and tuning require clarification. Sec. 5.1 uses an identical learning rate and training schedule across methods with substantially different parameterizations. In particular, adapting prompt-based classification methods to dense weather regression requires non-trivial design choices. The paper does not sufficiently specify prompt injection, trainable parameter counts, or per-method hyperparameter tuning. These omissions make it difficult to assess the fairness of the comparisons.

Questions for authors

Q1. How robust is GUST to different task orderings? Could the authors report results for a reversed or shuffled variable CL sequence?

Q2. What does two-hop propagation provide beyond directly learning adapter composition weights? Since the final update remains a convex combination of historical updates, two-hop propagation does not expand the set of possible mixtures relative to an unconstrained simplex router. Is its benefit primarily an inductive bias or improved optimization?

Q3. Were the baselines tuned individually? Could the authors clarify the prompt-based implementations and report the memory and inference overhead of GUST?

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

This paper investigates continual learning for weather and climate foundation models, formalizes three task-incremental settings, and proposes GUST, which stores each task as a frozen low-rank adapter and selectively composes historical updates through a directed task graph without replaying past data, evaluated with ClimaX on WeatherBench/ERA5 and ClimateBench. Nevertheless, the paper still has many problems, ranging from unsupported claims and missing baselines to inconsistencies between its abstract and the reported results.

Critical strengths

1、Continual learning in the climate domain is a valuable research direction.

Critical weaknesses
  1. The paper does not state which theory of causal inference the causal graph is grounded in, nor how the graph itself was constructed. The underlying assumptions and the concrete construction procedure are missing and need to be specified.

‌2. The experimental conclusions are questionable.‌ For example, the abstract claims that GUST is best on three of the four metrics for climate projection and that it achieves "zero forgetting" in the regional setting. These claims contradict the results in Table 2, where GUST is best on only one metric: GUST underperforms the baselines on NRMSEg, Total, and RMSE. Nevertheless, its entries are boldfaced, which is misleading and obscures the actual results.

‌3. The directional transfer and routing analysis lacks experimental support.‌ The claimed directional transfer and routing behavior is asserted rather than demonstrated; no experiments or results are provided to substantiate it.

‌4. Table 19 shows that GUST does exhibit forgetting.‌ The claim of achieving zero forgetting in the regional setting cannot be sustained.

‌5. The experimental baselines are problematic.‌ EASE is mentioned in the experimental setup, but no results for EASE appear in the main experiments.

‌6. The efficiency of the method is not reported.‌

‌7. There are numerous formatting problems.‌ For instance, Figure 3 is unclear.

Questions for authors

Please see the weaknesses.

Reviewer_JCNyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies replay-free continual adaptation of weather and climate foundation models across changing variables, regions, and climate-response targets. GUST freezes task-specific LoRA adapters and combines them through a directed task graph using relation priors, task embeddings, and online transfer feedback. The problem is worthwhile, and the variable-CL results suggest benefits over sequential finetuning.

Critical strengths
  1. The paper addresses a practical problem: adapting weather and climate foundation models as prediction needs evolve without revisiting historical training data. The three continual-learning settings give the study useful breadth.
  2. Keeping past adapters frozen while learning how to reuse them is a reasonable design. Online transfer feedback provides a way to select useful updates rather than relying solely on predefined task similarity.
  3. The variable-CL results are encouraging. Table 3 shows that selective reuse improves over naive adapter composition, while the stage-wise results help expose the trade-off between retaining earlier tasks and learning new ones.
Critical weaknesses
  1. The climate-projection results are internally inconsistent. Table 2's joint-training row reproduces the four target-wise NRMSEs values from Table 20 under four different metric headings. GUST's reported NRMSEg of 7.23 is actually the precipitation NRMSEs in Table 20. These entries also do not match the oracle-normalized averages described in Sec. 5.2. The claimed advantage needs to be reassessed using corrected aggregation.
  2. The regional zero-forgetting claim contradicts the appendix. Applying Eq. (16) to Table 19 gives approximately 6.0% mean forgetting across the five previously learned regions. This is not a rounding difference and directly affects a headline result.
  3. The graph-specific contribution is not sufficiently isolated. Table 3 supports selective reuse over naive composition, but does not separately test the relation prior, utility, and interference terms. A directly learned adapter mixture with the same parameter budget would help establish whether the proposed graph machinery is necessary.
  4. The experimental evidence is limited to one short task order per scenario, and only means over three runs are reported. This matters because transfer is explicitly order-dependent. Data split dates, the variable-CL forecast horizon, and adaptations of classification-oriented baselines to regression also need clearer specification.
Questions for authors
  1. Eq. (7) applies the latest composed update to all observed tasks. What is the exact inference path for an earlier task, and are its prediction head and other task-specific components frozen? Please distinguish preserving stored adapters from preserving earlier predictions.
  2. Can you provide corrected climate and regional summaries with the aggregation code, and test whether the gains over independent LoRA and a directly learned adapter mixture persist across task orders with uncertainty estimates?

Comments on the paper

1 comment

I thank the authors for their clarifications and additional experiments, which substantially address several of my concerns. I maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.