Pith. sign in

REVIEW 3 major objections 5 minor 68 references

Prompt components in multimodal entity alignment act as sequential fine-tuning steps, so curriculum-style prompts can replace black-box prompting.

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

2026-07-14 11:00 UTC pith:54BGKFNN

load-bearing objection Solid multi-component extension of ICL-as-fine-tuning plus a practical curriculum re-ranker that actually works; the multi-call pipeline sits outside the one-step dual-model math, but the empirical package is strong enough to keep. the 3 major comments →

arxiv 2607.10532 v1 pith:54BGKFNN submitted 2026-07-12 cs.IR

Implicit Fine-tuning via Context Engineering: A Curriculum Learning Framework for Multimodal Entity Alignment

Yunpeng Hong , Chenyang Bu , Di Wu , Yi He , Xindong Wu This is my paper
classification cs.IR
keywords Multimodal Entity AlignmentContext EngineeringFine-tuningCurriculum LearningIn-context LearningLarge Language ModelsKnowledge 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 argues that the usual practice of stuffing multimodal entity evidence into an LLM prompt is not a mysterious black box. Under a simplified attention model, each block of the prompt is mathematically equivalent to one gradient step of contrastive fine-tuning on a dual linear model. Because of that equivalence, the order in which information is shown matters: feeding easy evidence first and harder evidence later is the same as curriculum fine-tuning. The authors turn the insight into PTFEA, a three-stage progressive prompting scheme that starts with names only, then adds neighborhood structure, then full multimodal evidence, and that decides whether to escalate by an adaptive confidence gate. On five public datasets the scheme both raises Hits@1 and cuts token use and wall-clock time by more than 80 percent relative to a strong prior LLM baseline, while nearly closing the performance gap between 72B and 14B models. A sympathetic reader cares because the result supplies a first principled bridge between two previously separate toolkits—prompt engineering and parameter fine-tuning—for a practically important information-retrieval task.

Core claim

In multimodal entity alignment, the output of a single-layer softmax attention applied to a multi-component prompt is identical to the prediction of a dual linear-attention model after one gradient step under a contrastive loss whose terms correspond one-to-one with the prompt blocks; therefore curriculum-style selection and ordering of those blocks is formally equivalent to curriculum fine-tuning.

What carries the argument

Theorem 1 (and Remark 2): the equivalence that rewrites multi-component softmax attention as sequential contrastive gradient steps, with binary gates v_j that simultaneously select loss terms in fine-tuning and evidence blocks in the prompt.

Load-bearing premise

The equivalence is proved only for a single frozen softmax attention layer under a random-feature kernel approximation; if multi-layer residual or feed-forward interactions break that dual-model correspondence, the curriculum translation loses its theoretical warrant.

What would settle it

Train the dual linear model for one gradient step on the same multi-component token matrices used as prompts and check whether the L2 distance between its prediction and the real softmax-attention output remains near zero once the number of tokens reaches the paper’s reported N=18; a persistently large gap would falsify the claimed equivalence.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Any fine-tuning schedule that admits an ordered, re-weighted loss can be re-expressed as a corresponding progressive prompt schedule for MMEA.
  • Smaller LLMs can approach the alignment accuracy of much larger models once evidence is staged by difficulty rather than presented all at once.
  • Token budgets and wall-clock time for LLM re-ranking of multimodal entity pairs can be cut by more than 80 percent without sacrificing Hits@1.
  • The same dual-model view supplies a concrete criterion (confidence gates) for deciding when additional multimodal evidence is worth injecting.
  • Future work can systematically port other fine-tuning techniques (e.g., hard-example mining, multi-task curricula) into context-engineering designs for entity alignment.

Where Pith is reading between the lines

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

  • The same equivalence may extend, with suitable reformulation, to other multi-evidence LLM re-ranking tasks such as multimodal retrieval or cross-document coreference.
  • If residual multi-layer interactions can be absorbed into an effective dual model, the curriculum principle could be applied inside the frozen LLM itself rather than only at the prompt level.
  • Adaptive confidence gating offers a practical, model-agnostic alternative to fixed prompt templates that currently dominate production RAG systems.
  • The dramatic reduction in tokens suggests that many existing multimodal LLM pipelines are over-provisioning context; a curriculum audit could yield similar savings elsewhere.

