acceptodds
Under review as a conference paper at ICLR 2027

COLA: Complementary Context Retrieval with Utility Calibration for Graph Foundation Model Adaptation

Abstract

This paper studies the problem of few-shot adaptation for graph foundation models (GFMs), which aims to transfer pre-trained GFMs to unseen domains with limited support sets. Previous approaches usually introduce learnable task-aware prompts to adapt GFMs to new domains. Despite the progress, these approaches primarily exploit parametric knowledge in pre-trained models while neglecting rich contextual knowledge embedded in source graphs. Towards this end, we propose a novel approach named Complementary Context Retrieval with Utility Calibration (COLA) for graph foundation model adaptation. The core of our COLA is to retrieve relevant graph candidates from complementary knowledge sources and then refine them into fine-grained contexts guided by their estimated utility for graph representation adaptation. In particular, we first build a graph semantics database and a graph structure database from source domains to capture complementary knowledge. Given a target query, we retrieve relevant graph candidates from these two databases using graph representations and graph kernels, respectively. More importantly, we leverage reinforcement learning with pseudo-target calibration to train a lightweight policy model, which estimates the utility of retrieved graphs to guide fine-grained context selection from abundant candidates. The refined complementary contexts are then incorporated into target graph representations under the guidance of null-aware routing for domain adaptation. Extensive experiments on benchmark datasets validate the effectiveness of the proposed COLA in comparison with extensive baselines. Our code is available at https://anonymous.4open.science/r/COLA-D0DC.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Foundation Models, Retrieval-Augmented Generation, Utility-Aligned Retrieval, Few-Shot Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ivcSOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses few‑shot cross‑domain adaptation for graph foundation models (GFMs). The authors propose COLA, a retrieval‑augmented framework that builds two complementary databases: graph semantic database and graph structural database. Besides, the paper provides theoretical bound analysis for routing generalization, and conducts extensive node‑classification & ego‑graph classification experiments under limited settings. Ablations, hyperparameter sensitivity and efficiency analysis are also provided. Overall this is a technically complete work for retrieval‑augmented graph foundation model adaptation, though several practical and conceptual questions remain to be clarified.

Critical strengths
  1. The paper clearly identifies two key limitations for existing GFM adaptation: semantic‑structural heterogeneity, and conflating retrieval similarity with downstream utility. Distinguishing retrieval relevance and prediction utility is a meaningful conceptual contribution for graph RAG (Sec 1).

2.Theorem 3.1 provides generalization bound for adaptive routing under source‑target distribution shift, formally analyzing oracle gain, source routing loss and total‑variation domain shift; complementarity condition for semantic and structural branches is derived in appendix A.

Critical weaknesses
  1. The benchmark used in the article is a widely accepted historical standard for graph communities. However, though the n-shot setting increases the difficulty, in practical applications, such performance (such as accuracy around 50%) is almost means unacceptable, and not mature for usage in a large number of scenarios which require a robust systems.
  2. Although I agree with the intelligence of GFM in terms of generality, more extensive application scenarios should be used to verify its "generality", such as graph-related QA, even for graph-related agentic workflows. However, what has been confirmed in this article is limited to node classification and ego-graph classification on several limited datasets, and even key traditional tasks like link prediction are missing. I hope to see graph-based models that truly match the meaning of "foundation", to be honest, can complete a large number of different types of tasks, rather than being limited to the mature tasks in the field of graph research that have been studied for decades. This is what I enjoy about reading GFM's article. This is not required for specific score-target methods, they should focus on specific task and increase points. But for GFMs, the story should be wider to be okay.
  3. The main baseline RAG‑GFM also builds semantic‑structural memory but without utility re‑ranking. However, the paper does not add a straightforward ablation variant: applying simple supervised ranking / cross‑attention re‑ranker replacing the proposed RL group‑relative policy, to quantify actual gain brought by reinforcement‑learning policy versus simpler supervised re‑ranking alternatives (Sec4.3 Fig2(b)).
  4. The core conception GFM, is not commonly defined. And the GFMs used are limited. Which makes the scope of this paper are not spread. There are many useful and accptable GFMs including GraphLLMs like GraphGPt, LLaGA, Graph Pattern Machines, etc.
  5. COLA contains structural‑kernel training, dual database construction, RL‑based group‑relative utility policy, null‑aware router. Multiple hyper‑parameters are introduced (walk depth L, candidate pool size M, top‑K selection, RL clipping coefficient, KL‑entropy regularization weights). While hyper‑parameter sensitivity is tested for K and L, sensitivity for RL regularization hyper‑parameters (γ,ξ,ϵclip\gamma,\xi,\epsilon_{clip}) is not reported. It is unclear whether good performance heavily relies on careful hyper‑parameter tuning (Sec3.4, Fig3).
  6. Experiments are conducted on moderate‑size public citation/web/e‑commerce benchmarks. The paper does not analyze memory cost for building large‑scale semantic/structural database when source domains contain millions of graph/node instances. For real‑world large‑scale source corpus, storing embedding for every source instance will bring non‑trivial memory overhead, and approximate nearest‑neighbor retrieval may degrade utility‑calibration performance (Sec4).
Questions for authors

