Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

GRIP claims that an LLM can answer graph queries with no graph text in the prompt, because a one-time LoRA fine-tuning memorizes the graph into the model's parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 23:33 UTC pith:74YBKF4S

load-bearing objection A useful idea and solid engineering, but the KG numbers don't support the headline claim because test triplets were likely memorized during fine-tuning. the 3 major comments →

arxiv 2511.07457 v2 pith:74YBKF4S submitted 2025-11-06 cs.CL cs.AI

GRIP: In-Parameter Graph Reasoning through Fine-Tuning Large Language Models

Jiarui Feng , Donghong Cai , Yixin Chen , Muhan Zhang This is my paper
classification cs.CL cs.AI
keywords in-parameter learninggraph reasoningLoRAparameter-efficient fine-tuningknowledge graph completionquestion answeringtest-time adaptationtext-attributed graphs
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a graph does not have to appear in an LLM's prompt for the LLM to reason over it. GRIP fine-tunes a low-rank adapter (LoRA) on carefully generated tasks—reciting node/edge descriptions, summarizing sampled subgraphs, and answering retrieval and reasoning questions—so that the graph's relational knowledge is stored in parameters rather than in context. At inference, the query goes in alone, making token cost per query constant. On graphs too large for a context window, the paper reports GRIP consistently beating LLM baselines that are given subgraph context, and on smaller graphs matching full-context performance while cutting inference time by 30–60%. If this holds, one-time graph-specific fine-tuning turns a static graph into a reusable, lightweight adapter.

Core claim

GRIP's central discovery is that structural and relational graph information can be compressed into LLM parameters through LoRA, removing the need for graph serialization or a graph encoder at inference. The training pipeline decomposes the graph into nodes, edges, and sampled subgraphs, uses an LLM to generate summaries and diverse QA pairs, then fine-tunes in two stages: first to memorize graph context, then to learn retrieval and reasoning behavior. The resulting adapter lets the model answer a broad range of queries about the memorized graph with no graph context. The paper's experiments report that this approach outperforms no-context LLMs everywhere, outperforms graph-context baselines

What carries the argument

The central mechanism is LoRA (low-rank adaptation) used as a compact, reusable graph memory. Fine-tuning runs in two stages: stage one memorizes node and edge descriptions plus LLM-generated summaries of sampled nodes, edges, and subgraphs; stage two trains context-QA (masked node/edge/relation completion) and reasoning-QA (multi-hop, global, binary, k-shot) tasks. Applying LoRA to the MLP layers stores the graph while leaving the base LLM's retrieval abilities intact, so inference is just query plus adapter.

Load-bearing premise

The load-bearing premise is that the graph memorized into LoRA excludes the test questions' answers; for the knowledge-graph experiments this means test triplets must not be part of the injected graph. If test edges are included, the reported KG scores are recall of memorized facts, not reasoning over learned structure.

What would settle it

Compare the paper's reported total edge counts (Table 10) with the sum of train/val/test splits. For FB15K237 the total is 310,116, exactly train+val+test (272,115+17,535+20,466), indicating test triplet edges were in the memorized graph. A clean experiment would fine-tune GRIP on the graph with all test triplets removed and re-run evaluation; if accuracy drops sharply, the headline result is memorization rather than generalization.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Token cost per graph query becomes constant; repeated queries over the same graph no longer re-send the graph.
  • Graphs that exceed the context window become usable, because the full graph is (claimed to be) encoded rather than approximated by a sampled subgraph.
  • A trained LoRA is a portable, reusable artifact for a static graph, which lowers the amortized cost of GraphRAG and web-scale retrieval.
  • Supplying the original graph context after GRIP fine-tuning improves results further, suggesting the method also works as test-time adaptation.

Where Pith is reading between the lines

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

  • Beyond the paper: the same 'memorize-then-reason' recipe should transfer to other structured data—tables, hierarchies, or codebases—by generating summaries and QA pairs over that structure and storing it in a LoRA.
  • Beyond the paper: because GRIP never feeds an edge list, it may be less sensitive to node/edge permutation, a known failure mode of graph-to-sequence prompting; a permutation-robustness test on the Scene Graph data would settle this.
  • Beyond the paper: a trained graph adapter is a modular memory; merging or routing multiple adapters could let one base LLM handle many graphs without context, and incremental LoRA updates could support knowledge-graph refresh over time.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes GRIP, a method that internalizes graph knowledge into LLM parameters through LoRA fine-tuning on four task families: context memorization (Eq. 1), summarization, context QA, and reasoning QA. At inference, the graph is omitted from the input. Experiments on four KG completion datasets (FB15K237, WN18RR, CoDEx-Medium, NELL23K) and two reasoning benchmarks (Scene Graph, CLEGR-Reasoning) claim that for graphs exceeding the context window GRIP consistently outperforms LLM baselines, while for small graphs it achieves comparable accuracy with lower inference cost. The paper also reports efficiency gains and ablations over task components and LoRA hyperparameters.

