pith. sign in

arxiv: 2605.10503 · v3 · pith:CMH4WK7Dnew · submitted 2026-05-11 · 💻 cs.AI

SLASH the Sink: Sharpening Structural Attention Inside LLMs

Yiming Liu , Bin Lu , Xinbing Wang , Chenghu Zhou , Meng Jin This is my paper

Pith reviewed 2026-05-20 22:15 UTC · model grok-4.3

classification 💻 cs.AI
keywords LLMsgraph reasoningattention mechanismsattention sinkstructural understandingtopology reconstructionSLASHmolecular prediction
0
0 comments X

The pith

LLMs spontaneously reconstruct graph topology inside their attention maps as a sawtooth pattern matching the token adjacency matrix, but the attention sink dilutes this ability through conflict with anisotropic bias; a plug-and-play fix rev

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

Large language models process graphs by turning them into token sequences but often miss the overall connections. The paper shows these models actually rebuild the graph structure internally, visible as a sawtooth pattern in attention maps that lines up with the token-level adjacency matrix. This natural structural sense gets weakened by the attention sink because the model's bias toward language tasks clashes with the need for local graph aggregation. The authors introduce SLASH, a training-free redistribution of attention scores that amplifies the internal reconstruction. Experiments confirm this delivers consistent gains on graph tasks and molecular predictions across different models without any retraining.

Core claim

LLMs spontaneously reconstruct the graph's topology internally, evidenced by a distinct sawtooth pattern in their attention maps that structurally aligns with the token-level adjacency matrix. This intrinsic structural understanding is diluted by the attention sink, which the authors formalize as a representation bottleneck stemming from a fundamental conflict between the model's anisotropic bias, essential for language tasks, and the topology-aware local aggregation required for graph reasoning. To address this, they propose SLASH, a training-free plug-and-play attention redistribution that amplifies the internal structural understanding and yields significant performance gains on pure grap

What carries the argument

SLASH (StructuraL Attention SHarpening), a plug-and-play attention redistribution technique that amplifies the spontaneous sawtooth alignment with graph topology to overcome dilution from the attention sink.

Load-bearing premise

The sawtooth pattern in attention maps represents genuine internal reconstruction of graph topology, and its dilution stems specifically from a representation bottleneck caused by conflict between anisotropic bias and topology-aware local aggregation.

What would settle it

Showing that attention maps lack the described sawtooth alignment with the token-level adjacency matrix, or that applying the SLASH redistribution produces no measurable improvement on graph reasoning tasks.

Figures

Figures reproduced from arXiv: 2605.10503 by Bin Lu, Chenghu Zhou, Meng Jin, Xinbing Wang, Yiming Liu.

Figure 1
Figure 1. Figure 1: The core mechanism of SLASH. While a standard LLM dilutes the latent structural “sawtooth” pattern with a dominant attention sink (left), SLASH sharpens this topological signal, enhancing the model’s focus on the internal structure (right). the main contributions of our work are as follows: • We identify the conflict between semantic anisotropy and topology-aware local aggregation as a fundamen￾tal bottlen… view at source ↗
Figure 2
Figure 2. Figure 2: Mechanistic evidence of internal graph reconstruction. (a) Input graph. (b) Internal Topology Reconstruction: Comparison of ground-truth structures (token-level Mgt and node adjacency) with their attention-derived counterparts (sawtooth pattern and reconstructed adjacency). (c) Full attention map. (d) Attention Budget: Proportional breakdown of sink bias, structural aggregation, and noise. 3.1. Problem For… view at source ↗
Figure 3
Figure 3. Figure 3: Performance sensitivity to the sharpening factor γ. General-purpose LLMs show model-specific sensitivity, while the fine-tuned model’s performance remains stable [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Case study on connectivity detection. The vanilla model incorrectly answers ‘Yes’ by generating a hallucinated path. SLASH prevents hallucinations, ensuring a grounded ‘No’. 7. Conclusion In this paper, we reveal LLMs’ latent ability to reconstruct graph topology, identifying the conflict between semantic anisotropy and topology-aware local aggregation as the bot￾tleneck. We propose SLASH, a training-free … view at source ↗
Figure 5
Figure 5. Figure 5: Entropy distribution in Llama-3.1-8B. Active heads (high entropy) are concentrated in intermediate layers. (Left) Per-head entropy heatmap with active heads boxed. (Right) Layer-averaged entropy plot, where an automatic threshold (dashed line) isolates the active peak. C. Computational Environment and Overhead Computational Environment. Experiments were conducted on a server with eight NVIDIA RTX 4090 GPUs… view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity analysis of γ on the MolecularNet dataset. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity analysis of γ on the GraphInstruct dataset. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
read the original abstract