Please address my questions in weakness part.

5 comments

We are truly grateful for the time you have taken to review our paper and for your insightful comments. We address each concern below.

W1. The benchmark used in the article is a widely accepted historical standard for graph communities. However, though the n-shot setting increases the difficulty, in practical applications, such performance (such as accuracy around 50%) is almost means unacceptable, and not mature for usage in a large number of scenarios which require a robust systems.

R1. Thank you for your valuable feedback. COLA focuses on cross-domain few-shot graph adaptation, where a pretrained graph model must transfer to unseen target datasets or semantic domains with very limited labeled data. We follow the LODO-Dataset and LODO-Domain evaluation protocols adopted by RAG-GFM [1], where only one or five labeled instances per class are available for adaptation to unseen target datasets or semantic domains. We clarify the reported results from two perspectives:

  1. Performance under limited supervision and domain shift. Under the LODO protocols, the target dataset or entire semantic domain is excluded from source pretraining, and only one or five labeled examples per class are available for target adaptation. These conditions substantially constrain the target classifier, especially for multiclass prediction under distribution shifts. The difficulty is also reflected in existing methods evaluated under the same protocol. Thus, the absolute accuracies should be interpreted in the context of the limited supervision and cross-domain transfer setting.

  2. Improvements over existing methods. Under identical evaluation protocols, COLA consistently achieves competitive or superior performance compared with representative GFMs. On OGBN-Products, COLA improves upon RAG-GFM from 55.4% to 57.4% in 1-shot node classification and from 64.2% to 66.5% in 5-shot classification. Moreover, increasing supervision from one to five examples per class improves COLA's accuracy from 59.7% to 77.0% on Cora, 41.8% to 57.8% on CiteSeer, and 47.9% to 62.0% on Wiki-CS. The broader label-efficiency analysis in Fig. 3(a–b) further demonstrates that COLA effectively exploits additional target labels while maintaining its source-derived components frozen.

We have clarified the interpretation of the LODO few-shot setting in Section 4.1 and strengthened the discussion of label efficiency and cross-domain performance in Section 4.2.

W2. Although I agree with the intelligence of GFM in terms of generality, more extensive application scenarios should be used to verify its "generality", such as graph-related QA, even for graph-related agentic workflows. However, what has been confirmed in this article is limited to node classification and ego-graph classification on several limited datasets, and even key traditional tasks like link prediction are missing. I hope to see graph-based models that truly match the meaning of "foundation", to be honest, can complete a large number of different types of tasks, rather than being limited to the mature tasks in the field of graph research that have been studied for decades. This is what I enjoy about reading GFM's article. This is not required for specific score-target methods, they should focus on specific task and increase points. But for GFMs, the story should be wider to be okay.

R2. Thank you for your valuable feedback. COLA focuses on improving the cross-domain adaptability of pretrained graph foundation models through utility-calibrated retrieval. Beyond the node and ego-graph classification tasks evaluated in Table 1, we conducted an additional link prediction experiment to examine its applicability to a different graph learning task.

MethodROC-AUC (%) ↑AP (%) ↑
Frozen Backbone92.56 ± 0.1292.47 ± 0.12
Retrieval w/o Utility96.00 ± 0.3295.94 ± 0.27
COLA (Ours)98.09 ± 0.3397.88 ± 0.36

Under three independent edge splits with identical negative samples and downstream training budgets, COLA achieves the best performance on both metrics, outperforming the frozen backbone by 5.53 percentage points in ROC-AUC and 5.41 points in AP. Compared with retrieval without utility calibration, COLA further improves these metrics by 2.09 and 1.94 points, respectively. Notably, the target domain is excluded from source-side training, and the classification-calibrated utility scorers and null-aware router are directly transferred to link prediction without task-specific recalibration. These results provide additional evidence that COLA's retrieval-enhancement mechanisms can benefit graph prediction tasks beyond classification, complementing the cross-domain evaluation in Table 1.