Significance. If the empirical claims were sound, GRIP would be a practical and efficient test-time adaptation method for repeated queries over static graphs, with public code and detailed hyperparameter settings. However, the main evidence for the large-graph superiority claim is compromised by apparent test-set leakage and by overlap between a training objective and the evaluation task, so the current results do not establish the claimed generalization.

major comments (3)
  1. [§4.2, Eq. (1); §5.2; Table 10; Appendix B.1] The context memorization loss sums over all edges e in E. Table 10 lists FB15K237 edges as 310,116, which equals 272,115 + 17,535 + 20,466 (train+val+test) reported in B.1; the same equality holds for WN18RR, CoDEx-Medium, and NELL23K. Section 5.2 states GRIP 'injects knowledge from the entire graph into the model parameters.' No statement excludes val/test triplets from E or from the context-QA sampling (§4.3). Thus the reported KG accuracy (e.g., 88.75 vs 79.65 for Qwen on FB15K237) likely measures memorization recall of test triplets, not generalization. This invalidates the headline 'consistently outperforms' for large graphs, which relies on Tables 2–3.
  2. [§4.3; §5.2] The edge-level context QA task masks one of (s, r, t) and asks the model to predict the missing component; the KG evaluation is exactly masked-relation prediction. The authors acknowledge in §5.2 that this makes the evaluation 'effectively closer to in-domain evaluations,' yet still treat Tables 2–3 as evidence of generalization. The only out-of-domain evidence, Table 4, shows GRIP without context (Scene 53.27; Clegr 27.66) below the context-provided baselines (67.46; 33.48). Hence the claim that GRIP 'consistently outperforms' is unsupported by uncontaminated results.
  3. [Table 4; abstract and §5.2] The abstract claims 'comparable performance' on small-scale graphs. On Scene Graph, no-context GRIP (53.27 for Qwen) is 14.19 points below the context baseline (67.46); on Clegr the gap is 5.82 points. Calling this a 'minimal gap' in §5.2 is an overstatement. This matters because the 'comparable performance' claim is part of the paper's central contribution, and the comparison is confounded by the baseline receiving full graph context while GRIP receives none.
minor comments (5)
  1. [Throughout] The dataset name is inconsistent: 'Clegr-reasoning' vs 'CLEGR-Reasoning' appears in varying capitalization. Please standardize.
  2. [Table 9] Header reads 'FB5K237' — likely a typo for 'FB15K237'.
  3. [§5.1, Table 2] WN18RR results are reported as an average over five runs due to high variance, but no standard deviation or error bar is provided. Please include variance information to assess significance.
  4. [Tables 2–3] The 'full-context OOC' rows are informative, but 'OOC' should be defined in the caption (e.g., 'out of context window').
  5. [Eq. (1)] Clarify whether E includes all dataset edges or only training edges. If the latter, state it explicitly in the text and in Table 10.

Circularity Check

2 steps flagged

KG link prediction is trained on the test triplets and then reported as an outperformance result; the large-graph 'consistently outperforms' claim reduces to memorization recall.

specific steps
  1. fitted input called prediction [§4.2 Eq. (1); Table 10; §B.1]
    "we decompose the graph into smaller, independent components and encode them separately by fine-tuning the model to memorize individual features of nodes and edges directly. Formally, the task can be defined as: L_context = Σ_{v∈V} L_LM(x_v;θ) + Σ_{e∈E} L_LM(x_e;θ). (1)"

    For the KG benchmarks the input graph is the full dataset: Table 10 reports FB15K237 with 310,116 total relations, and §B.1 states the splits are 272,115/17,535/20,466 (sum = 310,116). Thus the memorization loss in Eq. (1) runs over every edge, including test triplets. The later evaluation asks the model to predict the relation of those same triplets, so the reported 'prediction' is a training input. No passage states that test or validation triplets were excluded from any fine-tuning task family.

