Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Harnessing Adaptive Topology Representations for Zero-Shot Graph Question Answering

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A small question-level router that picks among five graph drawings and three text formats improves zero-shot graph question-answering in both accuracy and brevity, with no access to model weights.

desk verdict A new routing idea with a self-contradicting ablation: the router loses to fixed visual TRFs on the very tasks it's supposed to help. read the letter →

arxiv 2508.06345 v1 pith:WFAVUBJE submitted 2025-08-08 cs.CL cs.AIcs.GRcs.LG

Yanbin Wei , Jiangyue Yan , Chun Kang , Yang Chen , Hua Liu , James T. Kwok , Yu Zhang This is my paper
classification cs.CLcs.AIcs.GRcs.LG
keywords zero-shotgraphQAtopologyrepresentationformslargemultimodalmodelsdynamicTRFroutingResponseEfficiencyreasoningbrevityselection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that the format in which a graph is presented to a large multimodal model is a choice that should be made per question, not a fixed template. It assembles a set of eight topology representation forms — five graph drawings using different layout algorithms and three text encodings (edge set, adjacency list, adjacency matrix) — and defines the Graph Response Efficiency (GRE) metric, $E^k_f(q) = 100 \times \text{acc}^k_f(q) / (\text{avg.tok}^k_f(q))^\alpha$, which scores each format by accuracy divided by a power of the token count. Using 7,000 algorithmically validated questions, it labels each question with the format that maximizes GRE, then trains a compact DeBERTaV3 classifier to predict that format from the question text alone. On seven algorithmic graph tasks and two downstream tasks, routing the representation this way outperforms every fixed representation and the CoT, NLGraph, GraphDPR, and GITA baselines on GRE, and does so for two closed-source models (GPT-4o and Gemini-2.5 Pro) without modifying the model. The practical stake is that closed large multimodal models can become better graph reasoners through input-stage routing alone.

What carries the argument

The load-bearing object is the GRE metric, $E^k_f(q) = \frac{100 \times \text{acc}^k_f(q)}{(\text{avg.tok}^k_f(q))^\alpha}$, which collapses the accuracy-versus-brevity trade-off of a representation $f$ on question $q$ into one scalar, and the TRF Preference dataset built from it. Because GRE ranks the eight TRFs independently per question, it yields multi-label preference sets $F^*_q = \arg\max_{f \in \mathcal{F}_{ZS}} E^k_f(q)$, and those sets are the training signal for the router — a DeBERTaV3-base multi-label classifier that maps question text to one of the eight TRFs. The Pareto-optimality theorem licenses the central claim: for any question distribution and any $\alpha > 0$, a router

What would settle it

Hold out graph sizes and densities beyond the training range of the TRFP dataset (for example, graphs with more than 100 nodes or edge probability below 0.1) and compare the question-text-only router against Ideal Routing on the same questions. If the router's accuracy falls toward the best fixed TRF while Ideal Routing stays high, the graph-blind assumption is the binding constraint; additionally feeding the router graph statistics (node count, density, weight range) and observing whether the gap closes would confirm the diagnosis.

Watch

Extended reading notes

Core claim

The central claim is that one-size-fits-all graph representations are a bottleneck in zero-shot graph QA, and that the mapping from question to representation is learnable and worth learning. Concretely, no single TRF in $\mathcal{F}_{ZS}$ dominates the others across tasks: visual TRFs win on perception-heavy tasks (connectivity, cycle detection, bipartite matching), while textual TRFs win on computation-heavy weighted tasks (shortest path, max flow) and ordered-decomposition tasks (topological sort, Hamilton path). DynamicTRF makes this preference per question: GRE ranks the eight TRFs for each question (with $k=10$ runs and $\alpha=0.5$), the top-ranked TRFs become multi-label training tar

Load-bearing premise

The method assumes that the question's text alone — its wording and task type — predicts which graph representation will work best, because the router never sees the graph it is routing for; the paper's own gap between the trained router and Ideal Routing shows the cost of that assumption.