We have reported the link-prediction protocol and results in Appendix D.6.

W3. The main baseline RAG-GFM also builds semantic-structural memory but without utility re-ranking. However, the paper does not add a straightforward ablation variant: applying simple supervised ranking / cross-attention re-ranker replacing the proposed RL group-relative policy, to quantify actual gain brought by reinforcement-learning policy versus simpler supervised re-ranking alternatives (Sec4.3 Fig2(b)).

R3. Thank you for your valuable feedback. To directly quantify the benefits of group-relative policy optimization over simpler supervised re-ranking methods, we conducted an additional controlled comparison with Pairwise MLP and Cross-attention re-rankers on OGBN-Products.

SettingPairwise MLPCross-attentionGRPO (Ours)
Node 1-shot54.955.957.4
Node 5-shot62.764.066.5
Graph 1-shot62.159.562.5
Graph 5-shot69.669.470.7

The results show that GRPO achieves the highest accuracy in all four settings, outperforming Pairwise MLP by 0.4–3.8 percentage points and Cross-attention by 1.3–3.0 points. Since Pairwise MLP and GRPO share the same scorer architecture and training budget, their comparison provides direct evidence of the additional benefit of group-relative policy optimization over supervised re-ranking. Together with the results in Fig. 2(b), these findings support the effectiveness of optimizing context selection through relative downstream rewards rather than relying solely on supervised utility ranking.

We have updated the utility-learning analysis in Section 4.3 and reported the controlled Pairwise MLP and Cross-attention comparisons in Appendix D.4.

W4. The core conception GFM, is not commonly defined. And the GFMs used are limited. Which makes the scope of this paper are not spread. There are many useful and accptable GFMs including GraphLLMs like GraphGPt, LLaGA, Graph Pattern Machines, etc.

R4. Thank you for your valuable feedback. We clarify the definition of GFMs adopted in this work and provide an additional backbone comparison to examine COLA's applicability across different graph modeling architectures.

  1. Definition and scope of GFMs. We consider GFMs as pretrained graph encoders that learn transferable representations for adaptation to unseen domains. As formulated in Eq. (1), COLA operates on representations extracted from a frozen backbone, with its retrieval and routing mechanisms designed for cross-domain adaptation.

  2. Generalization across different backbones. To directly evaluate architectural compatibility, we additionally applied COLA to the Graph Pattern Machine (GPM)[2], which differs substantially from the original backbone by explicitly modeling graph substructure patterns. The results are summarized below.

    BackboneTaskFrozen (%)+ COLA (%)Gain (pp)
    OriginalNode57.4 ± 4.966.2 ± 4.3+8.8
    OriginalGraph62.1 ± 3.968.9 ± 2.2+6.8
    GPMNode78.3 ± 4.081.5 ± 3.4+3.2
    GPMGraph81.7 ± 3.283.7 ± 3.1+2.0

    COLA consistently improves performance with both backbones. In particular, it provides additional gains of 3.2 and 2.0 percentage points on GPM for node and ego-graph classification, respectively. These results provide direct evidence that COLA's retrieval-based adaptation is not restricted to the original backbone architecture.

  3. Coverage of broader GFM architectures. Table 1 further compares COLA against representative graph foundation models, including GraphCLIP, UniGraph, and RAG-GFM. Graph-language models such as GraphGPT and LLaGA involve additional graph-language interfaces that are not evaluated in the current work. We therefore distinguish the demonstrated compatibility across representation-based backbones from broader applicability to graph-language generation architectures.

We have reported the GPM backbone comparison in Section 4.3 and Appendix D.5.

W5. COLA contains structural-kernel training, dual database construction, RL-based group-relative utility policy, null-aware router. Multiple hyper-parameters are introduced (walk depth L, candidate pool size M, top-K selection, RL clipping coefficient, KL-entropy regularization weights). While hyper-parameter sensitivity is tested for K and L, sensitivity for RL regularization hyper-parameters (γ,ξ,ϵclip\gamma,\xi,\epsilon_{clip}) is not reported. It is unclear whether good performance heavily relies on careful hyper-parameter tuning (Sec3.4, Fig3).