  2. self definitional [§5.2; Appendix A.2 (Rel Prompt)]
    "During fine-tuning, GRIP injects knowledge from the entire graph into the model parameters—a capability absent in baseline LLM methods, which rely on subgraph sampling to fit within the context window. ... Second, these tasks evaluate the model on link prediction. Since GRIP incorporates similar link prediction tasks through context QA as part of its training objectives, they are effectively closer to in-domain evaluations."

    The edge-level context QA objective trains the exact test format: given (s,t), ask 'what is the relation between node {src} and node {tgt}?' with answer {rel}. KG test-time completion asks precisely this masked-relation query. Therefore the test is, by construction, an in-domain copy of the training objective, and the reported gains on large graphs are not evidence of generalization to unseen relations. The paper's own 'in-domain evaluations' statement confirms that the predicted relation is equivalent to a training target.

full rationale

The method's core technique—memorizing graph text into LoRA parameters—is not circular in itself, and the reasoning-benchmark results (Table 4) provide some independent evidence. However, the headline claim that GRIP 'consistently outperforms' on graphs too large for the context window rests on the KG tables (Tables 2–3). There, Eq. (1) memorizes every edge in the dataset, including test triplets, and the edge-level context QA trains the identical masked-relation query used at test time. The paper explicitly calls these evaluations 'effectively closer to in-domain evaluations.' The result is therefore forced by construction: test accuracy on KG completion measures recall of memorized triplets, not generalization. The reasoning-task results do not rescue the headline, since without context GRIP (53.27 on Scene Graph) is below the context-provided baseline (67.46). No load-bearing self-citation chain is involved; the circularity is in the evaluation design rather than the bibliography.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

GRIP does not introduce new physical or mathematical entities. The load-bearing assumptions are the storage capacity of small LoRA adapters, the correctness of LLM-generated training tasks, the exclusion of test edges from the memorized graph, and the reliability of the LLM-as-judge metric. The dataset-specific hyperparameters act as free choices that shape the reported numbers.

free parameters (3)
  • Task-generation counts N_s, N_c, N_r = e.g., FB15K237: 6000/8000/2000; Scene Graph: 50/50/100
    Manually chosen per dataset in Table 11. Ablations in Table 6-7 show these counts materially affect accuracy, so the reported results depend on these hand-picked values.
  • LoRA rank r and alpha = r=16/24, alpha=32/48 depending on dataset
    Table 9 shows accuracy increases monotonically with rank on FB15K237, so the reported results use the largest tested rank. This is performance-driven hyperparameter selection.
  • Early-stop loss thresholds = 0.15-0.4 depending on dataset
    Used to stop stage 1 and stage 2 training. There is no stated principled criterion linking the threshold to final accuracy, making it a free choice that affects the trained adapter.
axioms (4)
  • domain assumption A rank-16 to 24 LoRA can store a graph of up to ~40k nodes and ~310k edges with enough fidelity to answer downstream questions.
    This is the central capacity assumption of in-parameter graph storage. It is only indirectly validated by downstream accuracy, and it is challenged by the observed need for larger rank on larger graphs.
  • domain assumption LLM-generated summaries and QA pairs from Qwen2.5-7B preserve graph facts accurately.
    All training signal—summaries, context QA, reasoning QA—is generated by a teacher LLM (§4.2, §4.3). Hallucinated facts would be memorized as ground truth. No verification or filtering of generated content is described.
  • ad hoc to paper The memorized KG contains only training edges, excluding test triplets.
    Required for the KG experiments to measure generalization rather than memorization. The paper never states this exclusion, and Table 10 reports total edges matching train+val+test totals, while §5.2 says the entire graph is injected into parameters.
  • domain assumption The Qwen2.5-32B judge reliably equates model outputs with ground truth on Scene Graph and CLEGR.
    Used to compute accuracy for the two reasoning datasets. No human agreement or calibration analysis is reported, so the metric itself is an unvalidated model judgment.