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 argues that, in multimodal entity alignment (MMEA), multi-component context engineering is mathematically equivalent to one-step contrastive fine-tuning of a dual linear-attention model (Theorem 1, Eq. 6), and that curriculum-style gating of prompt blocks corresponds to curriculum sample selection (Remark 2, Eqs. 7–8). Building on this, it proposes PTFEA: hybrid embedding retrieval followed by three-stage progressive LLM re-ranking (name-only → neighborhood → full multimodal) with adaptive confidence-based early exit. Experiments on five public datasets report SOTA Hits@1, a 0.6% H@1 gap between Qwen2.5-72B and 14B on ICWIKI, and large efficiency gains versus MM-ChatAlign (runtime ~21h→1h; tokens ~2200–3000→200–400).

Significance. If the claimed CE–fine-tuning unification holds under the stated assumptions, this is a useful first theoretical bridge for LLM-based MMEA and a concrete template for turning fine-tuning ideas into prompt design. Strengths that should be credited: a careful single-layer dual-model derivation extended to multi-component MMEA prompts (Appendix A.1); an empirical check that the dual-model vs. attention representation gap shrinks to zero after N steps (Fig. 3); strong multi-dataset results against ~20 baselines; stage-wise and model-scale ablations; and substantial, reproducible efficiency gains with public code. Even under a weaker reading (curriculum as design principle rather than multi-step dual dynamics), the practical contribution—capacity-robust progressive re-ranking with large token savings—is significant for the IR/KDD community.

major comments (3)
  1. [§3.2 Theorem 1, Remark 1–2; §4.3; abstract] Theorem 1 / Remark 1 establish a one-step dual-model equivalence for a single multi-component prompt under one frozen softmax attention layer (Eq. 6; Appendix A.1). PTFEA, however, is implemented as up to three independent LLM calls with adaptive early exit (§4.3, Eqs. 10–13). Dual weights are not carried across stages, so the multi-step curriculum trajectory suggested by “mirroring the gradient descent process” (abstract; §3.2.2; Fig. 2) is not literally realized. Remark 2’s binary gates (Eqs. 7–8) justify selecting which evidence blocks enter a prompt, but not sequential multi-call optimization. Please either (i) reframe PTFEA as curriculum-inspired gated composition justified by Remark 2 rather than multi-step dual GD, or (ii) extend the theory to multi-call early-exit and state the scope conditions explicitly.
  2. [§5.3.1 Fig. 4; Table 5; Remark 2] The curriculum claim is load-bearing for the method’s novelty, yet the ablations only remove stages or force all entities through Stage III (Fig. 4; Table 5). There is no reverse-order control (hard evidence first, then easy) or random-order control that would test whether easy-to-hard ordering—not merely “more information eventually”—drives the gains. Without that, the link from Remark 2 to the three-stage design remains under-supported. A reverse/random curriculum ablation on at least ICWIKI (and ideally one DBP15K subset) is needed.
  3. [§3.1–3.2; Appendix A.1; Fig. 3] The entire equivalence rests on reducing a multi-layer Transformer to a single softmax attention layer with frozen W_K, W_Q, W_V and a random-feature kernel approximation of softmax (Appendix A.1, following Ren & Liu 2024). Residual multi-layer interactions and FFN nonlinearities are outside the dual model. The manuscript should state this as a scope condition of Theorem 1, discuss what would falsify the correspondence in practice, and avoid language that presents the multi-layer LLM used in experiments as covered by the same derivation without qualification.
minor comments (5)
  1. [§4.2–4.3; §5.1.3; Fig. 7] Free parameters α, candidate size k/window, adaptive η rule, and Stage-III modality weights w_i (§4.2–4.3, Eq. 9, Eq. 13) are only partially justified. A short sensitivity table for α and η (beyond the fixed-threshold plot in Fig. 7) would strengthen robustness claims.
  2. [§5.2.1 Table 2] The drop in H@10 on DBP15K relative to some embedding baselines is acknowledged (§5.2.1) but left somewhat open. Clarify whether Stage-III re-ranking can optionally fall back to embedding ranks when LLM scores are low-confidence.
  3. [Fig. 2] Fig. 2 is dense; the three equivalence axes (data organization / sample selection / outputs) would be clearer with a short caption table mapping Eq. 6–8 to PTFEA stages.
  4. [§3.1–3.2] Notation: both C and c appear for confidence/normalization; 𝒉 vs h and W vs 𝑊 are mixed. Unify early in §3.1.
  5. [Abstract; §1] Typos/style: “fivepublic”, “𝐻 @1gap”, “from 21 hours to 1 hourand”, and occasional spacing around math operators in the abstract and §1.