Large Language Models (LLMs) show remarkable semantic understanding but often struggle with structural understanding when processing graph topologies in a serialized format. Existing solutions rely on training external graph-based adapters or fine-tuning, which incur high costs and lost generalizability. In this work, we investigate the internal mechanisms of LLMs and present a critical finding: LLMs spontaneously reconstruct the graph's topology internally, evidenced by a distinct "sawtooth" pattern in their attention maps that structurally aligns with the "token-level adjacency matrix". However, this intrinsic structural understanding is diluted by the attention sink. We theoretically formalize this dilution as a representation bottleneck, stemming from a fundamental conflict: the model's anisotropic bias, essential for language tasks, suppresses the topology-aware local aggregation required for graph reasoning. To address this, we propose a training-free solution, named StructuraL Attention SHarpening (SLASH), which amplifies this internal structural understanding via a plug-and-play attention redistribution. Experiments on pure graph tasks and molecular prediction validate that SLASH delivers significant and consistent performance gains across diverse LLMs.

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

3 major / 2 minor

Summary. The paper claims that LLMs spontaneously reconstruct graph topology when processing serialized inputs, as shown by a distinct 'sawtooth' pattern in attention maps that aligns with the token-level adjacency matrix. This intrinsic capability is diluted by the attention sink due to a conflict between the model's anisotropic bias (useful for language) and the need for topology-aware local aggregation. The authors formalize this as a representation bottleneck and propose SLASH, a training-free, plug-and-play attention redistribution technique to sharpen the structural signal. Experiments on pure graph tasks and molecular property prediction are reported to show consistent performance gains across diverse LLMs without fine-tuning or external adapters.

Significance. If the mechanistic interpretation of the sawtooth pattern as genuine topology reconstruction holds and SLASH's gains are robustly attributable to it, the work would be significant for enabling efficient, training-free improvements to LLMs' structural reasoning on graphs and molecules. This addresses a known limitation without the computational cost of adapters or fine-tuning, and the plug-and-play nature could have broad applicability. The absence of parameter-free derivations or machine-checked proofs limits the theoretical strength, but reproducible attention redistribution would be a practical contribution if validated.

major comments (3)
  1. [§3] §3 (or equivalent section on internal mechanisms): The central claim that the sawtooth pattern constitutes spontaneous reconstruction of graph topology (aligning with the token-level adjacency matrix) lacks controls to distinguish it from serialization artifacts. For example, no comparison is shown with node-order permutations, alternative linearizations, or non-graph sequences that preserve token positions; without these, the dilution-by-sink story and motivation for SLASH rest on an untested interpretation.
  2. [Experiments] Experiments section (likely §5): The reported performance gains on graph tasks and molecular prediction provide no details on statistical tests, variance across runs, or ablation of the redistribution parameters in SLASH. This makes it impossible to assess whether the gains are significant, consistent, or specifically due to amplifying the sawtooth rather than a generic attention adjustment.
  3. [§4] Theoretical formalization (likely §4): The representation bottleneck is described as arising from conflict between anisotropic bias and local aggregation, but no equation or derivation quantifies how the attention sink specifically suppresses the topology signal; the account remains qualitative and does not reduce to a testable model that predicts the observed sawtooth dilution.
minor comments (2)
  1. [Methods] The definition and construction of the 'token-level adjacency matrix' should include a small illustrative example in the main text or a dedicated figure to clarify how it is derived from the serialized graph input.
  2. [Figures] Figure captions for attention maps should explicitly state the graph size, serialization order, and layer/head indices shown, to allow readers to reproduce the sawtooth observation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment point by point below, indicating where revisions will be made to strengthen the work.

read point-by-point responses

  1. Referee: [§3] §3 (or equivalent section on internal mechanisms): The central claim that the sawtooth pattern constitutes spontaneous reconstruction of graph topology (aligning with the token-level adjacency matrix) lacks controls to distinguish it from serialization artifacts. For example, no comparison is shown with node-order permutations, alternative linearizations, or non-graph sequences that preserve token positions; without these, the dilution-by-sink story and motivation for SLASH rest on an untested interpretation.