Editorial extensions

If this is right

  • TRF selection becomes a plug-in, input-stage adapter: any closed-source large multimodal model can benefit without fine-tuning or architectural changes.
  • The task-level preference patterns reduce to a usable taxonomy: draw the graph for perception-heavy questions and use text for computation-heavy ones.
  • GRE offers a brevity-aware evaluation standard, so verbose-but-correct and concise-but-wrong answers no longer score as well as balanced ones.
  • Cross-model router transfer works in both directions (GPT-4o router to Gemini-2.5 Pro and vice versa), suggesting preference datasets can be reused rather than rebuilt for every new model.
  • The router's benefit transfers to out-of-domain downstream tasks (link prediction and node classification), indicating the learned preferences are not overfit to the seven algorithmic templates.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The router sees only the question, never the graph; the ideal-routing gap in the paper (for GPT-4o, 68.4 vs. 81.8 on shortest path and 36.6 vs. 53.8 on max flow) is the measurable price of that blindness. A testable extension is to feed the router cheap graph statistics such as node count, edge density, and weight range, and check whether the gap closes.
  • The perception-versus-computation preference taxonomy plausibly generalizes to other structured inputs such as tables, trees, and code, which are also typically encoded with a single fixed template today; this is my extrapolation, not a claim the paper makes.
  • Because TRFP labels are tied to a specific $\alpha$ and a specific LMM, the dataset is a preference snapshot; a rank-based router or a continuous-trade-off formulation could make brevity a dial at inference time rather than a trigger for retraining the dataset.
  • The synthetic training graphs (Erdős–Rényi, $N \le 30$) may under-represent the large, clustered, heavy-tailed networks found in real applications, so whether the router's preferences persist on such graphs is an open, testable question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes DynamicTRF, a framework for zero-shot graph QA with large multimodal models (LMMs). It defines a set F_ZS of eight topology representation forms (TRFs): five visual (Graphviz layouts) and three textual (edge set, adjacency list, adjacency matrix). It introduces a Graph Response Efficiency (GRE) metric, Eq. (1), that balances accuracy against output-token brevity. A TRF Preference (TRFP) dataset is constructed by labeling each question with the TRF(s) maximizing GRE, Eq. (2), and a DeBERTaV3-base router is trained on the question text to select a TRF at inference time. The method is evaluated on seven in-domain algorithmic graph QA tasks and two out-of-domain downstream tasks using GPT-4o and Gemini-2.5 Pro. The paper claims that DynamicTRF improves both accuracy and brevity over fixed-TRF baselines.

Significance. If the central claim held, the contribution would be practically significant: a model-agnostic, question-level TRF router for closed-source LMMs would be a useful plug-in for graph QA, and the TRFP dataset and GRE metric would be reusable resources. The paper also provides a useful catalog of existing TRFs and documents task-level preferences. However, the empirical support is undermined by the paper's own ablation: the trained router does not outperform a fixed visual TRF under the proposed GRE metric, and the TRFP label statistics are inconsistent with the definition of F* in Eq. (2). These issues strike at the core claim, so the significance is currently not realized.

