acceptodds
Under review as a conference paper at ICLR 2027

AgenticGNN: A Spectral-Inspired Agentic Framework for Automated Graph Learning Algorithm Design

Abstract

Existing large language model (LLM)-based frameworks for automated graph learning algorithm design typically rely on static performance feedback to conduct black-box iterative optimization. Consequently, they fail to establish explicit associations between target-graph features and algorithm design principle, leading to suboptimal performance and poor cross-task knowledge transfer. Inspired by spectral graph theory, which reveals the intrinsic filtering behaviors of graph learning operators via spectral energy changes, we propose AgenticGNN, an agentic framework for automated graph learning algorithm design that integrates multi-level graph feature perception, adaptive spectral-theory-guided context construction, and spectral-aware evolutionary feedback. AgenticGNN first quantifies target-graph characteristics at the neighborhood, community, and global-topology levels, and translates these measurements into an adaptive natural-language context informed by spectral graph theory. This context guides an LLM to generate candidate graph learning algorithms through structured reasoning. During iterative evolution, AgenticGNN analyzes spectral changes to attribute the effects of design decisions on model behavior. The resulting graph features, design rationales, generated code, and evaluation outcomes are distilled into a progressively evolving knowledge base, enabling explicit and reusable mappings from graph characteristics to algorithm design strategies. Experimental results on nine real-world datasets demonstrate that AgenticGNN consistently outperforms state-of-the-art automated design baselines, achieving superior task performance, higher design efficiency, and robust cross-task generalization. The source code is available at https://anonymous.4open.science/r/AgenticGNN-FDCC.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Representation Learning, Automated Machine Learning, Large Language Models, Agentic Systems, Spectral Graph Theory

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_3PSnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

AgenticGNN uses graph-feature perception, spectral context, and input/output spectral feedback to guide evolutionary generation of GNN propagation operators by an LLM. A knowledge base stores graph features, design rationales, and evaluation experience for cross-task reuse. With the surrounding backbone and training procedure fixed, the framework searches propagation operators and evaluates them on multiple node-classification datasets, with component and knowledge-base ablations.

Critical strengths

1.Graph features and spectral-change feedback provide targeted guidance for operator search, with initial support from the ablation experiments. 2.The method is competitive across the evaluated datasets. Knowledge-base experiments include a single-task control and exclude prior experience from the target dataset.

Critical weaknesses

1.Appendix F.4 does not construct the claimed cubic graph propagation, while F.9’s Fourier band-pass branch alone is a feature-channel linear map. These mismatches weaken the spectral-mechanism and interpretability claims. 2.The paper indicates validation feedback, but the candidate-ranking, early-stopping, and final-selection procedures still need clarification. Label masking for class-conditioned path statistics and whether repetitions include independent searches also remain unclear, limiting assessment and reproducibility.

Questions for authors

1.Are Appendices F.4 and F.9 the implementations used for the main results? Please provide the corresponding runnable versions and explain how repeatedly propagating the same input yields cubic graph filtering, and how the node-wise linear map implements Fourier band-pass filtering. If these are documentation errors, do corrections change the mechanism analysis or results? 2.Which labels are visible to every graph descriptor, especially class-conditioned path lengths and adjusted homophily? How are absent classes and unreachable or empty class pairs handled with very few training labels? Please specify the splits used for candidate fitness, early stopping, and final selection, and the exact level at which repetitions are independent. 3.At matched candidate-evaluation budgets and across independent searches, does genuine spectral feedback reliably outperform shuffled spectral feedback and performance-only feedback on test results selected through validation? Does cross-task memory still help under the main sparse split and full search budget? The existing knowledge-base experiment uses a different split and a shorter search budget, so it does not directly establish this.

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

This paper presents AgenticGNN, an LLM-driven evolutionary framework for automatically designing graph neural network operators. Within a fixed network backbone and training pipeline, it generates and evaluates candidate operators using target-graph descriptors, spectral-energy changes, and predictive performance as feedback. A cross-task knowledge base stores distilled design experience to guide subsequent searches. Experiments on nine node-classification benchmarks show competitive overall performance, with larger gains on heterophilic graphs.