    Authors: We agree that the current presentation would benefit from explicit controls to isolate topology reconstruction from potential serialization effects. While the sawtooth pattern is shown to align quantitatively with the token-level adjacency matrix across multiple datasets and models, we will add a dedicated subsection with new experiments: attention maps under node-order permutations, alternative linearizations (e.g., BFS versus DFS traversals), and non-graph sequences with matched token positions. These will be used to test whether the pattern persists specifically for graph-structured inputs. revision: yes

  2. Referee: [Experiments] Experiments section (likely §5): The reported performance gains on graph tasks and molecular prediction provide no details on statistical tests, variance across runs, or ablation of the redistribution parameters in SLASH. This makes it impossible to assess whether the gains are significant, consistent, or specifically due to amplifying the sawtooth rather than a generic attention adjustment.

    Authors: We acknowledge that the experiments section lacks reported variance, statistical tests, and parameter ablations. In the revised manuscript we will report mean results with standard deviations over at least five random seeds, include paired statistical tests (e.g., Wilcoxon signed-rank) to establish significance of gains, and add ablations on SLASH hyperparameters such as redistribution strength and sink suppression ratio. These will demonstrate that performance improvements track the amplification of the structural attention signal rather than generic redistribution. revision: yes

  3. Referee: [§4] Theoretical formalization (likely §4): The representation bottleneck is described as arising from conflict between anisotropic bias and local aggregation, but no equation or derivation quantifies how the attention sink specifically suppresses the topology signal; the account remains qualitative and does not reduce to a testable model that predicts the observed sawtooth dilution.

    Authors: The current Section 4 offers a conceptual formalization of the representation bottleneck arising from the tension between anisotropic language bias and topology-aware aggregation. To make this more quantitative, we will introduce a simplified mathematical model in the revision that expresses the dilution of the sawtooth pattern as a function of sink-induced attention mass and the model's anisotropy parameter, including a short derivation showing how this predicts reduced local topology signal. This will render the account more directly testable against the observed attention maps. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on empirical attention patterns and external performance validation

full rationale

The paper reports an observed 'sawtooth' pattern in attention maps as evidence of spontaneous topology reconstruction, then proposes a training-free redistribution method (SLASH) that yields measured gains on graph and molecular tasks. No equations or derivations are shown that reduce a claimed prediction back to fitted inputs or self-citations by construction. The theoretical framing of a 'representation bottleneck' is presented as post-hoc explanation of the observed dilution rather than a closed mathematical loop. Self-citations, if present, are not load-bearing for the core empirical finding or the plug-and-play intervention.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard transformer attention assumptions plus the domain-specific premise that anisotropic bias is essential for language modeling and inherently conflicts with local topology aggregation.

axioms (1)
  • domain assumption LLMs possess an anisotropic bias that is essential for language tasks but suppresses topology-aware local aggregation
    Invoked in the abstract to explain the representation bottleneck and attention-sink dilution.

pith-pipeline@v0.9.0 · 5722 in / 1216 out tokens · 36919 ms · 2026-05-20T22:15:53.048810+00:00 · methodology

Review history (3 revisions) →

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 2 internal anchors

  1. [1]

    2024 , editor =

    Chen, Runjin and Zhao, Tong and Jaiswal, Ajay Kumar and Shah, Neil and Wang, Zhangyang , booktitle =. 2024 , editor =

    work page 2024

  2. [2]

    HiGPT: Heterogeneous Graph Language Model , booktitle =

    Jiabin Tang and Yuhao Yang and Wei Wei and Lei Shi and Long Xia and Dawei Yin and Chao Huang , editor =. HiGPT: Heterogeneous Graph Language Model , booktitle =. 2024 , timestamp =

    work page 2024

  3. [3]

    The Eleventh International Conference on Learning Representations,

    Jianan Zhao and Meng Qu and Chaozhuo Li and Hao Yan and Qian Liu and Rui Li and Xing Xie and Jian Tang , title =. The Eleventh International Conference on Learning Representations,. 2023 , timestamp =

    work page 2023

  4. [4]

    ICML 2024 AI for Science Workshop , year=

    A multi-view mixture-of-experts based on language and graphs for molecular properties prediction , author=. ICML 2024 AI for Science Workshop , year=

    work page 2024

  5. [5]

    The Twelfth International Conference on Learning Representations,

    Hao Liu and Jiarui Feng and Lecheng Kong and Ningyue Liang and Dacheng Tao and Yixin Chen and Muhan Zhang , title =. The Twelfth International Conference on Learning Representations,. 2024 , timestamp =

    work page 2024

  6. [6]

    Pure Transformers are Powerful Graph Learners , booktitle =

    Jinwoo Kim and Dat Nguyen and Seonwoo Min and Sungjun Cho and Moontae Lee and Honglak Lee and Seunghoon Hong , editor =. Pure Transformers are Powerful Graph Learners , booktitle =. 2022 , timestamp =

    work page 2022

  7. [7]