Circularity Check

1 steps flagged

No load-bearing circularity: Theorem 1 is a one-way dual derivation from attention; Remark 2's curriculum map is a constructed correspondence, not a forced prediction.

specific steps
  1. self definitional [Sec. 3.2.2 Remark 2, Eqs. 7–8]
    "With a difficulty-aware gate v_j, curriculum learning and CE admit a direct correspondence in how training signals are scheduled and injected: L=−1/(ξC) ∑_{j=1}^k v_j L_j ⇔ X=∑_{j=1}^k v_j X_j where curriculum learning uses v_j to select/reweight the j-th loss term, while CE uses the same v_j to schedule whether the j-th evidence block is composed into the prompt. Moreover, v_j can be determined by a confidence-based difficulty criterion: v_j=I(c_j^{(t)}≥τ^{(t)}) ⇔ v_j=I(c_j^{(t)}≥η)"

    The claimed 'mathematical equivalence' between curriculum learning weights and context sample selection is obtained by writing the identical indicator gate v_j on both sides of a biconditional. The correspondence is therefore true by the authors' definition of the map, not by an independent derivation that could have failed. Mild and non-load-bearing: it designs PTFEA's stage gates; it does not force Theorem 1 or the reported H@1 numbers.

full rationale

The paper's central theoretical claim (Theorem 1 / Eq. 6) is obtained by rewriting multi-component softmax attention under a frozen single-layer model and exhibiting an equivalent one-step gradient update on the dual linear-attention model f(x)=W φ(x), following the external Ren & Liu (2024) / Dai et al. (2023) ICL-as-GD framework. Training pairs, labels, and W0 are read off the prompt tokens by construction of the dual, then the attention output is shown to match y_test; this is a standard dualization, not reverse-engineering a target metric. Experimental Fig. 3 checks that the representation gap shrinks under that dual setup; it does not fit a parameter and re-label it as a prediction. Self-citations are limited to embedding baselines used only for candidate retrieval and are not load-bearing for the equivalence. The only mild self-definitional flavor is Remark 2 (Eqs. 7–8), where the same binary gate v_j is placed on both the curriculum loss and the prompt composition and then called a mathematical equivalence between curriculum weights and context sample selection—the correspondence is defined rather than independently derived. That is framework design, not a circular prediction of the experimental H@1 numbers. The skeptic's theory–practice gap (one-step dual vs multi-call PTFEA stages that never carry dual weights forward) is a warrant/correctness concern, not circularity under the stated patterns. Overall circularity is negligible.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claim rests on the single-layer dual-model construction imported from Ren & Liu 2024, the random-feature kernel, frozen attention weights, and a handful of engineering thresholds chosen for the progressive stages. No new physical entities are postulated; the free parameters are ordinary hyper-parameters of the retrieval and gating pipeline.

free parameters (4)
  • hybrid-retrieval weight α = 0.0 / 0.2
    Balances MEAformer vs Simple-HHEA similarities; set to 0.0 on DBP15K and 0.2 on ICWIKI/ICYAGO by hand.
  • adaptive confidence threshold η = dynamic (last-10 avg or Q/2)
    Computed from the average of the last ten confidence scores or Q/2; controls stage transitions and therefore which evidence blocks enter the prompt.
  • candidate list size k / window size = window=5
    Top-k retrieval and neighborhood window size 5 determine how much structural evidence is available for Stage II.
  • Stage-III modality weights w_i
    Linear combination of name, image, relation scores; values not exhaustively ablated.
axioms (4)
  • domain assumption Softmax attention with random-feature kernel is dual to a linear attention model whose one-step gradient update matches the attention output (imported from Ren & Liu 2024).
    Theorem 1 and Appendix A.1 rest entirely on this duality; multi-layer residuals are ignored.
  • domain assumption Attention-layer parameters W_K, W_Q, W_V remain fixed after initialization (no actual parameter update occurs during prompting).
    Stated in §3.2.1; required for the dual-model interpretation of frozen LLMs.
  • domain assumption Entity evidence admits a natural easy-to-hard ordering (names < local structure < full multimodal triples).
    Justifies the curriculum schedule; empirically plausible for MMEA but not proved.
  • ad hoc to paper LLM-reported confidence scores are sufficiently calibrated to serve as a difficulty gate.
    Adaptive Difficulty Modulation (§4.3.1) treats raw confidence as a reliable sample-selection signal.
invented entities (1)
  • PTFEA three-stage progressive inference pipeline independent evidence
    purpose: Operationalizes the curriculum equivalence by injecting name-only, then neighborhood, then full multimodal evidence under confidence gates.
    The concrete algorithm is new; it is an engineering construct rather than a postulated physical entity, and its independent evidence is the empirical ablation suite.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Implicit Fine-tuning via Context Engineering: A Curriculum Learning Framework for Multimodal Entity Alignment." pith.science (2026). https://pith.science/paper/54BGKFNN