pith-pipeline@v1.3.0-alltime-deepseek · 18174 in / 12841 out tokens · 110385 ms · 2026-08-03T23:33:24.013615+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in modeling sequential textual data and generalizing across diverse tasks. However, effectively adapting LLMs to structural data, such as knowledge graphs or web graphs, remains a fundamental challenge. Some approaches adopt complex strategies to convert graphs into text sequences, resulting in significant token overhead and rendering them impractical for large-scale graphs. Others introduce additional modules to encode graphs into fixed-size token representations for LLMs. However, these methods typically require large-scale post-training on graph-text corpus and complex alignment procedures, yet often yield sub-optimal results due to poor modality alignment. In this work, we propose GRIP. Instead of relying on heavy graph serialization or specialized graph encoding modules, GRIP directly internalizes complex relational knowledge from graphs into the parameters of LLM through carefully designed fine-tuning tasks. The acquired structural knowledge is compactly stored in lightweight LoRA modules, enabling the fine-tuned LLM to perform a wide range of tasks over the internalized graph without requiring access to the original graph as context at inference time. Extensive experiments validate our approach. For graphs that cannot fit within the LLMs context window, GRIP consistently outperforms LLM baselines by leveraging internalized graph knowledge, while for small-scale graphs, it achieves comparable performance with substantially lower inference cost.

Figures

Figures reproduced from arXiv: 2511.07457 by Donghong Cai, Jiarui Feng, Muhan Zhang, Yixin Chen.

Figure 1
Figure 1. Figure 1: Overview of GRIP. During the fine-tuning phase, we design a variety of tasks to inject graph context into the LoRA parameters and explicitly instruct the model to utilize this context for solving downstream tasks. In the inference phase, GRIP can directly answer user queries without requiring explicit graph context. In contrast, standard LLM-based inference over graphs relies on providing explicit graph co… view at source ↗
Figure 2
Figure 2. Figure 2: Relative inference time comparison between base [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Beyond One-Size-Fits-All: Adaptive Subgraph Denoising for Zero-Shot Graph Learning with Large Language Models

    cs.LG 2026-03 unverdicted novelty 7.0

    GraphSSR introduces an adaptive SSR pipeline with SSR-SFT data synthesis and SSR-RL (Authenticity-Reinforced and Denoising-Reinforced stages) to overcome one-size-fits-all subgraph noise in zero-shot LLM graph reasoning.

Reference graph

Works this paper leans on

50 extracted references · 28 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Zeyuan Allen-Zhu and Yuanzhi Li. 2023. Physics of language models: Part 3.1, knowledge storage and extraction.arXiv preprint arXiv:2309.14316(2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Ok- sana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data.Advances in neural information processing systems26 (2013)

    2013

  3. [3]

    Andrew Carlson, Justin Betteridge, Bryan Kisiel, Burr Settles, Estevam Hruschka, and Tom Mitchell. 2010. Toward an architecture for never-ending language learning. InProceedings of the AAAI conference on artificial intelligence, Vol. 24. 1306–1313

    2010

  4. [4]

    Rujikorn Charakorn, Edoardo Cetin, Yujin Tang, and Robert Tjarko Lange. 2025. Text-to-LoRA: Instant Transformer Adaption.arXiv preprint arXiv:2506.06105 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Runjin Chen, Tong Zhao, Ajay Kumar Jaiswal, Neil Shah, and Zhangyang Wang

  6. [6]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168(2021)

    Pith/arXiv arXiv 2021

  7. [7]

    Tim Dettmers, Minervini Pasquale, Stenetorp Pontus, and Sebastian Riedel. 2018. Convolutional 2D Knowledge Graph Embeddings. InProceedings of the 32th AAAI Conference on Artificial Intelligence. 1811–1818. https://arxiv.org/abs/1707.01476

    Pith/arXiv arXiv 2018

  8. [8]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130(2024)

    Pith/arXiv arXiv 2024

  9. [9]

    Jiarui Feng, Hao Liu, Lecheng Kong, Mingfang Zhu, Yixin Chen, and Muhan Zhang. 2024. TAGLAS: An atlas of text-attributed graph datasets in the era of large graph and language models.arXiv preprint arXiv:2406.14683(2024)

    Pith/arXiv arXiv 2024

  10. [10]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783 (2024)

    Pith/arXiv arXiv 2024

  11. [11]

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. 2024. DeepSeek-Coder: When the Large Language Model Meets Programming–The Rise of Code Intelligence.arXiv preprint arXiv:2401.14196(2024)

    Pith/arXiv arXiv 2024

  12. [12]

    Xiaoxin He, Xavier Bresson, Thomas Laurent, Adam Perold, Yann LeCun, and Bryan Hooi. 2024. Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning. InThe Twelfth International Conference on Learning Representations

    2024

  13. [13]

    Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi. 2024. G-retriever: Retrieval-augmented gen- eration for textual graph understanding and question answering.Advances in Neural Information Processing Systems37 (2024), 132876–132907

    2024

  14. [14]