major comments (4)
  1. [§5.3, Table 5] The text states: 'With the TRF Router, DynamicTRF achieves the highest GRE scores across all tasks beyond all individual TRFs.' This is directly contradicted by Table 5. For GPT-4o on Connectivity, fixed Vneato has GRE 34.3, Vsfdp 33.3, and Vfdp 33.6, while the TRF Router has GRE 15.4. On Cycle, Vsfdp has GRE 32.2 versus the router's 10.3. Averaging the seven in-domain GRE columns, Vneato and Vsfdp each average approximately 10.7, whereas the TRF Router averages approximately 6.7. For Gemini-2.5 Pro (Table A16), Vneato averages about 11.0 versus the router's 10.2. Thus a single fixed TRF from F_ZS beats the adaptive router on average and on the perceptual tasks. This undermines the abstract's claim of simultaneous accuracy and brevity gains for the trained method, not merely the gap to Ideal Routing.
  2. [§4.4, Eq. (2), Table A11] The TRFP label definition in Eq. (2) is F*_q = arg max_f E^k_f(q). For Connectivity and Cycle, visual TRFs have GRE around 30-34 while textual TRFs have GRE around 5, so textual TRFs cannot be in the arg max set. Yet Table A11 reports high label frequencies for textual TRFs in these tasks: for GPT-4o on Connectivity, Tset appears in 47.7%, Tmat in 58.0%, and Tlist in 27.0% of F* sets; similar patterns appear for Gemini-2.5 Pro. This indicates that the labels were not generated by a strict argmax, or that a non-argmax construction (e.g., thresholding, ranking window, or multi-label tolerance) was used without being reported. Because the router is trained on these labels, the degraded Connectivity/Cycle performance in Table 5 may be a direct artifact of training on suboptimal labels. The paper must specify the exact label-construction rule and retrain with labels that strictly follow Eq. (
  3. [§4.3, Eq. (1)] The GRE metric uses avg.tok, the average number of output tokens, as the only cost term. Visual TRFs that elicit short answers such as 'Yes' or 'No' in roughly 8 output tokens receive GRE values above 30, while the cost of the input image and any API pricing for image tokens is ignored. This systematically biases the metric in favor of visual TRFs and makes the paper's 'brevity' and 'computational cost' claims incomparable across modalities. For closed-source LMMs, input token cost is a real part of the computational budget. At minimum, the paper should report input-token or API cost alongside output tokens, or redefine the cost term to include both, and check whether the main comparative results survive.
  4. [§4.5] The router receives only the question text q, not the graph G. The paper does not test whether question surface features are sufficient to predict the GRE-optimal TRF. The gap between the TRF Router and Ideal Routing in Table 5 (e.g., Connectivity 15.4 vs 35.6; Maximum Flow 1.9 vs 2.8) could be attributed to missing graph features, but this is not investigated. More seriously, the observation that fixed Vneato/Vsfdp beat the router on average shows that the limitation is not only graph-specific: the router fails to reproduce even task-level preferences. This is a load-bearing weakness of the proposed architecture and should be addressed either by including graph-derived features in the router or by demonstrating, with controlled experiments, that question text alone carries sufficient signal.
minor comments (5)
  1. [Tables 3 and 5] The DynamicTRF Topological Sort accuracy for GPT-4o is 40.8 in Table 3 but the TRF Router row in Table 5 reports 41.4 for the same configuration. Please reconcile.
  2. [Appendix G, Table A16] Table A16 repeats GPT-4o rows and gives the Gemini-2.5 Pro rows an 'Ideal Routing' line identical to GPT-4o's. Ideal Routing is model-specific and should be recomputed for Gemini-2.5 Pro.
  3. [Appendix A] Typo: 'barpartite' should be 'bipartite' in the sentence about node shape for bipartite graph matching.
  4. [Appendix B] The Erdős–Rényi model is spelled 'Erd6s' in the text; use the correct Unicode spelling.
  5. [§5.3] The sentence 'We leave Gemini-2.5 Pro results in Appendix G' is slightly misleading because Appendix G contains a complete duplicate of the GPT-4o comparison as well. It would be clearer to state that the full comparison for both models is in Appendix G.

Circularity Check

1 steps flagged · score 2.0 of 10

The trained-router evaluation is a genuine held-out prediction; the only circular element is the tautological 'optimal routing' theorem, whose R* is defined as the GRE-argmax selector.

  1. self definitional [Sec. 4.4 Eq. (2); Sec. 4.5 Theorem 1; Appendix D]
    "F ∗ q = arg max f∈F ZS E k f (q).(2) ... Theorem 1 (GRE-based Dynamic Routing Pareto Optimality). For any question distribution D q and tradeoff parameter α > 0, the optimal router R ∗ satisfying ∀f∈ FZS , R∗ ≻R f , where R ∗ always select a TRF f R∗ q ∈ F ∗ q"

    R* is defined as a router that always selects from F*_q, and F*_q is defined as the set of TRFs maximizing GRE. The theorem's conclusion that R* dominates every fixed router is simply the pointwise maximality built into Eq. (2): the proof's 'pointwise optimality' step restates the construction rather than deriving an independent result. It is a definitional tautology, and it cannot by itself establish that the trained DeBERTaV3 router (which is not R*) improves GRE.