@misc{pith2026260710532,
  author       = {Pith},
  title        = {Pith review of: Implicit Fine-tuning via Context Engineering: A Curriculum Learning Framework for Multimodal Entity Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/54BGKFNN}},
  note         = {Machine review of arXiv:2607.10532}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multimodal Entity Alignment (MMEA) aims to identify equivalent entities across different modalities. While existing methods enhance MMEA performance through black-box context engineering strategies, their reliance on LLM parameter capacity and lack of theoretical interpretability remain unresolved. To this end, we first theoretically validate the mathematical equivalence between context engineering and model fine-tuning in MMEA tasks, demonstrating that prompt components simulate contrastive learning-based sequential fine-tuning in MMEA. Building on this foundation, we then propose PTFEA, a curriculum-learning-inspired framework that translates fine-tuning strategies into interpretable context engineering. Specifically, adaptive difficulty modulation dynamically adjusts information injection stages using confidence thresholds, establishing mathematical equivalence between curriculum learning weights and context sample selection; and three-stage progressive inference incorporates entity information from simple to complex cases, mirroring the gradient descent process in fine-tuning. Experiments on five public datasets demonstrate that PTFEA consistently outperforms strong baselines. In particular, on the ICWIKI dataset, PTFEA narrows the H@1 gap between Qwen2.5-72B and 14B to 0.6%. Moreover, compared with the representative context-engineering-based MMEA method MM-ChatAlign, PTFEA reduces the runtime of Qwen2.5-72B from 21 hours to 1 hour and lowers token consumption from 2200-3000 to 200-400, achieving over 80% reduction on the ICWIKI dataset. This work provides the first theoretical framework unifying context engineering and fine-tuning in MMEA, paving the way for future research that seeks to translate additional fine-tuning strategies into context engineering paradigms. Our code is available at https://github.com/DMiC-Lab-HFUT/PTFEA.

Figures

Figures reproduced from arXiv: 2607.10532 by Chenyang Bu, Di Wu, Xindong Wu, Yi He, Yunpeng Hong.

Figure 1
Figure 1. Figure 1: Example illustrating that the black-box nature of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the equivalence between curriculum learning-based fine-tuning and the proposed context engineering [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The representation gap between fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Token consumption statistics of PTFEA and MM [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance comparison of different LLMs on the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison results of fixed confidence thresholds [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

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

68 extracted references · 1 canonical work pages

  1. [1]

    Chirag Agarwal, Daniel D’souza, and Sara Hooker. 2022. Estimating example difficulty using variance of gradients. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 10368–10378

    2022

  2. [2]

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Cur- riculum learning. InProceedings of International Conference on Machine Learning. ACM, 41–48

    2009

  3. [3]

    Chenyang Bu, Guojie Chang, Zihao Chen, CunYuan Dang, Zhize Wu, Yi He, and Xindong Wu. 2025. Query-Driven Multimodal GraphRAG: Dynamic Local Knowledge Graph Construction for Online Reasoning. InFindings of the Associa- tion for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 - August 1,

    2025

  4. [4]

    Liyi Chen, Zhi Li, Yijun Wang, Tong Xu, Zhefeng Wang, and Enhong Chen

  5. [5]

    InProceeding of Knowledge Science, Engineering and Management

    Mmea: Entity alignment for multi-modal knowledge graph. InProceeding of Knowledge Science, Engineering and Management. Springer, 134–147

  6. [6]

    Liyi Chen, Zhi Li, Tong Xu, Han Wu, Zhefeng Wang, Nicholas Jing Yuan, and Enhong Chen. 2022. Multi-modal siamese network for entity alignment. In Proceedings of SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 118–126