    Yufei He, Yuan Sui, Xiaoxin He, and Bryan Hooi. 2024. UniGraph: Learning a Unified Cross-Domain Foundation Model for Text-Attributed Graphs.arXiv preprint arXiv:2402.13630(2024)

    Pith/arXiv arXiv 2024

  15. [15]

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2021. LoRA: Low-Rank Adaptation of Large Language Models. InInternational Conference on Learning Representations

    2021

  16. [16]

    Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin

  17. [17]

    Lecheng Kong, Jiarui Feng, Hao Liu, Chengsong Huang, Jiaxin Huang, Yixin Chen, and Muhan Zhang. 2025. GOFA: A Generative One-For-All Model for Joint Graph Language Modeling. InThe Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=mIjblC9hfm GRIP : In-Parameter Graph Reasoning through Fine-Tuning Large Languag...

    2025

  18. [18]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient Mem- ory Management for Large Language Model Serving with PagedAttention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles

    2023

  19. [19]

    Jakub Lála, Odhran O’Donoghue, Aleksandar Shtedritski, Sam Cox, Samuel G Rodriques, and Andrew D White. 2023. Paperqa: Retrieval-augmented generative agent for scientific research.arXiv preprint arXiv:2312.07559(2023)

    Pith/arXiv arXiv 2023

  20. [20]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems33 (2020), 9459–9474

    2020

  21. [21]

    Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun Liu. 2024. LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods. arXiv:2412.05579 [cs.CL] https://arxiv.org/abs/2412.05579

    Pith/arXiv arXiv 2024

  22. [22]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning. PMLR, 19730–19742

    2023

  23. [23]

    Jintang Li, Ruofan Wu, Yuchang Zhu, Huizhe Zhang, Liang Chen, and Zibin Zheng. 2025. Are Large Language Models In-Context Graph Learners? arXiv:2502.13562 [cs.LG] https://arxiv.org/abs/2502.13562

    Pith/arXiv arXiv 2025

  24. [24]

    Tianqianjin Lin, Pengwei Yan, Kaisong Song, Zhuoren Jiang, Yangyang Kang, Jun Lin, Weikang Yuan, Junjie Cao, Changlong Sun, and Xiaozhong Liu. 2024. LangGFM: A Large Language Model Alone Can be a Powerful Graph Foundation Model.arXiv preprint arXiv:2410.14961(2024)

    Pith/arXiv arXiv 2024

  25. [25]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437(2024)

    Pith/arXiv arXiv 2024

  26. [26]

    Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang. 2023. One For All: Towards Training One Graph Model For All Classification Tasks. InThe Twelfth International Conference on Learning Representations

    2023

  27. [27]

    Xin Lv, Xu Han, Lei Hou, Juanzi Li, Zhiyuan Liu, Wei Zhang, Yichi Zhang, Hao Kong, and Suhui Wu. 2020. Dynamic anticipation and completion for multi-hop reasoning over sparse knowledge graph.arXiv preprint arXiv:2010.01899(2020)

    Pith/arXiv arXiv 2020

  28. [28]

    Yansheng Mao, Jiaqi Li, Fanxu Meng, Jing Xiong, Zilong Zheng, and Muhan Zhang. 2024. LIFT: Improving Long Context Understanding Through Long Input Fine-Tuning.arXiv preprint arXiv:2412.13626(2024)

    Pith/arXiv arXiv 2024

  29. [29]

    OpenAI and et al. 2024. GPT-4 Technical Report. arXiv:2303.08774 [cs.CL] https://arxiv.org/abs/2303.08774

    Pith/arXiv arXiv 2024

  30. [30]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Des- maison, Andreas Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. PyTorch: An Imperative Style, Hi...

    2019

  31. [31]

    Soham Petkar, Anirudh Vempati, Akshit Sinha, Ponnurangam Kumarauguru, Chirag Agarwal, et al. 2025. A Graph Talks, But Who’s Listening? Rethinking Evaluations for Graph-Language Models.arXiv preprint arXiv:2508.20583(2025)

    Pith/arXiv arXiv 2025

  32. [32]

    Tara Safavi and Danai Koutra. 2020. Codex: A comprehensive knowledge graph completion benchmark.arXiv preprint arXiv:2009.07810(2020)

    Pith/arXiv arXiv 2020

  33. [33]

    Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2023. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face.Advances in Neural Information Processing Systems36 (2023), 38154–38180

    2023

  34. [34]