    Can Language Models Solve Graph Problems in Natural Language? , booktitle =

    Heng Wang and Shangbin Feng and Tianxing He and Zhaoxuan Tan and Xiaochuang Han and Yulia Tsvetkov , editor =. Can Language Models Solve Graph Problems in Natural Language? , booktitle =. 2023 , timestamp =

    work page 2023

  8. [8]

    McAuley , editor =

    Jianing Wang and Junda Wu and Yupeng Hou and Yao Liu and Ming Gao and Julian J. McAuley , editor =. InstructGraph: Boosting Large Language Models via Graph-centric Instruction Tuning and Preference Alignment , booktitle =. 2024 , timestamp =

    work page 2024

  9. [9]

    GraphGPT: Graph Instruction Tuning for Large Language Models , booktitle =

    Jiabin Tang and Yuhao Yang and Wei Wei and Lei Shi and Lixin Su and Suqi Cheng and Dawei Yin and Chao Huang , editor =. GraphGPT: Graph Instruction Tuning for Large Language Models , booktitle =. 2024 , timestamp =

    work page 2024

  10. [10]

    GraphWiz: An Instruction-Following Language Model for Graph Computational Problems , booktitle =

    Nuo Chen and Yuhan Li and Jianheng Tang and Jia Li , editor =. GraphWiz: An Instruction-Following Language Model for Graph Computational Problems , booktitle =. 2024 , timestamp =

    work page 2024

  11. [11]

    Haoyu Peter Wang and Peihao Wang and Mufei Li and Shikun Liu and Siqi Miao and Zhangyang Wang and Pan Li , booktitle=. Graph-

    work page

  12. [12]

    The Twelfth International Conference on Learning Representations,

    Bahare Fatemi and Jonathan Halcrow and Bryan Perozzi , title =. The Twelfth International Conference on Learning Representations,. 2024 , timestamp =

    work page 2024

  13. [13]

    What Does BERT Learn about the Structure of Language?

    Jawahar, Ganesh and Sagot, Beno \^i t and Seddah, Djam \'e. What Does BERT Learn about the Structure of Language?. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019

    work page 2019

  14. [14]

    Measures of Entropy From Data Using Infinitely Divisible Kernels , journal =

    Luis Gonzalo S. Measures of Entropy From Data Using Infinitely Divisible Kernels , journal =. 2015 , timestamp =

    work page 2015

  15. [15]

    The Twelfth International Conference on Learning Representations , year=

    Efficient Streaming Language Models with Attention Sinks , author=. The Twelfth International Conference on Learning Representations , year=

    work page

  16. [16]

    Why do llms attend to the first token?arXiv preprint arXiv:2504.02732, 2025

    Federico Barbero and Alvaro Arroyo and Xiangming Gu and Christos Perivolaropoulos and Michael M. Bronstein and Petar Velickovic and Razvan Pascanu , title =. CoRR , volume =. 2025 , eprinttype =. 2504.02732 , timestamp =

    work page arXiv 2025

  17. [17]

    How Contextual are Contextualized Word Representations? C omparing the Geometry of BERT , ELM o, and GPT -2 Embeddings

    Ethayarajh, Kawin. How Contextual are Contextualized Word Representations? C omparing the Geometry of BERT , ELM o, and GPT -2 Embeddings. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019

    work page 2019

  18. [18]

    A Threshold Selection Method from Gray-Level Histograms , year=

    Otsu, Nobuyuki , journal=. A Threshold Selection Method from Gray-Level Histograms , year=

    work page

  19. [19]

    Gomez and Lukasz Kaiser and Illia Polosukhin , title =

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , title =

    work page

  20. [20]

    What Does BERT Look at? An Analysis of BERT ' s Attention

    Clark, Kevin and Khandelwal, Urvashi and Levy, Omer and Manning, Christopher D. What Does BERT Look at? An Analysis of BERT ' s Attention. Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. 2019

    work page 2019

  21. [21]

    Forty-first International Conference on Machine Learning , year=

    Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration , author=. Forty-first International Conference on Machine Learning , year=

    work page

  22. [22]

    arXiv preprint arXiv:2510.06477 , year=

    Enrique Queipo. Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coin , journal =. 2025 , eprinttype =. 2510.06477 , timestamp =

    work page arXiv 2025

  23. [23]

    The Thirteenth International Conference on Learning Representations , year=

    When Attention Sink Emerges in Language Models: An Empirical View , author=. The Thirteenth International Conference on Learning Representations , year=

    work page

  24. [24]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    work page

  25. [25]