R5. Thank you for your valuable feedback. We assessed sensitivity to the GRPO regularization parameters γ\gamma, ξ\xi, and ϵclip\epsilon_{\mathrm{clip}} (Eq. 17) on OGBN-Products 1-shot node classification. We varied one parameter at a time across 13 configurations, using identical source-side training budgets and 10 fixed evaluation splits. The default is (0.01,0.001,0.2)(0.01,0.001,0.2).

ParameterValueAccuracy (%)
Default(0.01, 0.001, 0.2)57.4 ± 6.4
γ\gamma056.3 ± 7.4
0.00556.3 ± 7.7
0.0256.1 ± 7.6
0.0455.8 ± 8.6
ξ\xi055.3 ± 8.7
0.000556.2 ± 7.4
0.00255.7 ± 7.7
0.00455.6 ± 5.6
ϵclip\epsilon_{\mathrm{clip}}0.0554.0 ± 7.1
0.156.7 ± 7.4
0.356.4 ± 7.6
0.454.8 ± 11.2

Relative to the default, accuracy decreases by at most 1.6 pp when varying γ\gamma and 2.1 pp when varying ξ\xi. Clipping values of 0.1–0.3 yield 56.4–57.4% accuracy, while 0.05 and 0.4 perform worse. Thus, performance is relatively stable across the tested regularization settings but more sensitive to extreme clipping values.

We have extended the hyperparameter discussion in Section 4.3 and reported the RL regularization and clipping sensitivity analysis in Appendix D.7.

W6. Experiments are conducted on moderate-size public citation/web/e-commerce benchmarks. The paper does not analyze memory cost for building large-scale semantic/structural database when source domains contain millions of graph/node instances. For real-world large-scale source corpus, storing embedding for every source instance will bring non-trivial memory overhead, and approximate nearest-neighbor retrieval may degrade utility-calibration performance (Sec4).

R6. Thank you for your valuable feedback. We evaluated database storage, million-instance retrieval efficiency, and ANN effects on utility calibration.

1. Storage and Retrieval Scalability. COLA uses 256D semantic embeddings and 32D structural keys (Eqs. 3 and 7). Storage grows linearly with indexed instances. Retained artifacts occupy 0.160 GB (node/ego-graph representations, Exact/IVF indices, and metadata; excluding models and raw graphs), extrapolating to 6.23 GB for one million instances under the same multi-index configuration. We benchmarked FAISS IVF-Flat [3] against exact retrieval on one million synthetic vectors per database, with matching dimensions, the same CPU, and four threads.

Database (1M)Exact (ms)IVF (ms)Recall@M (%)Speedup
Semantic (256D)43.2812.1153.913.58×
Structural (32D)5.401.8095.223.00×

IVF accelerates retrieval by 3.00–3.58×, with lower recall, especially for semantic vectors. Synthetic benchmarks do not measure downstream accuracy at million-instance scale.

2. ANN Impact on Calibration. We compared exact and IVF retrieval on OGBN-Products Home/Tech (5-shot node classification, 10 support splits).

Calibration / InferenceSem. Recall@32 (%)Str. Recall@16 (%)Accuracy (%)
Exact / Exact100.00100.0066.5 ± 2.6
Exact / IVF92.2699.9966.6 ± 3.5
IVF / IVF92.2699.9963.0 ± 3.6

Inference-only IVF preserves accuracy (66.6% vs. 66.5%), whereas IVF during calibration and inference reduces it to 63.0%. This suggests greater calibration sensitivity; however, policy retraining prevents attributing the entire decrease solely to ANN.

We have reported these analyses in Appendix D.8.

Reference

[1] Haonan Yuan, et al. "Rag-gfm: Overcoming in-memory bottlenecks in graph foundation models via retrieval-augmented generation." In Proceedings of the ACM Web Conference 2026, pp. 626–637, 2026b.

[2] Wang, Zehong, et al. "Beyond message passing: Neural graph pattern machine." arXiv preprint arXiv:2501.18739 (2025).

[3] Douze, Matthijs, et al. "The faiss library." IEEE Transactions on Big Data (2025).

In light of these responses, we hope we have addressed your concerns and that you will consider raising your score. If any concern remains, we would be grateful for the opportunity to clarify it further.

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

This paper proposes COLA for few-shot adaptation of graph foundation models to unseen datasets and domains. COLA augments a frozen graph encoder with source instances retrieved from semantic and structural databases. Semantic retrieval uses backbone representations, while structural retrieval uses source-specific encoders based on random-walk descriptors and propagation. Candidate contexts are refined by utility scorers trained through group-relative policy optimization on source-only pseudo-target episodes. A router combines semantic, structural and unchanged backbone representations according to their estimated usefulness. During target adaptation, all source-derived components remain frozen and only a lightweight predictor is fitted. Experiments cover few-shot node and ego-graph classification on five benchmark datasets. The reported results show that COLA achieves promising improvements over the strongest baseline on the evaluated product graphs.