Critical strengths
  1. Practical automation of operator design. The framework provides a working end-to-end loop for generating, training, and refining GNN operators, potentially reducing the manual effort required to develop task-specific architectures.
  2. Combining graph descriptors with spectral-energy feedback gives the LLM more informative guidance than accuracy alone. Ablations support the utility of these components.
Critical weaknesses
  1. Limited methodological novelty and insufficient graph-specific insight. The central generate–evaluate–select–refine loop closely follows established LLM-driven evolutionary design. AutoSGNN already generates spectral GNN operators using mutation, crossover, preference-based prompting, elite selection, and parallel evaluation. The incremental contribution therefore rests on graph descriptors, spectral feedback, and experience retrieval. These primarily incorporate established tools into the search context, without establishing a new filtering principle or a validated relationship between graph characteristics and effective operator designs.
  2. An insufficiently established cost–benefit trade-off. The default search evaluates approximately 240 candidates, requiring repeated GNN training, LLM generation, and spectral analysis. Yet improvements over AutoSGNN are only 0.03 percentage points on Cora and 0.04 on PubMed, with a slight decrease on Photo. Although several heterophilic graphs show larger gains, per-generation timing does not establish overall cost effectiveness. Total GPU-hours, LLM usage, knowledge-base construction costs, and accuracy under matched computational budgets are needed to substantiate the efficiency claims.
  3. Experimental comparisons and evaluation details leave significant uncertainty. Many baseline means and uncertainties exactly match AutoSGNN’s published results, without clearly distinguishing reused results from fresh runs. AgenticGNN uses GPT-4o, whereas the published AutoSGNN configuration uses GPT-3.5-turbo, leaving the generation backbone as an important confounding factor. Moreover, knowledge-base ablations switch from the main 2.5%/2.5%/95% split and 20 iterations to a 60%/20%/20% split and five iterations. \
    Additional concerns remain unresolved. The reported repetitions do not clearly separate search, training, and split variability. Chameleon and Squirrel are evaluated without explaining whether known duplicate-node issues were addressed.
  4. The anonymous repository contains only a README, with no executable framework, evaluation scripts, configurations, prompts, or search records. This conflicts with Section 7’s statement that all code and datasets have been made available.
  5. The manuscript contains substantial unfinished material and reporting inconsistencies. Section 4.2 claims approximately five percentage points of improvement over the previous best method on Chameleon, although Table 1 shows only 1.83 points over SANE. Section 7 retains unfilled placeholders, including “Your contributiong” and “Specific dataset names,” alongside malformed punctuation. Appendix B incorrectly lists PubMed as having five classes. Appendix C describes power iteration and Hotelling deflation, contradicting the Lanczos procedure specified in Section 3.1. Finally, Figure 7 leaves its curves identified only as “Method A/B/C.” Together with frequent grammatical errors, these issues indicate insufficient verification and materially reduce confidence in the paper’s precision.
  6. Several spectral interpretations are unsupported or inconsistent with the supplied code. Appendix F.6 labels one-, two-, and four-hop normalized-adjacency propagation as low-, middle-, and high-frequency channels. However, propagation depth does not define frequency bands: the four-hop branch performs repeated diffusion, rather than an established high-pass operation. Appendix F.9 describes adaptive Fourier band-pass filtering, while its implementation is simply torch.mm(x, self.fourier_weights), a feature-channel projection without an explicit graph-frequency-selective operation.
Questions for authors

See weaknesses.

Reviewer_8QoHOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AgenticGNN, an LLM-driven framework for automated GNN operator design. It incorporates graph structural and feature statistics into prompts and uses predictive performance and spectral changes in node representations to guide iterative refinement. A knowledge base stores and retrieves design experience to support subsequent searches. Experiments on nine node classification datasets show strong overall performance, with ablations examining the contributions of the main components.

Critical strengths
  1. The motivation is clear. The method incorporates structural and feature statistics of the target graph, along with spectral changes in candidate models, into LLM-driven GNN operator design, providing richer feedback for iterative optimization than a single performance score.

  2. The connection between the method and the experiments is reasonably clear. The ablation studies cover the main components, including graph characteristics, spectral feedback, and the knowledge base, providing evidence for their effectiveness.

Critical weaknesses
  1. The discussion of closely related work is insufficient. For example, DesiGNN (Proficient Graph Neural Network Design by Accumulating Knowledge on Large Language Models) already combines graph property understanding, retrieval of prior design knowledge, and architecture refinement. The paper should clarify how it differs from this work in motivation, mechanism, and contribution.