full rationale

The central empirical pipeline is not circular: TRFP labels are generated by running the target LMM on a 7K probe set and computing GRE for each TRF; the router is then trained on (q, F*_q) pairs and evaluated on the separate GVLQA-BASE in-domain tasks and two external out-of-domain datasets. The router's selections are genuine predictions, and the gap between the trained router and Ideal Routing in Table 5 shows the result is not forced. Raw accuracy and token counts are reported alongside GRE, so the 'accuracy and brevity' claim has independent content. The self-citations to the authors' prior GITA work supply the visual TRF recipes and the in-domain benchmark, but they are used as components and baselines, not as a uniqueness argument, and the downstream OOD data are external. Appendix B appropriately discloses that D_TRFP is model- and GRE-specific, which limits but does not invalidate the comparison. The one genuinely circular item is Theorem 1/Appendix D: the 'optimal router' is defined as the GRE-argmax selector, so its dominance over fixed TRFs is true by construction and carries no predictive force. That tautology is not load-bearing for the empirical claims, which stand or fall on Tables 3-5. The apparent contradiction in Table 5 (fixed Vneato/Vsfdp outscoring the router in GRE on Conn/Cyc and on average) is a correctness/evaluation issue, not circularity. Overall: minor self-definitional step, no significant circularity in the main derivation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 3 invented entities

The empirical contributions rest on a small number of hand-picked choices: the GRE exponent alpha (Eq. 1), the number of runs k, the synthetic Erdős-Rényi training distribution, and the assumption that question text alone suffices for routing. The dominant costs in this ledger are the transfer assumption from small synthetic graphs to the evaluation benchmarks, and the definition of cost as output tokens only.

free parameters (3)
  • GRE exponent alpha = 0.5
    Eq. (1) defines GRE = 100*accuracy/(avg_tokens)^alpha. alpha = 0.5 is used to build the TRFP dataset and for the main router. It directly controls how much output brevity is rewarded, and ablations with alpha=0 and alpha=1 change the results.
  • Number of runs k = 10 for TRFP labels, 3 for evaluation
    k determines the stability of accuracy estimates in GRE. The choice is motivated only by cost, and no sensitivity analysis for k is provided.
  • TRFP graph generation ranges = N in [3,30], edge probability in [0.1,0.7], weights in {1,...,10}
    The synthetic data distribution used to define preferred TRFs is hand-picked. The router's transfer to larger and denser benchmarks (e.g., TS with 114 edges, real graphs with thousands of nodes) is not guaranteed.
assumptions (4)
  • domain assumption Each TRF faithfully encodes the graph so that differences in QA performance are attributable to representation format rather than information loss or rendering artifacts.
    The TRF analysis in Sec. 3.1 and the use of visual TRFs assume Graphviz images preserve enough topology for an LMM to recover, including for larger graphs.
  • domain assumption Question text alone suffices for routing decisions; graph content is not needed.
    The router in Sec. 4.5 takes only q as input. Any question-specific variation in TRF preference driven by graph density, size, or weights cannot be exploited. The Ideal Routing gap in Table 5 shows this limitation.
  • domain assumption The TRFP synthetic distribution transfers to GVLQA-BASE and large real-world graphs.
    Router training (Sec. 4.4) uses 7K small synthetic ER graphs, while evaluation includes larger and real graphs with no explicit specification of how those large graphs are rendered or subsampled.
  • domain assumption Automated algorithmic verification of LMM responses is a valid correctness measure.
    TRFP label construction in Appendix B and evaluation rely on exact-match or parser-based validation against deterministic algorithms such as Union-Find, Dijkstra, and Hopcroft-Karp.
invented entities (3)
  • GRE metric (Graph Response Efficiency)
    purpose: Defines the scalar reward used to rank TRFs and train the router; measures accuracy divided by output tokens to the power alpha.
    New evaluation metric with no external validation. The choice of alpha is user-dependent, and the metric is used both to generate labels and to score final results, creating a self-referential loop.
  • TRF Router (DeBERTaV3-base classifier) independent evidence
    purpose: Maps a question to a preferred TRF at inference; the core decision module of DynamicTRF.
    Its predictions are evaluated on held-out in-domain and out-of-domain benchmarks, giving a falsifiable handle outside the training distribution.
  • TRFP dataset
    purpose: Multi-label dataset of (question, preferred TRF set) used to train the router.
    No public release or external reuse is described; generated and evaluated by the same authors and models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Harnessing Adaptive Topology Representations for Zero-Shot Graph Question Answering." pith.science (2026). https://pith.science/paper/WFAVUBJE