Critical strengths
  1. The method explicitly distinguishes retrieval similarity from downstream usefulness. It provides a concrete way to learn context selection without requiring target query labels for few-shot adaptation.

  2. The experiments compare utility calibration strategies under identical candidate pools and training budgets. Further analyses examine whether predicted utility reflects downstream usefulness and whether semantic and structural contexts provide complementary benefits, supporting the design of utility-guided context refinement.

  3. The paper is well organized and easy to follow.

Critical weaknesses
  1. The semantic and structural databases are already introduced in existing studies (e.g., RAG-GFM), which retrieve from both stores to augment representations. The paper should more precisely distinguish COLA from existing retrieval-based GFMs to clarify its methodological contributions.

  2. Pseudo-target domains are excluded from retrieval during calibration but have already been seen by the pretrained backbone. This mismatch leaves uncertainty about how well the learned utility estimates transfer to genuinely unseen domains. The utility analysis on Cora provides limited evidence for this transfer under stronger domain shifts.

  3. The efficiency comparison excludes RAG-GFM and reports adaptation episodes rather than end-to-end runtime. The additional cost of context retrieval and augmentation should also be included. A direct comparison with similar retrieval-based methods under consistent cost accounting is needed to substantiate the efficiency advantage.

Questions for authors
  1. What are the main methodological contributions of COLA beyond existing retrieval-based GFMs such as RAG-GFM? A more precise comparison would help distinguish the proposed advances from established designs.

  2. How well do the predicted utilities reflect actual downstream gains under stronger domain shifts? Extending the utility analysis to a domain-holdout setting would help assess whether source-only calibration transfers to unseen domains.

  3. The backbone pretraining objective and the representations used to train the calibration predictor need further clarification. Specifying the target early-stopping criterion would also improve reproducibility.

  4. What is the end-to-end target adaptation time when context retrieval and augmentation are included? A direct comparison with RAG-GFM under consistent cost accounting would help establish the efficiency of COLA relative to similar methods.

5 comments

We are truly grateful for the time you have taken to review our paper and for your insightful comments. We address each concern below.

W1 & Q1. Existing retrieval-based GFMs, such as RAG-GFM, already employ semantic and structural databases for representation augmentation. The methodological contributions of COLA beyond these established approaches should be more clearly distinguished.

R1. Thank you for your valuable feedback. COLA addresses two fundamental challenges in retrieval-augmented graph adaptation: retrieval relevance does not necessarily reflect downstream predictive utility, and retrieved evidence may introduce negative transfer when incorporated indiscriminately. These challenges motivate a utility-oriented framework that learns which source contexts benefit prediction and adaptively determines whether their contributions should be incorporated into target representations. The main methodological contributions are threefold:

  1. Utility-oriented hierarchical retrieval. COLA explicitly separates relevance-based candidate retrieval from downstream utility-based context selection (Eqs. 8–14). Rather than treating semantic or structural similarity as sufficient evidence of usefulness, the second-stage refinement prioritizes contexts according to their estimated contribution to downstream prediction.

  2. Source-only group-relative utility calibration. Estimating downstream context utility is challenging when target-domain labels are unavailable. COLA addresses this by constructing source-only pseudo-target episodes and optimizing context selection policies through group-relative downstream rewards (Eqs. 15–17). This enables the utility scorers to learn predictive usefulness without requiring labeled target data during calibration.

  3. Utility-informed null-aware routing. Even utility-selected contexts may contribute differently across target instances, making unconditional representation augmentation potentially harmful. COLA introduces an adaptive router that considers evidence reliability and agreement between semantic and structural branches (Eqs. 18–21). An explicit null route allows the original backbone representation to be retained when retrieval-based enhancement is less beneficial.

These contributions are supported by the component and utility-learning ablations in Fig. 2(a–b), together with the utility and branch-complementarity analysis in Fig. 4. Unlike retrieval-based GFMs that primarily rely on semantic and structural relevance, COLA jointly addresses how context utility is learned, how useful contexts are selected, and when retrieved evidence should influence target representations. This distinction extends beyond the construction of complementary retrieval databases.

We have sharpened the distinction between relevance-based retrieval and utility-calibrated adaptation in the Introduction, with the corresponding mechanisms detailed in Sections 3.4–3.5.

