Dynamic Bundling with Large Language Models for Zero-Shot Inference on Text-Attributed Graphs

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following.

Q1. The significance of this paper is limited. There are some technical concerns. Please refer to questions for details.

A1. Thank you for the comment. The contribution of this paper is three-fold:

• The idea of bundling nodes/texts: We propose to bundle relevant nodes and text attributes, and use these bundles instead of individual nodes to query LLMs and supervise GNNs.
• Unified Framework: We propose a novel framework consisting of bundle sampling, bundle query, bundle supervision, and bundle refinement. The proposed framework outperforms a number of baselines.
• Theoretical Analysis: We provide rigorous theoretical analysis of our method, showing its tolerance to outlier nodes and the convergence properties.

We will address your technical concerns in the following.

Q2. The theoretical analysis focuses on bundle, which is quite limited as supervised tasks are mainly performed by Large language models.

A2. Thank you for the comment. The bundle is an important concept of this paper. Bundles are used for supervising the GNN, which is then used for the downstream task. The two theorems in this paper focus on bundle supervision, which is one of the core components of this paper.

Q3. Can a text bundle be interpreted as context window size? Any insight? If a bundle is interpreted as a proxy for this window—composed of graph nodes and their descriptions—this has important implications for scaling, quality of encoding, and model performance.

A3. Thank you for the question. A text bundle is not the context window size. In addition to providing extra information, text bundling has the following advantages. Firstly, text bundling reduces the task difficulty for LLMs, as predicting the overall property of a set of closely related text items is easier than predicting one of them. This makes the supervision signals more robust. Secondly, text bundling is more efficient in supervising GNNs, as all nodes in a bundle can provide supervision signals.

Q4. Any other way to tell graphs with high homophily or heterophily other than graph type? What to do with medium homophily and heterophily? The design of the bundle sampling mechanism may need to be adaptive, possibly using homophily-aware selection or weighting strategies.

A4. Thank you for the question. The homophilic ratio can be estimated by sampling a set of connected node pairs and directly asking the LLM whether they belong to the same category. Below are the results for ground truth homophily and the prediction by GPT4o-mini. The predicted results are close to the ground truth, which is enough to roughly decide the type of bundle sampling method to use.

We also provide an adaptive approach of bundle sampling, where we take the union of topological sampling and semantic sampling, with the ratio of each determined by the predicted homophily. The results are shown below, where we can see that adaptive sampling also outperforms the baseline.

Q5. The paper states: “During optimization of the graph neural network gθ, the bundle B may include nodes that do not belong to the category of yˆB.” Given that this is a supervised learning setup and labels are available during training, it’s unclear why the bundles are not explicitly constructed to include only same-class nodes. Such a strategy could improve signal consistency in the text prompts and reduce label noise within the bundle. If there are limitations that prevent such a strategy—such as low label density or design constraints—these need to be explicitly stated. Otherwise, this decision appears to introduce noise and reduces the clarity of the supervision signal.

A5. Thank you for the comment. We want to clarify that the setting of this paper is zero-shot inference instead of supervised learning. In zero-shot inference, we do not have access to the ground truth labels.

Q6. "Bundle" seems to be the central component of the proposed method: (1) Edges in a bundle is implicitly input to Large language models as graph structure information (2) Multiple nodes in a bundle provides a large context, which will reduce hallucination However, more nodes are preferred in a bundle to incorporate more structural information, which also increases "outliers". As the core component, analysis in 4.4 is not sufficient. Structural information incorporated in this way is quite limited.

A6. Thank you for the comment. There is a balance between incorporating more information and excluding outliers. We empirically find that the bundle size of 5 is optimal in terms of overall accuracy. Additionally, by comparing bundle query and individual query in Section 4.3, we show that bundling improves performance. We also provide an additional experiment where we manually include one outlier into the bundles. The results are shown below, and we can see that our method is robust to outliers, which is also demonstrated by Theorem 3.1.

Q7. It is better to include dataset split details directly in the paper instead of referring to other papers.

A7. Thank you for your suggestion. As our setting is zero-shot inference, there is no training set, and we list the test set split details as follows.