@misc{pith2026250806345,
  author       = {Pith},
  title        = {Pith review of: Harnessing Adaptive Topology Representations for Zero-Shot Graph Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFAVUBJE}},
  note         = {Machine review of arXiv:2508.06345}
}
abstract

Large Multimodal Models (LMMs) have shown generalized zero-shot capabilities in diverse domain question-answering (QA) tasks, including graph QA that involves complex graph topologies. However, most current approaches use only a single type of graph representation, namely Topology Representation Form (TRF), such as prompt-unified text descriptions or style-fixed visual styles. Those "one-size-fits-all" approaches fail to consider the specific preferences of different models or tasks, often leading to incorrect or overly long responses. To address this, we first analyze the characteristics and weaknesses of existing TRFs, and then design a set of TRFs, denoted by $F_{ZS}$, tailored to zero-shot graph QA. We then introduce a new metric, Graph Response Efficiency (GRE), which measures the balance between the performance and the brevity in graph QA. Built on these, we develop the DynamicTRF framework, which aims to improve both the accuracy and conciseness of graph QA. To be specific, DynamicTRF first creates a TRF Preference (TRFP) dataset that ranks TRFs based on their GRE scores, to probe the question-specific TRF preferences. Then it trains a TRF router on the TRFP dataset, to adaptively assign the best TRF from $F_{ZS}$ for each question during the inference. Extensive experiments across 7 in-domain algorithmic graph QA tasks and 2 out-of-domain downstream tasks show that DynamicTRF significantly enhances the zero-shot graph QA of LMMs in terms of accuracy

Figures

Figures reproduced from arXiv: 2508.06345 by the authors.

Figure 1
Figure 1. Illustrating the impacts of diverse TRFs on the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An Illustration of eight candidate TRFs in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the DynamicTRF framework, where the TRF Router guides the LMM Reasoner to use the most appro [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Are Large Language Models Suitable for Graph Computation? Progress and Prospects

    cs.CL 2026-06 unverdicted novelty 4.0 of 10

    A survey of LLMs for graph computation introduces a role-based taxonomy of executors versus planners and concludes that current models suit simple small-scale tasks but remain unreliable for large-scale exact computation.

Reference graph

Works this paper leans on

47 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    A.; and Glance, N

    Adamic, L. A.; and Glance, N. 2005. The political blogosphere and the 2004 US election: divided they blog. In Proceedings of the 3rd international workshop on Link discovery, 36--43

    work page 2005

  2. [2]

    Akoglu, L.; Tong, H.; and Koutra, D. 2015. Graph based anomaly detection and description: a survey. Data mining and knowledge discovery, 29(3): 626--688

    work page 2015

  3. [3]

    Bellman, R. 1962. Dynamic programming treatment of the travelling salesman problem. Journal of the ACM (JACM), 9(1): 61--63

    work page 1962

  4. [4]

    Cao, Q.; Shen, H.; Gao, J.; Wei, B.; and Cheng, X. 2020. Popularity prediction on social platforms with coupled graph neural networks. In Proceedings of the 13th International Conference on Web Search and Data Mining, 70--78

    work page 2020

  5. [5]

    Chai, Z.; Zhang, T.; Wu, L.; Han, K.; Hu, X.; Huang, X.; and Yang, Y. 2023. Graphllm: Boosting graph reasoning ability of large language model. arXiv preprint arXiv:2310.05845

    arXiv 2023

  6. [6]

    Chen, N.; Li, Y.; Tang, J.; and Li, J. 2024 a . Graphwiz: An instruction-following language model for graph problems. arXiv preprint arXiv:2402.16029

    arXiv 2024

  7. [7]

    Chen, R.; Zhao, T.; Jaiswal, A.; Shah, N.; and Wang, Z. 2024 b . LLaGA: Large Language and Graph Assistant. arXiv preprint arXiv:2402.08170

    arXiv 2024

  8. [8]

    Cui, P.; Wang, X.; Pei, J.; and Zhu, W. 2018. A survey on network embedding. IEEE transactions on knowledge and data engineering, 31(5): 833--852

    work page 2018

Show all 47 references
  1. [9]

    Daniel, K. 2017. Thinking, fast and slow

    2017

  2. [10]