    Forty-second International Conference on Machine Learning , year=

    Layer by Layer: Uncovering Hidden Representations in Language Models , author=. Forty-second International Conference on Machine Learning , year=

    work page

  26. [26]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Do Language Models Use Their Depth Efficiently? , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    work page

  27. [27]

    The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models

    Razzhigaev, Anton and Mikhalchuk, Matvey and Goncharova, Elizaveta and Oseledets, Ivan and Dimitrov, Denis and Kuznetsov, Andrey. The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models. Findings of the Association for Computational Linguistics: EACL 2024. 2024

    work page 2024

  28. [28]

    and Gomes, Joseph and Geniesse, Caleb and Pappu, Aneesh S

    Wu, Zhenqin and Ramsundar, Bharath and Feinberg, Evan N. and Gomes, Joseph and Geniesse, Caleb and Pappu, Aneesh S. and Leswing, Karl and Pande, Vijay. MoleculeNet: a benchmark for molecular machine learning. Chem. Sci. 2018

    work page 2018

  29. [29]

    CoRR , volume =

    Yuyan Liu and Sirui Ding and Sheng Zhou and Wenqi Fan and Qiaoyu Tan , title =. CoRR , volume =. 2024 , eprinttype =. 2406.12950 , timestamp =

    work page arXiv 2024

  30. [30]

    The Llama 3 Herd of Models

    Llama Team , title =. CoRR , volume =. 2024 , eprinttype =. 2407.21783 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [31]

    Qwen3 Technical Report

    An Yang and Anfeng Li and Baosong Yang and Beichen Zhang and Binyuan Hui and Bo Zheng and Bowen Yu and Chang Gao and Chengen Huang and Chenxu Lv and Chujie Zheng and Dayiheng Liu and Fan Zhou and Fei Huang and Feng Hu and Hao Ge and Haoran Wei and Huan Lin and Jialong Tang and Jian Yang and Jianhong Tu and Jianwei Zhang and Jian Yang and Jiaxi Yang and Ji...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    2023 , timestamp =

    Hao Yuan and Haiyang Yu and Shurui Gui and Shuiwang Ji , title =. 2023 , timestamp =

    work page 2023

  33. [33]

    ICML 2025 Workshop on Methods and Opportunities at Small Scale , year=

    ZeroTuning: Unlocking the Initial Token's Power to Enhance Large Language Models Without Training , author=. ICML 2025 Workshop on Methods and Opportunities at Small Scale , year=

    work page 2025

  34. [34]

    2025 , eprint=

    Identifying and Evaluating Inactive Heads in Pretrained LLMs , author=. 2025 , eprint=

    work page 2025

  35. [35]

    Towards Mechanistic Interpretability of Graph Transformers via Attention Graphs , journal =

    Batu El and Deepro Choudhury and Pietro Li. Towards Mechanistic Interpretability of Graph Transformers via Attention Graphs , journal =. 2025 , eprinttype =. 2502.12352 , timestamp =

    work page arXiv 2025

  36. [36]

    Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Wen, Zhihao and Fang, Yuan , title =. Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2023 , isbn =

    work page 2023

  37. [37]

    The Fourteenth International Conference on Learning Representations,

    Wu, Jingyao and Lu, Bin and Di, Zijun and Gan, Xiaoying and Jin, Meng and Fu, Luoyi and Wang, Xinbing and Zhou, Chenghu , title =. The Fourteenth International Conference on Learning Representations,

    work page

  38. [38]

    CoRR , volume =

    Jingyao Wu and Bin Lu and Zijun Di and Xiaoying Gan and Meng Jin and Luoyi Fu and Xinbing Wang and Chenghu Zhou , title =. CoRR , volume =. 2026 , eprinttype =. 2602.01771 , timestamp =

    work page arXiv 2026

  39. [39]

    Graph Out-of-Distribution Generalization With Controllable Data Augmentation , year=

    Lu, Bin and Zhao, Ze and Gan, Xiaoying and Liang, Shiyu and Fu, Luoyi and Wang, Xinbing and Zhou, Chenghu , journal=. Graph Out-of-Distribution Generalization With Controllable Data Augmentation , year=

    work page

  40. [40]

    I nstruct G raph: Boosting Large Language Models via Graph-centric Instruction Tuning and Preference Alignment

    Wang, Jianing and Wu, Junda and Hou, Yupeng and Liu, Yao and Gao, Ming and McAuley, Julian. I nstruct G raph: Boosting Large Language Models via Graph-centric Instruction Tuning and Preference Alignment. Findings of the Association for Computational Linguistics: ACL 2024. 2024

    work page 2024