Q8. Performance shown in experiments often is far below traditional GNN methods (e.g., GCN achieves 81.5% accuracy on Cora, GraphSage achieves 83.67% on Wiki CS). As the proposed LLMs-based approach does not modify LLMs in any way, simple prompt-level work may have limitations. Please justify the motivation using LLMs, just because they are popular?

A8. Thank you for your question. The setting of this paper is zero-shot inference, while the performance you mentioned (81.5% and 83.67%) is semi-supervised learning. LLMs have strong zero-shot generalization ability, and they can understand the semantics of text attributes and class labels, enabling preliminary classification, which is beyond the capabilities of traditional GNNs.

In light of these responses, we hope that we have addressed your concerns and that you will consider raising your score. If there are any additional notable points of concern that we have not yet addressed, please do not hesitate to share them, and we will promptly attend to those points.

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following.

Q1. The core idea of bundling similar items for better LLM performance is not particularly novel. The contribution is more incremental, combining existing techniques (LLM querying + GNN training) rather than introducing fundamentally new concepts;

A1. Thank you for the comment. The contribution of this paper is three-fold:

• The idea of bundling nodes/texts: We propose to bundle relevant nodes and text attributes, and use these bundles instead of individual nodes to query LLMs and supervise GNNs.
• Unified Framework: We propose a novel framework consisting of bundle sampling, bundle query, bundle supervision, and bundle refinement. The proposed framework outperforms a number of baselines.
• Theoretical Analysis: We provide rigorous theoretical analysis of our method, showing its tolerance to outlier nodes and the convergence properties.

Our method is not a simple combination of existing techniques. We introduce the new concept of bundles, while prior studies adopt individual supervision on GNNs.

Q2. The approach treats LLMs and GNNs as separate components rather than achieving true integration. The LLM serves merely as a labeling tool, and there's limited synergy between the two components;

A2. Thank you for the comment. This paper explores graph annotation from the perspective of bundles, which are used to facilitate downstream GNN training.

Q3. The bundle refinement process is quite mechanical, removing only the least confident nodes at fixed epochs (300 and 400), which may not be optimal for all scenarios.

A3. Thank you for the comment. As the bundle size of our paper is 5, removing one item is a reasonable choice. We also provide the results of removing more items below. For the epochs, we also provide results of alternative configurations. The results suggest the current default configuration achieves satisfactory performance.

Q4. Although the experimental results were comprehensive and effective, I still have doubts about the reliability of the experimental results (as shown in table 1 and table 3). Table3 v2 and v5 correspond to single-node annotation and individual supervision respectively. Their results are also superior to most baselines, which seems to be contrary to the core contribution of this paper.

A4. Thank you for the comment. For V2, our method achieves significant improvement compared to individual querying. For V5, while it uses $\mathcal L_{IE}$ (defined in Theorem 3.1), it still adopts the overall framework of DENSE, including bundle sampling/query, ranking-based supervision, and bundle refinement. Many baseline methods suffer from poor generalization capability, leading to low accuracy under the zero-shot setting. For example, GOFA and ZeroG adopt joint training of LLM and graphs. Their performance worsens when the test distribution is different from the training distribution.

Q5. Regarding weaknesses 4, could you clarify if the experimental settings are consistent across all comparisons? What specific configurations were used for baseline methods?

A5. Thank you for the question. The experimental setting is zero-shot inference on TAGs without access to node labels, and it is consistent across all comparisons. We provide the details as follows:

• Text encoders (SBERT, RoBERTa, TE-3-Large, LLM2Vec): The text attributes and class labels are encoded by the text encoders, and the nodes are assigned according to the cosine similarity between their text embeddings and class embeddings.
• Generative LLMs (GPT-3.5-Turbo, GPT-4o): The class labels are obtained by directly prompting the LLMs.
• Graph SSL methods (DGI, GraphMAE): The implementation follows the TSGFM [1] benchmark, and the models are trained on the ogbn-arxiv dataset.
• OFA, ZeroG, GraphGPT, LLAGA are trained on the ogbn-arxiv dataset.
• For GOFA, the pretrained weights are adopted, and the ogbn-arxiv dataset is used for finetuning.
• For UniGLM, the pretrained model is directly adopted for inference.
• For LLMBP, it is a pretraining-free solution, and we directly use their official implementation.