    2022

  7. [7]

    Liyi Chen, Ying Sun, Shengzhe Zhang, Yuyang Ye, Wei Wu, and Hui Xiong. 2024. Tackling uncertain correspondences for multi-modal entity alignment.Proceed- ings of Advances in Neural Information Processing Systems37 (2024), 119386– 119410

    2024

  8. [8]

    Shengyuan Chen, Qinggang Zhang, Junnan Dong, Wen Hua, Qing Li, and Xiao Huang. 2024. Entity alignment with noisy annotations from large language models.Proceedings of Advances in Neural Information Processing Systems37 (2024), 15097–15120

    2024

  9. [9]

    Xuan Chen, Tong Lu, and Zhichun Wang. 2024. LLM-Align: Utilizing Large Language Models for Entity Alignment in Knowledge Graphs. arXiv:2412.04690 https://arxiv.org/abs/2412.04690

    Pith/arXiv arXiv 2024

  10. [10]

    Zihao Chen, Chenyang Bu, Shengwei Ji, and Xindong Wu. 2026. Dual-Branch Multi-Granularity Network with Structured Contrastive Ranking for Cross-Modal Retrieval. InProceedings of the ACM Web Conference 2026. ACM, 1959–1970

    2026

  11. [11]

    Pan, Wenting Song, and Huajun Chen

    Zhuo Chen, Jiaoyan Chen, Wen Zhang, Lingbing Guo, Yin Fang, Yufeng Huang, Yichi Zhang, Yuxia Geng, Jeff Z. Pan, Wenting Song, and Huajun Chen. 2023. Meaformer: Multi-modal entity alignment transformer for meta modality hybrid. InProceedings of International Conference on Multimedia. ACM, 3317–3327

    2023

  12. [12]

    Zhuo Chen, Yin Fang, Yichi Zhang, Lingbing Guo, Jiaoyan Chen, Jeff Z Pan, Huajun Chen, and Wen Zhang. 2025. Noise-powered multi-modal knowledge graph representation framework. InProceedings of International Conference on Computational Linguistics. ACM, 141–155

    2025

  13. [13]

    Jingwei Cheng, Mingxiao Guo, and Fu Zhang. 2025. SGMEA: Structure-guided multimodal entity alignment. InProceedings of International Conference on Com- putational Linguistics. ACL, 7851–7861

    2025

  14. [14]

    Gonzalez

    Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, Nicholas Thumiger, Aditya Desai, Ion Stoica, Ana Klimovic, Graham Neubig, and Joseph E. Gonzalez. 2025. The danger of overthinking: Examining the reasoning-action dilemma in agentic tasks. https://arxiv.org/abs/2502.08235

    Pith/arXiv arXiv 2025

  15. [15]

    Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Shuming Ma, Zhifang Sui, and Furu Wei

  16. [16]

    InFindings of the Association for Computational Linguistics

    Why can gpt learn in-context? Language models secretly perform gradient descent as meta-optimizers. InFindings of the Association for Computational Linguistics. ACL, 4005–4019

  17. [17]

    Yasha Ektefaie, George Dasoulas, Ayush Noori, Maha Farhat, and Marinka Zitnik

  18. [18]

    Multimodal learning with graphs.Nature Machine Intelligence5, 4 (2023), 340–350

    2023

  19. [19]

    Guan Dong Feng, Tao Ren, Jun Hu, and Dan dan Wang. 2025. CateEA: En- hancing Entity Alignment via Implicit Category Supervision. InProceedings of International Conference on Computational Linguistics. ACL, 5975–5986

    2025

  20. [20]

    Junliang Guo, Xu Tan, Linli Xu, Tao Qin, Enhong Chen, and Tie-Yan Liu. 2020. Fine-tuning by curriculum learning for non-autoregressive neural machine trans- lation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. AAAI Press, 7839–7846

    2020

  21. [21]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep resid- ual learning for image recognition. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 770–778

    2016

  22. [22]

    Yunpeng Hong, Chenyang Bu, Jie Zhang, Yi He, Di Wu, and Xindong Wu. 2026. PSQE: A Theoretical-Practical Approach to Pseudo Seed Quality Enhancement for Unsupervised Multimodal Entity Alignment. InProceedings of ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 464–475

    2026

  23. [23]