  2. The allocation of space in the methodology section could be improved. Sections 3.1 and 3.3 devote considerable space to basic graph metrics and standard computational procedures, whereas the context construction in Section 3.2 and the generation strategy in Section 3.5 are described relatively briefly. The paper does not sufficiently explain how this information guides specific design decisions, leaving the core novelty insufficiently highlighted.

  3. The experimental analysis is not sufficiently comprehensive. Although the method ranks first overall in the main experiments, its improvements on some datasets are marginal. Performance degradation also occurs in the knowledge base and cold-start experiments, for example, on Computers and PubMed in Table 4 and Computers in Table 5. The main text does not adequately discuss these results or the conditions under which the method is effective.

  4. The knowledge base ablations in Table 4 and Figure 4 change both the label split and the search budget relative to the main experiments, moving from fewer labeled samples and twenty search rounds to more labeled samples and five search rounds, without adequately explaining the rationale. Consequently, it remains unclear whether the knowledge base is effective under the main experimental settings and whether its benefits arise from faster progress in the early search stages or improvements in final performance.

  5. The manuscript needs further checking. Section 7, Reproducibility Statement, still contains template placeholder text, and the linked code repository is empty as of this review, preventing verification of implementation details. The markings for the best and second-best results in Table 1 should also be checked for consistency. Figure 2 and Table 1 present largely overlapping information; consolidating them or moving one to the appendix could free up space for a clearer explanation of the core mechanisms.

Questions for authors
  1. How is spectral feedback translated into specific suggestions for modifying operators? Could the authors provide an actual example of operator evolution to illustrate what additional information it provides beyond accuracy or simpler smoothness metrics?

  2. How do the authors explain the cases where the knowledge base does not improve performance in Table 4, as well as the substantial performance drop on Computers in Table 5? What limitations of the method do these results reveal?

  3. How was the twenty-round search budget in the main experiments determined? Why do the knowledge base ablations use a different label split and a five-round budget? Could the authors provide complete search curves with and without the knowledge base under the main experimental settings to distinguish search acceleration from improvements in final performance?

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

AgenticGNN proposes an LLM-based framework for graph AutoML composed of spectral graph metrics, a knowledge base for storing previous experimental results and design insights, and an LLM for proposing new designs.

Critical strengths

The study compares to 16 other models on 9 datasets, providing extensive benchmarking. Ablations provided in the appendix show that AgenticGNN can address spectral errors in the proposed algorithm designs.

Critical weaknesses

It seems like AgenticGNN is incrementally better at best. Their method performs best on 5 of the 9 datasets and when it does best, its improvement is always within one standard deviation. Furthermore, many experimental details are left out of the paper that make it hard to understand if AgenticGNN is actually a better method. Specifically, I have two concerns about comparison fairness.

  1. It is not clear why certain models have accuracy numbers reused in table 1 while others are rerun for recording accuracy. For example, table 1 reuses the accuracy AutoSGNN reports in their paper but reruns experiments for GNAS-LLM and reports significantly lower results on Cora and Citeseer. If accuracy numbers from the GNAS-LLM paper were used, GNAS-LLM would beat AgenticGNN on two of the three datasets they share.

  2. The experiments say that GPT-4o was used in AgenticGNN. Are the compared agentic search methods using a model with comparable reasoning ability? The best alternative, AutoSGNN, uses GPT-3.5-turbo. Since this model is substantially weaker (https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/), it is unclear if AgenticGNN will remain best when controlling for model reasoning capability.

Other feedback (will not affect decision):

Numerous spelling and grammar problems are present, but they are not the main concern. A few: Figure 1 misspells “spectral.” Section 3.3 should be “thereby guiding the optimization of [the] graph learning operator.” The sentence after should use “its” instead of “it’s.” Discussion section titles should use “analysis” instead of “analyze.”

The source code repository provided (https://anonymous.4open.science/r/AgenticGNN-FDCC) seems to only have a readme file.

Questions for authors

How was it determined what models to re-run results on vs. what models to report original numbers from?

Why was GPT-4o used for AgenticGNN, and does its performance significantly depend on the reasoning capabilities of the LLM used?

How were the graph features chosen? For example, spectral radius is chosen but why this over other spectral features? To my knowledge there is no precedent for choosing the selected features over others.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.