pith. sign in

arxiv: 2512.08964 · v4 · pith:7VAW5CSZnew · submitted 2025-11-30 · 💻 cs.LG

T2T-LA: A Topology-to-Topology LLM Agent for Graph Learning with Neither Feature Access nor Task Knowledge

Yongyu Wang This is my paper

Pith reviewed 2026-05-21 18:38 UTC · model grok-4.3

classification 💻 cs.LG
keywords graph learningLLM agenttopology generationmachine learning for CADgraph representation learningzero-shot inferenceconnectivity patterns
0
0 comments X

The pith

An LLM agent infers useful graph topologies from failed attempts and their scores alone, without task knowledge or feature access.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether a large language model can reason about graph connectivity to produce effective topologies for downstream graph learning tasks in CAD settings. T2T-LA receives only a collection of previously unsuccessful graph structures along with numerical scores from an unknown scorer and must produce a new topology in one attempt. The method avoids all conventional requirements such as feature matrices, task descriptions, hand-crafted construction rules, or parameter tuning. A sympathetic reader would care because many engineering and design problems model states as graphs yet struggle to specify good connectivity patterns manually. If the approach holds, it suggests LLMs can discover task-relevant structure through pattern inference from negative examples.

Core claim

T2T-LA is given only failed graph topologies and the scores assigned to them by a private evaluator; the agent is told nothing about the underlying task, the feature matrix, or the process that generated the topologies. From this limited input the model infers hidden relationships between connectivity patterns and the observed scores, then outputs a new topology that, when used by the downstream algorithm, yields a sufficiently good solution.

What carries the argument

T2T-LA, the Topology-to-Topology LLM Agent, which takes failed topologies and their scores as sole input and generates a new topology by inferring connectivity-score relationships.

If this is right

  • Graph construction for CAD problems no longer requires task-specific rules or extensive tuning.
  • LLM-based agents become viable for topology design when useful structures are hard to specify by hand.
  • Representation learning pipelines can operate with minimal or zero knowledge of the downstream objective and data features.
  • One-shot generation of workable topologies from negative examples alone is feasible in ML-for-CAD workflows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same inference mechanism could apply to other graph-structured design domains such as circuit layout or molecular configuration.
  • Performance may improve if the agent is allowed to observe a small number of successful topologies rather than only failures.
  • The approach raises the question of whether similar agents can operate across multiple unrelated graph tasks without retraining.

Load-bearing premise

The LLM can discover meaningful links between graph connectivity patterns and the observed scores even though it receives no information about the task, the features, or how the topologies were created.

What would settle it

Run the downstream algorithm on many T2T-LA-generated topologies across different CAD tasks and check whether any of them produce solutions that meet the success threshold; consistent failure would falsify the central claim.

Figures

Figures reproduced from arXiv: 2512.08964 by Yongyu Wang.

Figure 1
Figure 1. Figure 1: Overview of the proposed method. 3.1 Phase 1: Learning Latent Representations Machine learning models can be viewed as transformation engines that convert inputs into outputs by successively reshaping representations. For example, deep neural networks compress and refine raw feature vectors through multiple nonlinear layers. Likewise, methods such as Support Vector Machines (SVM) and Support Vector Cluster… view at source ↗
Figure 2
Figure 2. Figure 2: provides a visual comparison of the degradation caused by SEA and by the baseline attack for all three models. The figure clearly shows that SEA consistently yields stronger attack effects than random reweighting. This confirms that spectral adversarial robustness evaluation is able to pinpoint structurally weak links that are particularly vulnerable to such subtle perturbations. Method Original ↓ SEA ↓ Ra… view at source ↗
read the original abstract

Graph learning aims to convert data into graph representations, which are fundamental to many problems in machine learning for CAD, where circuits, layouts, designs, and optimization states are often modeled as graph-structured objects. Existing graph learning methods usually rely on carefully designed graph construction rules, extensive parameter tuning, and sophisticated mathematical theory; moreover, achieving good performance often requires task-specific graph construction tailored to the downstream objective. In this work, we study whether a large language model (LLM) can reason about graph structure and infer a useful topology without observing the feature matrix, without knowing the downstream task, and without relying on any carefully designed graph construction algorithm or parameter tuning process. To this end, we propose T2T-LA, a Topology-to-Topology LLM Agent that receives no input other than a set of previously failed topologies and the scores assigned to them by a private scorer. The agent is not told what task or algorithm produces the scores, how these topologies are generated, or what the scores mean. Since none of the observed topologies is satisfactory, T2T-LA cannot simply imitate a good example. Instead, it is forced to infer hidden relationships between graph connectivity patterns and the observed scores, a capability that is particularly relevant to CAD scenarios where useful design structures may be difficult to specify manually. Experimental results show that T2T-LA can generate, in one shot, a graph topology that enables the downstream algorithm to produce a sufficiently good solution, suggesting a new LLM-driven direction for topology reasoning and graph representation learning in ML-for-CAD workflows.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes T2T-LA, a Topology-to-Topology LLM Agent for graph learning in ML-for-CAD. Given only a set of previously failed graph topologies and their scores from a private scorer—with no access to the feature matrix, no knowledge of the downstream task or algorithm, and no information on how the topologies were generated—the agent is claimed to infer hidden relationships between connectivity patterns and scores. It then generates, in one shot, a new topology that enables the downstream algorithm to reach a sufficiently good solution. This is positioned as a new direction for topology reasoning without task-specific rules or parameter tuning.