    Manzong Huang, Chenyang Bu, Yi He, and Xindong Wu. 2025. How to Mitigate Information Loss in Knowledge Graphs for GraphRAG: Leveraging Triple Context Restoration and Query-Driven Feedback. InProceedings of International Joint Conference on Artificial Intelligence. 8104–8112

    2025

  24. [24]

    Manzong Huang, Chenyang Bu, Yi He, Xingrui Zhuo, and Xindong Wu. 2026. Relink: Constructing Query-Driven Evidence Graph On-the-Fly for GraphRAG. InProceedings of the AAAI Conference on Artificial Intelligence. AAAI Press, 31202– 31210

    2026

  25. [25]

    Yani Huang, Xuefeng Zhang, Richong Zhang, Junfan Chen, and Jaein Kim. 2024. Progressively modality freezing for multi-modal entity alignment. InProceedings of Association for Computational Linguistics. ACM, 3477–3489

    2024

  26. [26]

    Xuhui Jiang, Yinghan Shen, Zhichao Shi, Chengjin Xu, Wei Li, Zixuan Li, Jian Guo, Huawei Shen, and Yuanzhuo Wang. 2024. Unlocking the Power of Large Language Models for Entity Alignment. InProceedings of Association for Computational Linguistics. ACL, 7566–7583

    2024

  27. [27]

    Xuhui Jiang, Yinghan Shen, Zhichao Shi, Chengjin Xu, Wei Li, Huang Zihe, Jian Guo, and Yuanzhuo Wang. 2024. MM-ChatAlign: A novel multimodal reasoning framework based on large language models for entity alignment. InFindings of the Association for Computational Linguistics: EMNLP. ACM, 2637–2654

    2024

  28. [28]

    Xuhui Jiang, Chengjin Xu, Yinghan Shen, Yuanzhuo Wang, Fenglong Su, Zhichao Shi, Fei Sun, Zixuan Li, Jian Guo, and Huawei Shen. 2024. Toward practical entity alignment method design: Insights from new highly heterogeneous knowledge graph datasets. InProceedings of the ACM Web Conference. ACM, 2325–2336

    2024

  29. [29]

    Chenxiao Li, Jingwei Cheng, Qiang Tong, and Fu Zhang. 2025. Exploring the Im- pacts of Feature Fusion Strategy in Multi-modal Entity Alignment. InProceedings of International Conference on Computational Linguistics. ACL, 7809–7818

    2025

  30. [30]

    Qian Li, Shu Guo, Yangyifei Luo, Cheng Ji, Lihong Wang, Jiawei Sheng, and Jianxin Li. 2023. Attribute-consistent knowledge graph representation learning for multi-modal entity alignment. InProceedings of the ACM Web Conference. ACM, 2499–2508

    2023

  31. [31]

    Zhenxi Lin, Ziheng Zhang, Meng Wang, Yinghui Shi, Xian Wu, and Yefeng Zheng

  32. [32]

    In Proceedings of International Conference on Computational Linguistics

    Multi-modal contrastive representation learning for entity alignment. In Proceedings of International Conference on Computational Linguistics. ACL, 2572– 2584

  33. [33]

    Fangyu Liu, Muhao Chen, Dan Roth, and Nigel Collier. 2021. Visual pivoting for (unsupervised) entity alignment. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. AAAI Press, 4257–4266

    2021

  34. [34]

    Zhiyuan Liu, Yixin Cao, Liangming Pan, Juanzi Li, Zhiyuan Liu, and Tat-Seng Chua. 2020. Exploring and evaluating attributes, values, and structures for entity alignment. InProceedings of Empirical Methods in Natural Language Processing. ACL, 6355–6364

    2020

  35. [35]

    Chenglong Lu, Chenxiao Li, Jingwei Cheng, Yongquan Ji, Guoqing Chen, and Fu Zhang. 2025. Breaking the Noise Barrier: LLM-Guided Semantic Filtering and Enhancement for Multi-Modal Entity Alignment. InProceedings of Conference on Empirical Methods in Natural Language Processing. ACM, 33141–33155

    2025

  36. [36]

    Xin Mao, Wenting Wang, Yuanbin Wu, and Man Lan. 2021. Boosting the speed of entity alignment 10×: Dual attention matching network with normalized hard sample mining. InProceedings of the ACM Web Conference. ACM, 821–832

    2021

  37. [37]