    Dijkstra, E. W. 1959. A note on two problems in connexion with graphs. Numerische Mathematik

    1959

  3. [11]

    Edmonds, J.; and Karp, R. M. 1972. Theoretical improvements in algorithmic efficiency for network flow problems. Journal of the ACM (JACM), 19(2): 248--264

    1972

  4. [12]

    Erd6s, P.; and R \'e nyi, A. 1960. On the evolution of random graphs. Publ. Math. Inst. Hungar. Acad. Sci, 5: 17--61

    1960

  5. [13]

    Evans, J. S. B. 1974. Dual processes in reasoning? Cognition, 3(2): 141--154

    1974

  6. [14]

    R.; and Fulkerson, D

    Ford, L. R.; and Fulkerson, D. R. 1956. Maximal flow through a network. Canadian Journal of Mathematics

    1956

  7. [15]

    M.; and Reingold, E

    Fruchterman, T. M.; and Reingold, E. M. 1991. Graph drawing by force-directed placement. Software: Practice and experience, 21(11): 1129--1164

    1991

  8. [16]

    R.; and North, S

    Gansner, E. R.; and North, S. C. 2000. An open graph visualization system and its applications to software engineering. Software: practice and experience, 30(11): 1203--1233

    2000

  9. [17]

    Google. 2025. Gemini-2.5 Pro . Technical report

    2025

  10. [18]

    Gould, R. J. 2003. Advances on the Hamiltonian problem--a survey. Graphs and Combinatorics

    2003

  11. [19]

    Guo, J.; Du, L.; Liu, H.; Zhou, M.; He, X.; and Han, S. 2023. Gpt4graph: Can large language models understand graph structured data? an empirical evaluation and benchmarking. arXiv preprint arXiv:2305.15066

    2023 arXiv

  12. [20]

    He, X.; Deng, K.; Wang, X.; Li, Y.; Zhang, Y.; and Wang, M. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In International Conference on Research and Development in Information Retrieval, 639--648

    2020

  13. [21]

    Hopcroft, J.; and Tarjan, R. 1973. Algorithm 447: efficient algorithms for graph manipulation. Communications of the ACM, 16(6): 372--378

    1973

  14. [22]

    E.; and Karp, R

    Hopcroft, J. E.; and Karp, R. M. 1973. An n\^ 5/2 algorithm for maximum matchings in bipartite graphs. SIAM Journal on computing, 2(4): 225--231

    1973

  15. [23]

    X.; and Wen, J.-R

    Jiang, J.; Zhou, K.; Dong, Z.; Ye, K.; Zhao, W. X.; and Wen, J.-R. 2023. StructGPT: A General Framework for Large Language Model to Reason over Structured Data. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 9237--9251

    2023

  16. [24]

    Jin, B.; Liu, G.; Han, C.; Jiang, M.; Ji, H.; and Han, J. 2024. Large language models on graphs: A comprehensive survey. IEEE Transactions on Knowledge and Data Engineering

    2024

  17. [25]

    Kahn, A. B. 1962. Topological sorting of large networks. Communications of ACM

    1962

  18. [26]

    M.; Vazirani, U

    Karp, R. M.; Vazirani, U. V.; and Vazirani, V. V. 1990. An optimal algorithm for on-line bipartite matching. In Proceedings of the twenty-second annual ACM symposium on Theory of computing

    1990

  19. [27]

    Kuang, J.; Shen, Y.; Xie, J.; Luo, H.; Xu, Z.; Li, R.; Li, Y.; Cheng, X.; Lin, X.; and Han, Y. 2025. Natural language understanding and inference with mllm in visual question answering: A survey. ACM Computing Surveys, 57(8): 1--36

    2025

  20. [28]

    Leskovec, J.; Kleinberg, J.; and Faloutsos, C. 2007. Graph evolution: Densification and shrinking diameters. ACM transactions on Knowledge Discovery from Data (TKDD), 1(1): 2--es

    2007

  21. [29]

    Li, Y.; Hu, B.; Shi, H.; Wang, W.; Wang, L.; and Zhang, M. 2024. Visiongraph: Leveraging large multimodal models for graph theory problems in visual context. arXiv preprint arXiv:2405.04950

    2024 arXiv

  22. [30]