Significance. If the central empirical claim holds, the work demonstrates a viable LLM-based approach to graph topology generation that operates without feature access or task knowledge. This could be significant for CAD applications where manual specification of effective graph structures is difficult, potentially reducing reliance on hand-crafted construction rules and opening an empirical, feedback-driven path for representation learning in graph-structured design problems.

major comments (2)
  1. [Abstract] Abstract: The abstract reports positive experimental outcomes ('T2T-LA can generate, in one shot, a graph topology that enables the downstream algorithm to produce a sufficiently good solution') but supplies no information on the specific tasks or CAD domains tested, the baselines used for comparison, the number of trials or random seeds, statistical significance, or how the private scorer was implemented. These omissions leave the central empirical claim without visible quantitative support.
  2. [Method] Approach description: The claim that T2T-LA infers hidden connectivity-score relationships solely from failed topologies and scores (without task or feature information) is not supported by any described control experiments that would isolate this inference from the LLM's pre-trained knowledge of effective graph structures in CAD domains such as circuits and layouts. Without such controls, it remains possible that generation succeeds via implicit priors rather than the claimed score-driven reasoning.
minor comments (1)
  1. [Abstract] The abstract and introduction could more explicitly state the precise definition of 'sufficiently good solution' used to evaluate the generated topologies.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and specify the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The abstract reports positive experimental outcomes ('T2T-LA can generate, in one shot, a graph topology that enables the downstream algorithm to produce a sufficiently good solution') but supplies no information on the specific tasks or CAD domains tested, the baselines used for comparison, the number of trials or random seeds, statistical significance, or how the private scorer was implemented. These omissions leave the central empirical claim without visible quantitative support.

    Authors: We agree that the abstract is insufficiently specific. In the revised manuscript we will expand the abstract to name the CAD domains and downstream tasks evaluated, list the baselines, report the number of trials and random seeds, note statistical significance where applicable, and briefly describe the private scorer. revision: yes

  2. Referee: [Method] Approach description: The claim that T2T-LA infers hidden connectivity-score relationships solely from failed topologies and scores (without task or feature information) is not supported by any described control experiments that would isolate this inference from the LLM's pre-trained knowledge of effective graph structures in CAD domains such as circuits and layouts. Without such controls, it remains possible that generation succeeds via implicit priors rather than the claimed score-driven reasoning.

    Authors: This observation is correct: the current manuscript does not present explicit control experiments that isolate score-driven reasoning from the LLM's pre-trained priors on CAD graphs. We will add a dedicated limitations paragraph that acknowledges this possibility and outlines control experiments (e.g., score ablation, out-of-distribution graph tasks, or models without CAD pre-training) that could be performed in follow-up work. We maintain that the one-shot setting with only failed topologies and scores already constrains the agent more tightly than typical prompting, but we accept that stronger isolation would be desirable. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical agent method with no derivation chain

full rationale

The paper presents T2T-LA as an LLM-based empirical agent that takes failed topologies and private scores as input to generate a new topology, without any mathematical derivation, equations, fitted parameters, or self-citation chains. The abstract and description emphasize experimental results on CAD-related graphs rather than a first-principles proof or closed-form prediction that reduces to its inputs by construction. No load-bearing steps match the enumerated circularity patterns; the approach is self-contained as a black-box inference procedure evaluated externally.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the untested premise that an off-the-shelf LLM possesses latent graph-topology reasoning ability that can be elicited by failure-score pairs alone. No free parameters or invented physical entities are introduced; the agent itself is the novel component.

axioms (1)
  • domain assumption Large language models can perform non-trivial reasoning over graph connectivity patterns when given only negative examples and scalar scores.
    Invoked in the abstract when stating that the agent is forced to infer hidden relationships.
invented entities (1)
  • T2T-LA agent no independent evidence
    purpose: To map sets of failed topologies and scores to a new useful topology without task or feature information.
    The agent is introduced as the core technical contribution; no independent evidence for its capabilities is supplied beyond the abstract claim.

pith-pipeline@v0.9.0 · 5814 in / 1199 out tokens · 37785 ms · 2026-05-21T18:38:54.826284+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 1 internal anchor