W2 & Q2. The pseudo-target domains are excluded from retrieval during calibration but remain included in backbone pretraining, raising concerns about whether the learned utility estimates generalize to genuinely unseen domains. The utility analysis on Cora alone provides limited evidence of their correlation with actual downstream gains under stronger domain shifts.

R2. Thank you for your valuable feedback. To test whether the utility scorers generalize to unseen domains without backbone exposure to the pseudo-target, we conducted a strict holdout experiment. COLA learns utility preferences from group-relative downstream rewards in source-only pseudo-target episodes. Standard COLA pretrains on Cora, CiteSeer, PubMed, and Wiki-CS; Strict COLA excludes Wiki-CS from pretraining. Both construct retrieval databases from Cora, CiteSeer, and PubMed and calibrate the utility scorers and router on Wiki-CS. Home and Tech remain unseen throughout source-side training.

TaskShotCOLA (Std.)COLA (Strict)Base (Strict)Sem. ρ\rhoStr. ρ\rho
Node157.4 ± 6.454.6 ± 6.149.7 ± 4.70.1440.355
Node566.5 ± 2.666.2 ± 4.357.4 ± 4.90.2340.446
Graph162.5 ± 3.459.0 ± 2.256.5 ± 3.40.1860.371
Graph570.7 ± 2.668.9 ± 2.262.1 ± 3.90.2630.495

Strict COLA outperforms its frozen backbone by 2.5–8.8 percentage points. On the unseen Home/Tech domains, Spearman correlations (ρ\rho) between predicted utility and actual downstream loss reduction are 0.144–0.263 (semantic) and 0.355–0.495 (structural). These positive correlations demonstrate that the source-calibrated utility estimates retain meaningful ranking capability beyond Cora under stronger domain shift, despite the calibration domain being absent from backbone pretraining. Target query labels are used only for post-hoc analysis, not calibration.

We have provided the strict holdout and utility-correlation analysis in Appendix D.1.

W3 & Q4. The efficiency evaluation excludes a direct comparison with RAG-GFM and reports adaptation episodes rather than actual runtime. A consistent comparison that includes target adaptation, context retrieval, and representation augmentation is needed to substantiate COLA's computational efficiency.

R3. Thank you for your valuable feedback. To directly assess the performance–efficiency trade-off, we conducted a controlled comparison with RAG-GFM on OGBN-Products (Home and Tech) under 5-shot node classification. Both methods use the same pretrained backbone and hardware, with target-side timing covering initialization, adaptation, and inference, including retrieval and augmentation where applicable.

MetricRAG-GFMCOLA
Accuracy (%)64.266.5
Initialization (s)1.9611.042
Target Adaptation (s)39.6020.553
Inference (s / 1K queries)0.0090.420
Total Target-side Time (s)41.5722.015
Deployment Storage (GB)0.6410.470

COLA improves accuracy by 2.3 percentage points while reducing measured target-side runtime by 20.6× and deployment storage by 26.7%. Deployment storage includes the model components and retrieval artifacts required for execution, rather than the database alone. Although COLA incurs higher inference latency due to online retrieval and utility refinement, this cost is included in the total runtime and remains below 0.5 ms per query. Source-side training and database construction are excluded as reusable offline operations; thus, the reported speedup characterizes target-side execution rather than the complete training pipeline.

We have expanded the target-side efficiency discussion in Section 4.2 and provided the direct RAG-GFM runtime and storage comparison in Appendix D.3.

Q3. The backbone pretraining objective and the representations used to train the calibration predictor need further clarification. Specifying the target early-stopping criterion would also improve reproducibility.

R4. Thank you for your valuable feedback. COLA follows a stage-wise training procedure, as detailed in Algorithms 1 and 2 of Appendix B, consisting of backbone pretraining, source-only utility calibration, and target few-shot adaptation. We clarify the corresponding training objectives, representations, and stopping criteria as follows:

  1. Backbone pretraining objective. The backbone is pretrained using a neighbor-based contrastive objective rather than supervised classification. Each anchor node is associated with one positive sample and 50 negative samples, and cosine similarity is used to distinguish positive from negative representations. The objective is defined as