[1] Text-space graph foundation models: Comprehensive benchmarks and new insights. Chen et al. NeurIPS 2024.

Q6. The method proposed in this paper is a kind of LLMs for Data Augmentation, but the baselines do not include similar experimental settings.（such as Chen, Z., Mao, H., Wen, H., Han, H., Jin, W., Zhang, H., Liu, H., and Tang, J. Label-free node classification on graphs with large language models (llms). ICLR, 2024e.）

A6. Thank you for bringing this excellent work to our attention! We will include this work in the revised version. We want to clarify that our method is not data augmentation. We provide a comparison between our method and Chen et al. as follows:

Q7. Why is the bundle refinement limited to only two fixed epochs (300 and 400) with removal of just the least confident node? Have you explored more adaptive refinement strategies or different refinement frequencies? What is the sensitivity to these hyperparameters?

A7. Thank you for the question. For the time of refinement, the general principle is to perform refinement after the training converges. In practice, we observe that the bundle supervision converges within 300 epochs, and that after refinement, it converges within an additional 100 epochs. Additionally, since the optimal bundle size is 5 (Section 4.4), removing one item each time is reasonable. We also provide results of different configurations in A3. The model's performance is not very sensitive to the configuration.

Q8. For heterophilic graphs, you use semantic proximity for sampling, but this essentially ignores the graph topology entirely. How does this approach truly address the claimed limitation of "limited information on graph structure"? Could you provide more analysis on how graph structure is actually utilized?

A8. Thank you for the question. For the homophilic graphs, graph topology is important for understanding the semantics of nodes, and we incorporate the topological information in bundle sampling. For heterophilic graphs, the topology provides less information, and we find semantic proximity more effective in bundle sampling.

Q9. The method requires multiple LLM queries (100 bundles × datasets). Could you provide analysis of the computational cost and scalability compared to baselines?

A9. Thank you for the question. We provide the computational cost and scalability on the Cora dataset below. The results show that our method is better than the baseline. The computational cost is generally not a concern, as GNNs are lightweight models and our method maintains good performance even with cheaper LLMs (e.g. DeepSeek-V3 or Gemini-2.5-flash in Table 2).

In light of these responses, we hope that we have addressed your concerns and that you will consider raising your score. If there are any additional notable points of concern that we have not yet addressed, please do not hesitate to share them, and we will promptly attend to those points.

Thank you for your detailed rebuttal. I have reviewed the authors' responses, and they have addressed most of my concerns. Based on the authors' clarifications in the rebuttal, I will slightly increase my original rating.

We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! In the following response, we would like to address your major concern and provide additional clarification.

Q1. Scope expansion: Method currently requires textual attributes; exploring LLM-generated pseudo-descriptors for non-text-attributed graphs could broaden applicability.

A1. Thank you for your suggestion! We will add the following discussion to the revised manuscript: The proposed DENSE framework can be extended to a more general setting where text attributes are not provided. Non-text-attributed graphs can be converted to text-attributed graphs, which have been explored by Wang et al. [1]. We will introduce these techniques to expand the scope of our method in future work.

[1] Can LLMs Convert Graphs to Text-Attributed Graphs? Wang et al. 2024

Q2. Heterophily refinement: While semantic sampling handles heterophily, explicitly prompting LLMs to resolve topological proximity vs. semantic conflicts (e.g., "identify dominant label in semantically divergent bundles") may further boost robustness.

A2. Thank you for your suggestion! We will extend our work by exploring the prompting techniques in bundle query when handling heterophilic graphs, explicitly informing the LLMs about the homophilic/heterophilic information of the current graph to improve the robustness of our method.

We will properly incorporate your suggestions into our revised version. Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions.

We are truly grateful for the time you have taken to review our paper, your insightful comments and support. Your positive feedback is incredibly encouraging for us! In the following response, we would like to address your major concern and provide additional clarification.

Q1. The authors adopt two bundle sampling techniques: topological proximity and semantic proximity, according to graph homophily. In practice, it may be non-trivial to determine the homophily degree of graphs, and this paper lacks discussion on this. The authors should discuss how to decide graph homophily when it is not obvious from the graph's meta information.