  1. [1]

    Adversarial attack and defense on graph data: A survey,

    L. Sun, Y. Dou, C. Yang, K. Zhang, J. Wang, P. S. Yu, L. He, and B. Li, “Adversarial attack and defense on graph data: A survey,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 8, pp. 7693–7711, 2022

    work page 2022

  2. [2]

    Adversarial attack on graph structured data,

    H. Dai, H. Li, T. Tian, X. Huang, L. Wang, J. Zhu, and L. Song, “Adversarial attack on graph structured data,” inInternational conference on machine learning. PMLR, 2018, pp. 1115–1124

    work page 2018

  3. [3]

    Adversarial attacks on graph neural networks via meta learning,

    D. Zügner and S. Günnemann, “Adversarial attacks on graph neural networks via meta learning,” 2024. [Online]. Available:https://arxiv.org/abs/1902.08412

    work page arXiv 2024

  4. [4]

    All you need is low (rank) defending against adversarial attacks on graphs,

    N. Entezari, S. A. Al-Sayouri, A. Darvishzadeh, and E. E. Papalexakis, “All you need is low (rank) defending against adversarial attacks on graphs,” inProceedings of the 13th international conference on web search and data mining, 2020, pp. 169–177

    work page 2020

  5. [5]

    Graph structure learning for robust graph neural networks,

    W. Jin, Y. Ma, X. Liu, X. Tang, S. Wang, and J. Tang, “Graph structure learning for robust graph neural networks,” inProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 2020, pp. 66–74

    work page 2020

  6. [6]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,”Advances in neural information processing systems, vol. 30, 2017

    work page 2017

  7. [7]

    Zoobp:Belief propagation for heterogeneous networks,

    D.Eswaran,S.Günnemann,C.Faloutsos,D.Makhija,andM.Kumar,“Zoobp:Belief propagation for heterogeneous networks,”Proceedings of the VLDB Endowment, vol. 10, no. 5, pp. 625–636, 2017

    work page 2017

  8. [8]

    Semi-Supervised Classification with Graph Convolutional Networks

    T. Kipf, “Semi-supervised classification with graph convolutional networks,”arXiv preprint arXiv:1609.02907, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  9. [9]

    von Luxburg, U.: A tutorial on spectral clustering.Statistics and Computing, 17(4):395–416 (2007)

    work page 2007

  10. [10]

    The graph neural network model,

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini, “The graph neural network model,”IEEE transactions on neural networks, vol. 20, no. 1, pp. 61–80, 2008

    work page 2008

  11. [11]

    Adversarial attacks on neural networks for graph data,

    D. Zügner, A. Akbarnejad, and S. Günnemann, “Adversarial attacks on neural networks for graph data,” inProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 2847–2856

    work page 2018

  12. [12]

    Making machine learning robust against adversarial inputs,

    I. Goodfellow, P. McDaniel, and N. Papernot, “Making machine learning robust against adversarial inputs,”Communications of the ACM, vol. 61, no. 7, pp. 56–66, 2018

    work page 2018

  13. [13]

    Analysis of classifiers’ robustness to adversarial perturbations,

    A. Fawzi, O. Fawzi, and P. Frossard, “Analysis of classifiers’ robustness to adversarial perturbations,”Machine learning, vol. 107, no. 3, pp. 481–508, 2018

    work page 2018

  14. [14]

    Topology attack and defense for graph neural networks: An optimization perspective,

    K. Xu, H. Chen, S. Liu, P.-Y. Chen, T.-W. Weng, M. Hong, and X. Lin, “Topology attack and defense for graph neural networks: An optimization perspective,”arXiv preprint arXiv:1906.04214, 2019

    work page arXiv 1906

  15. [15]

    F. R. Chung,Spectral graph theory. American Mathematical Soc., 1997, vol. 92

    work page 1997

  16. [16]

    G. H. Golub and C. F. Van Loan,Matrix computations. JHU press, 2013. 10 Yongyu Wang

    work page 2013

  17. [17]

    Spade: A spectral methodforblack-boxadversarialrobustnessevaluation,

    W. Cheng, C. Deng, Z. Zhao, Y. Cai, Z. Zhang, and Z. Feng, “Spade: A spectral methodforblack-boxadversarialrobustnessevaluation,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 1814–1824

    work page 2021

  18. [18]

    Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs,

    Y. A. Malkov and D. A. Yashunin, “Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs,”IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 4, pp. 824–836, 2018

    work page 2018

  19. [19]

    Support vector clustering,

    A. Ben-Hur, D. Horn, H. T. Siegelmann, and V. Vapnik, “Support vector clustering,” Journal of machine learning research, vol. 2, no. Dec, pp. 125–137, 2001

    work page 2001