    = -\frac{1}{N}\sum_{i=1}^{N} \log \left( \frac{\exp(\cos(z_i,z_i^{+}))} {\sum_{j=1}^{50}\exp(\cos(z_i,z_{ij}^{-}))} \right).$$ The backbone is optimized using Adam with a learning rate of $10^{-5}$, exclusively on the source graphs permitted by the corresponding LODO protocol.
  2. Calibration predictor and representations. For each pseudo-target calibration split, a linear predictor is trained on L2-normalized, 256-dimensional representations extracted from the frozen backbone, without retrieval augmentation. For ego-graph classification, the corresponding pooled graph representations are used. The predictor is optimized using cross-entropy loss and Adam for 400 steps with a learning rate of 10−310^{-3}. Once trained, its parameters remain frozen throughout subsequent utility calibration. The same predictor evaluates both the original and context-enhanced representations, ensuring that the estimated downstream gains reflect changes in retrieved evidence rather than changes in the predictor.

  3. Target adaptation and fixed training budget. During target adaptation, all source-derived components remain frozen, and only the downstream predictor is optimized using the labeled support set. We use Adam for a fixed budget of 400 adaptation steps without early stopping. Target query labels are used exclusively for final evaluation and do not influence model optimization or model selection.

We have specified backbone pretraining, calibration-predictor construction, and fixed-step target adaptation in Appendix C.1 and aligned the training description in Appendix B.

In light of these responses, we hope we have addressed your concerns and that you will consider raising your score. If any concern remains, we would be grateful for the opportunity to clarify it further.

Dear Reviewer ptsJ,

We thank you for your careful reviews and constructive suggestions. With the rebuttal-discussion period coming to a close, may we kindly ask you to review them and confirm if they have effectively addressed your concerns? Any further insights or comments you wish to share would be warmly welcomed. We extend our heartfelt gratitude for the time and effort you've invested in reviewing our work.

Best regards,

the Authors

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

This paper proposes COLA, a retrieval-augmented approach for graph foundation model adaptation. The method constructs semantic and structural databases from source graphs and retrieves useful contexts through utility calibration. It then uses Null-aware Router to fuse the retrieved context with the graph representation. Experiment are conducted on five graph datasets. The results show COLA can improve the performance on node and graph classification.

Critical strengths

S1. This paper studies a meaningful problem: how to utilize source graphs to adapt to unseen domains with few labeled samples.

S2. The method is clear and easy to understand. It is an interesting idea to distinguish retrieval relevance from downstream task utility.

S3. The experiments cover several different transfer settings and compare COLA with many baselines.

Critical weaknesses

W1. Pseudo-target calibration might not fully simulate adaptation to truly unseen domains. The authors exclude the pseudo-target graph from the retrieval database, but the backbone has already been pretrained on this graph. Therefore, the backbone may have learned information about the pseudo-target, which could affect the generalization of the utility scorers and router to unseen domains. I believe excluding pseudo-target graph from pre-training will make the experiment more convincing.

W2. According to Eq 3 and 7, althogh both semantic and structural branches use different retrieval methods, they both use backbone representations as retrieved contexts. However, the full model may retrieve more samples than any single-branch one. Therefore, it is unclear how much of the improvement comes from omplementary structural retrieval. I believe a comparison under the same total retrieval budget would make the claim of structural complementarity more convincing.

W3. Compared to RAG-GFM, COLA achieves only small improvement in many settings. However, COLA may require more computational resources for model training and database construction. I understand this might be a trade-off, but I wander know whether these performance gains justify the additional computational cost.

Questions for authors

See W1~3

5 comments

We are truly grateful for the time you have taken to review our paper and for your insightful comments. We address each concern below.

W1. Pseudo-target calibration might not fully simulate adaptation to truly unseen domains. The authors exclude the pseudo-target graph from the retrieval database, but the backbone has already been pretrained on this graph. Therefore, the backbone may have learned information about the pseudo-target, which could affect the generalization of the utility scorers and router to unseen domains. I believe excluding pseudo-target graph from pre-training will make the experiment more convincing.

R1. Thank you for your valuable feedback. To directly examine the generalization of utility calibration to genuinely unseen domains, we conducted an additional strict domain-holdout experiment in which the pseudo-target domain (Wiki-CS) is entirely excluded from backbone pre-training. Specifically, Cora, CiteSeer, and PubMed are used for backbone pre-training and retrieval database construction, while Wiki-CS is used exclusively for source-side utility and router calibration. OGBN-Products remains unseen throughout source-side training. This setting allows us to evaluate COLA without prior backbone exposure to the calibration domain.