    George Stoica, Pratik Ramesh, Boglarka Ecsedi, Leshem Choshen, and Judy Hoffman. 2024. Model merging with svd to tie the knots.arXiv preprint arXiv:2410.19735(2024)

    Pith/arXiv arXiv 2024

  35. [35]

    Weihang Su, Yichen Tang, Qingyao Ai, Junxi Yan, Changyue Wang, Hongning Wang, Ziyi Ye, Yujia Zhou, and Yiqun Liu. 2025. Parametric Retrieval Augmented Generation.arXiv preprint arXiv:2501.15915(2025)

    Pith/arXiv arXiv 2025

  36. [36]

    Yuqiao Tan, Shizhu He, Huanxuan Liao, Jun Zhao, and Kang Liu. 2025. Dy- namic Parametric Retrieval Augmented Generation for Test-time Knowledge Enhancement. arXiv:2503.23895 [cs.CL] https://arxiv.org/abs/2503.23895

    Pith/arXiv arXiv 2025

  37. [37]

    Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Lixin Su, Suqi Cheng, Dawei Yin, and Chao Huang. 2024. Graphgpt: Graph instruction tuning for large language models. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 491–500

    2024

  38. [38]

    Heng Wang, Shangbin Feng, Tianxing He, Zhaoxuan Tan, Xiaochuang Han, and Yulia Tsvetkov. 2023. Can language models solve graph problems in natural language?Advances in Neural Information Processing Systems36 (2023), 30840– 30861

    2023

  39. [39]

    Haoyu Wang, Shikun Liu, Rongzhe Wei, and Pan Li. 2025. Model generalization on text attribute graphs: Principles with large language models.arXiv preprint arXiv:2502.11836(2025)

    Pith/arXiv arXiv 2025

  40. [40]

    Yan Wang, Dongyang Ma, and Deng Cai. 2024. With greater text comes greater necessity: Inference-time training helps long text generation.arXiv preprint arXiv:2401.11504(2024)

    Pith/arXiv arXiv 2024

  41. [41]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. HuggingFace’s Transforme...

    Pith/arXiv arXiv 2020

  42. [42]

    Xixi Wu, Yifei Shen, Caihua Shan, Kaitao Song, Siwei Wang, Bohang Zhang, Jiarui Feng, Hong Cheng, Wei Chen, Yun Xiong, et al. 2024. Can Graph Learning Improve Planning in LLM-based Agents?. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  43. [43]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng- peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfen...

    Pith/arXiv arXiv 2024

  44. [44]

    Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. 2024. Language is All a Graph Needs. InEACL (Findings)

    2024

  45. [45]

    Mengmei Zhang, Mingwei Sun, Peng Wang, Shen Fan, Yanhu Mo, Xiaoxiao Xu, Hong Liu, Cheng Yang, and Chuan Shi. 2024. GraphTranslator: Aligning Graph Model to Large Language Model for Open-ended Tasks.arXiv preprint arXiv:2402.07197(2024)

    Pith/arXiv arXiv 2024

  46. [46]

    A green apple is on the table at position (5, 10)

    Xi Zhu, Haochen Xue, Ziwei Zhao, Wujiang Xu, Jingyuan Huang, Minghao Guo, Qifan Wang, Kaixiong Zhou, and Yongfeng Zhang. 2025. LLM as GNN: Graph Vocabulary Learning for Text-Attributed Graph Foundation Models. arXiv:2503.03313 [cs.LG] https://arxiv.org/abs/2503.03313 Conference’17, July 2017, Washington, DC, USA Trovato and Tobin, et al. A Implementation ...

    arXiv 2025

  47. [49]

    Two sample questions must differ in at least one di- mension: Interrogative, Format, Focus

    Select two sample questions that best fit the provided text snippets. Two sample questions must differ in at least one di- mension: Interrogative, Format, Focus

  48. [50]

    how many

    For each selected sample question, write a new question with the same type and style, but the answer is fully grounded in the text snippets. Specifically, - The new questions should focus on reasoning over the pro- vided text snippets. - Avoid trivial lookups (e.g., complete an edge) and avoid quan- tity questions (“how many”). - Use natural phrasing; do ...

    2017

  49. [2023]

    arXiv preprint arXiv:2307.13269(2023)

    Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269(2023)

    Pith/arXiv arXiv 2023

  50. [2024]

    InInternational Conference on Machine Learning

    LLaGA: Large Language and Graph Assistant. InInternational Conference on Machine Learning. PMLR, 7809–7823