A1. Thank you for your suggestion! We will discuss the graph homophily in the revised version. Specifically, one approach is to sample a set of connected node pairs and directly ask the LLM whether they belong to the same category. Below are the results for ground truth homophily and the prediction by GPT4o-mini.

Q2. The motivation of the proposed ranking-based supervision is not very clear. Why do you use both entropy-based and ranking-based supervision? Why using a min operator? The authors should explain the rationale behind this technique.

A2. Thank you for your question! We use entropy-based supervision because nodes in the bundle are more likely to fall into the mode category of the bundle. Since there may be nodes from other categories, the ranking-based supervision is designed to penalize bundles in which the majority of nodes are not classified into the mode category. The min operator is used to achieve this functionality. We will provide more explanation of the rationale in the revised version.

Q3. The authors only use one GNN architecture for homophilic graphs, while there are many GNN architectures available. The authors should perform experiments on different types of GNNs like GraphSAGE or GIN to show the influence of GNN architectures.

A3. Thank you for your suggestion! We provide additional results using GraphSAGE as follows. The results suggest that our method achieves good performance with different GNNs.

We will properly incorporate your suggestions into our revised version. Thanks again for appreciating our work and for your constructive suggestions. Please let us know if you have further questions.

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following.

Q1. The exploration of heterophily is limited. While the paper includes heterophilic graphs (Cornell, Texas, Wisconsin, Washington), the mechanism for handling them relies solely on switching from "topological proximity" to "semantic proximity" for bundle sampling. This semantic sampling is based on the initial text embeddings. The paper could benefit from a deeper exploration of how the bundling strategy interacts with strong heterophily, where even semantically similar nodes might belong to different classes. The performance gains on heterophilic graphs, while strong, might stem more from the power of the base GNN designed for heterophily (GloGNN) rather than the bundling strategy itself being optimally adapted for this challenging scenario.

A1. Thank you for the comment. While some semantically similar nodes might belong to different classes, they are generally close in the label space. Additionally, the proposed bundle supervision is tolerant to outliers (Theorem 3.1). As for the performance gain, we compare the accuracy of individual query and bundle query below. The results suggest that the bundling strategy using semantic proximity is effective.

Q2. The method depends on High-Quality Initial Embeddings: The "semantic proximity" sampling, crucial for heterophilic graphs, relies on the quality of initial node embeddings generated by a text encoder. The paper uses a task-adaptive embedding method from a prior work. If these initial embeddings are poor or do not capture the relevant semantic nuances for classification, the resulting bundles could be noisy from the outset, potentially hampering the entire process. The paper does not analyze the sensitivity of the method to the quality of these initial text representations.

A2. Thank you for the comment. We show that our method is robust to the quality of initial text embeddings. Specifically, for embedding $\textbf x$, we introduce noisy using $\textbf x \leftarrow (1 + \beta \textbf n)\odot \textbf x$, where $\textbf n\sim\mathcal N(\textbf 0,\textbf I)$ is gaussian noise. We show the performance of our method under different noise ratios $\beta$ on the WikiCS dataset below. As can be seen from the results, our method is robust when the quality of text embeddings degrades.

Q3. The experiment lacks sufficient large-scale datasets for evaluation. Most datasets used for evaluation are relatively small. Four of the ten datasets (Cornell, Texas, Wisconsin, Washington) have fewer than 300 nodes each, and two others (Cora, CiteSeer) are also considered small by modern standards. While the inclusion of datasets like Sportsfit (173k nodes) is commendable, the method's scalability and effectiveness on truly large-scale graphs with millions of nodes remain under-explored. The default setting uses 100 bundles for supervision, which may provide very sparse coverage on a much larger graph, potentially requiring a significant increase in costly LLM queries to achieve good results, as evidenced by the finding that more bundles yield better performance.

A3. Thank you for your comment. We provide results on ogbn-products with 2,449,029 nodes and 61,859,140 edges. The baseline LLMBP uses 940 queries (default setup), while the proposed DENSE uses 100 and 200 queries. The results suggest that DENSE outperforms the baseline with only 100 bundles (default setup, much less than the baseline), and the performance further improves with additional bundles.

In light of these responses, we hope that we have addressed your concerns and that you will consider raising your score. If there are any additional notable points of concern that we have not yet addressed, please do not hesitate to share them, and we will promptly attend to those points.