TaskShotCOLA (Std.)COLA (Strict)Base (Strict)Sem. ρ\rhoStr. ρ\rho
Node157.4 ± 6.454.6 ± 6.149.7 ± 4.70.1440.355
Node566.5 ± 2.666.2 ± 4.357.4 ± 4.90.2340.446
Graph162.5 ± 3.459.0 ± 2.256.5 ± 3.40.1860.371
Graph570.7 ± 2.668.9 ± 2.262.1 ± 3.90.2630.495

Under strict domain holdout, COLA maintains strong performance across all four settings, consistently outperforming the corresponding frozen backbone by 2.5–8.8 percentage points. Furthermore, the predicted semantic and structural utilities remain positively correlated with actual downstream gains, indicating that the learned utility scorers retain meaningful ranking capability even when the calibration domain is unseen during backbone pre-training. These findings demonstrate that COLA continues to provide effective cross-domain adaptation without relying on prior backbone exposure to the pseudo-target domain.

We have reported the strict domain-holdout results and utility correlations in Appendix D.1.

W2. According to Eq 3 and 7, althogh both semantic and structural branches use different retrieval methods, they both use backbone representations as retrieved contexts. However, the full model may retrieve more samples than any single-branch one. Therefore, it is unclear how much of the improvement comes from omplementary structural retrieval. I believe a comparison under the same total retrieval budget would make the claim of structural complementarity more convincing.

R2. Thank you for your valuable feedback. COLA is designed to capture complementary source evidence through two distinct retrieval criteria: semantic similarity in the pretrained representation space and structural correspondence measured by the learned graph kernel (Eqs. 8–9). To further validate that this complementarity translates into performance gains beyond simply using more contexts, we conducted a controlled comparison with a fixed total budget of 10 selected contexts. Specifically, Semantic-only and Structural-only each select 10 contexts, while Full COLA selects 5 from each branch. All variants follow the same evaluation protocol.

TaskShotSemantic-only (10)Structural-only (10)Full COLA (5+5)
Node154.8 ± 3.454.3 ± 2.957.4 ± 6.4
Node561.4 ± 3.560.9 ± 3.366.5 ± 2.6
Graph158.4 ± 4.356.6 ± 3.962.5 ± 3.4
Graph565.3 ± 3.660.2 ± 2.570.7 ± 2.6

Under the same total context budget, Full COLA consistently outperforms both single-branch variants across all four settings, improving accuracy by 2.6–5.4 percentage points over Semantic-only and 3.1–10.5 points over Structural-only. These results demonstrate that combining semantic and structural retrieval provides benefits beyond simply increasing the number of selected contexts.

We have clarified the fixed-context-budget analysis in Section 4.3 and provided the controlled comparison in Appendix D.2.

W3. Compared to RAG-GFM, COLA achieves only small improvement in many settings. However, COLA may require more computational resources for model training and database construction. I understand this might be a trade-off, but I wander know whether these performance gains justify the additional computational cost.

R3. Thank you for your valuable feedback. To assess the performance–efficiency trade-off, we conducted a direct comparison with RAG-GFM on OGBN-Products under 5-shot node classification. Both methods use the same pretrained backbone and were evaluated under identical hardware and timing conditions.

MetricRAG-GFMCOLA
Accuracy (%)64.266.5
Initialization (s)1.9611.042
Target Adaptation (s)39.6020.553
Inference (s / 1K queries)0.0090.420
Total Target-side Time (s)41.5722.015
Deployment Storage (GB)0.6410.470

COLA improves accuracy by 2.3 percentage points while reducing the measured target-side execution time by 20.6× and deployment storage by 26.7%. The storage comparison accounts for the model components and retrieval artifacts required for deployment, rather than database files alone. Although COLA incurs higher inference latency due to online retrieval and utility refinement, the additional cost remains below 0.5 ms per query. Source-side training and database construction are excluded from this timing comparison, as these offline operations produce reusable components that remain frozen during target adaptation. Thus, the results demonstrate a favorable accuracy–efficiency trade-off at the target adaptation stage.

We have reported the direct RAG-GFM comparison in Appendix D.3.

In light of these responses, we hope we have addressed your concerns and that you will consider raising your score. If any concern remains, we would be grateful for the opportunity to clarify it further.

Dear Reviewer buzE,

We thank you for your careful reviews and constructive suggestions. With the rebuttal-discussion period coming to a close, may we kindly ask you to review them and confirm if they have effectively addressed your concerns? Any further insights or comments you wish to share would be warmly welcomed. We extend our heartfelt gratitude for the time and effort you've invested in reviewing our work.

Best regards,

the Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.