    E.; and Girvan, M

    Newman, M. E.; and Girvan, M. 2004. Finding and evaluating community structure in networks. Physical review E, 69(2): 026113

    2004

  23. [31]

    OpenAI. 2024. GPT-4o . Technical report

    2024

  24. [32]

    Perozzi, B.; Fatemi, B.; Zelle, D.; Tsitsulin, A.; Kazemi, M.; Al-Rfou, R.; and Halcrow, J. 2024. Let your graph do the talking: Encoding structured data for llms. arXiv preprint arXiv:2402.05862

    2024 arXiv

  25. [33]

    Sedgewick, R. 2001. Algorithms in C, part 5: graph algorithms. Pearson Education

    2001

  26. [34]

    Tang, J.; Yang, Y.; Wei, W.; Shi, L.; Su, L.; Cheng, S.; Yin, D.; and Huang, C. 2023. Graphgpt: Graph instruction tuning for large language models. arXiv preprint arXiv:2310.13023

    2023 arXiv

  27. [35]

    Tang, J.; Zhang, Q.; Li, Y.; Chen, N.; and Li, J. 2025. Grapharena: Evaluating and exploring large language models on graph computation. In The Thirteenth International Conference on Learning Representations

    2025

  28. [36]

    Tarjan, R. E. 1975. Efficiency of a good but not linear set union algorithm. Journal of the ACM (JACM), 22(2): 215--225

    1975

  29. [37]

    Wang, H.; Feng, S.; He, T.; Tan, Z.; Han, X.; and Tsvetkov, Y. 2023. Can language models solve graph problems in natural language? Advances in Neural Information Processing Systems, 36: 30840--30861

    2023

  30. [38]

    Wasserman, S.; and Faust, K. 1994. Social network analysis: Methods and applications

    1994

  31. [39]

    H.; Le, Q

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Ichter, B.; Xia, F.; Chi, E. H.; Le, Q. V.; and Zhou, D. 2022. Chain of Thought Prompting Elicits Reasoning in Large Language Models. In Proceedings of Neural Information Processing Systems

    2022

  32. [40]

    T.; and Zhang, Y

    Wei, Y.; Fu, S.; Jiang, W.; Kwok, J. T.; and Zhang, Y. 2024 a . Rendering graphs for graph reasoning in multimodal large language models. arXiv preprint arXiv:2402.02130, 1

    2024 arXiv

  33. [41]

    Wei, Y.; Fu, S.; Jiang, W.; Zhang, Z.; Zeng, Z.; Wu, Q.; Kwok, J.; and Zhang, Y. 2024 b . Gita: Graph to visual and textual integration for vision-language graph reasoning. Advances in Neural Information Processing Systems, 37: 44--72

    2024

  34. [42]

    Yamanishi, Y.; Araki, M.; Gutteridge, A.; Honda, W.; and Kanehisa, M. 2008. Prediction of drug--target interaction networks from the integration of chemical and genomic spaces. Bioinformatics, 24(13): i232--i240

    2008

  35. [43]

    Yang, Z.; Cohen, W.; and Salakhudinov, R. 2016. Revisiting semi-supervised learning with graph embeddings. In International conference on machine learning, 40--48. PMLR

    2016

  36. [44]

    Ye, R.; Zhang, C.; Wang, R.; Xu, S.; and Zhang, Y. 2023. Language is all a graph needs. arXiv preprint arXiv:2308.07134

    2023 arXiv

  37. [45]

    Zhang, J. 2023. Graph-toolformer: To empower llms with graph reasoning ability via prompt augmented by chatgpt. arXiv preprint arXiv:2304.11116

    2023 arXiv

  38. [46]

    Zhang, Q.; Hong, X.; Tang, J.; Chen, N.; Li, Y.; Li, W.; Tang, J.; and Li, J. 2024. Gcoder: Improving large language model for generalized graph problem solving. arXiv preprint arXiv:2410.19084

    2024 arXiv

  39. [47]

    Zhao, J.; Zhuo, L.; Shen, Y.; Qu, M.; Liu, K.; Bronstein, M.; Zhu, Z.; and Tang, J. 2023. Graphtext: Graph reasoning in text space. arXiv preprint arXiv:2310.01089

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.