    Marwa Naïr, Kamel Yamani, Lynda Lhadj, and Riyadh Baghdadi. 2024. Curriculum learning for small code language models. InProceedings of the Association for Computational Linguistics. ACL, 390–401

    2024

  38. [38]

    Wenxin Ni, Qianqian Xu, Yangbangyan Jiang, Zongsheng Cao, Xiaochun Cao, and Qingming Huang. 2023. Psnea: Pseudo-siamese network for entity align- ment between multi-modal knowledge graphs. InProceedings of International Conference on Multimedia. ACM, 3489–3497

    2023

  39. [39]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual mod- els from natural language supervision. InInternational Conference on Machine Learning. PMLR, 8748–8763

    2021

  40. [40]

    Ruifeng Ren and Yong Liu. 2024. Towards Understanding How Transformers Learn In-context Through a Representation Learning Lens. InProceedings of Advances in Neural Information Processing Systems, Vol. 37. Curran Associates, Inc., 892–933

    2024

  41. [41]

    Subham Sekhar Sahoo, Justin Deschenaux, Aaron Gokaslan, Guanghan Wang, Justin T Chiu, and Volodymyr Kuleshov. 2025. The Diffusion Duality. InProceed- ings of International Conference on Machine Learning. https://openreview.net/ forum?id=9P9Y8FOSOk

    2025

  42. [42]

    Taoyu Su, Jiawei Sheng, Duohe Ma, Xiaodong Li, Juwei Yue, Mengxiao Song, Yingkai Tang, and Tingwen Liu. 2025. Mitigating Modality Bias in Multi-modal Entity Alignment from a Causal Perspective. InProceedings of International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 1186–1196

    2025

  43. [43]

    Taoyu Su, Jiawei Sheng, Shicheng Wang, Xinghua Zhang, Hongbo Xu, and Tingwen Liu. 2024. Ibmea: Exploring variational information bottleneck for multi-modal entity alignment. InProceedings of ACM International Conference on Multimedia. ACM, 4436–4445

    2024

  44. [44]

    Taoyu Su, Xinghua Zhang, Jiawei Sheng, Zhenyu Zhang, and Tingwen Liu

  45. [45]

    InProceedings of European Conference on Artificial Intelligence, Vol

    LoginMEA: Local-to-Global Interaction Network for Multi-Modal Entity Alignment. InProceedings of European Conference on Artificial Intelligence, Vol. 392. 1173–1180

  46. [46]

    Zequn Sun, Wei Hu, Qingheng Zhang, and Yuzhong Qu. 2018. Bootstrapping en- tity alignment with knowledge graph embedding. InProceedings of International Joint Conference on Artificial Intelligence. AAAI, 4396–4402. Implicit Fine-tuning via Context Engineering: A Curriculum Learning Framework for MMEA KDD ’26, August 09–13, 2026, Jeju Island, Republic of Korea

    2018

  47. [47]

    Cunda Wang, Weihua Wang, Xinyu Li, Qiuyu Liang, and Feilong Bao. 2025. OTMEA: Multi-modal Entity Alignment via Optimal Transport. InProceedings of IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 1–5

    2025

  48. [48]

    Yiru Wang, Weihao Gan, Jie Yang, Wei Wu, and Junjie Yan. 2019. Dynamic curriculum learning for imbalanced data classification. InProceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, 5017–5026

    2019

  49. [49]

    Yuanyi Wang, Haifeng Sun, Jiabo Wang, Jingyu Wang, Wei Tang, Qi Qi, Shaoling Sun, and Jianxin Liao. 2024. Towards semantic consistency: Dirichlet energy driven robust multi-modal entity alignment. InProceedings of IEEE International Conference on Data Engineering. IEEE, 3559–3572

    2024

  50. [50]

    Zilong Wang, Jiongda Wang, Xiaoyong Chen, Meng Wang, Ming Ma, Zhipeng Wang, Zhenyu Zhou, Tianming Yang, and Wang-Zhou Dai. 2025. From end-to- end to step-by-step: Learning to abstract via abductive reinforcement learning. InProceedings of International Joint Conference on Artificial Intelligence. IJCAI Organization, 6515–6523

    2025

  51. [51]

    Tianle Xia, Liang Ding, Guojia Wan, Yibing Zhan, Bo Du, and Dacheng Tao

  52. [52]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Improving complex reasoning over knowledge graph with logic-aware curriculum tuning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. AAAI Press, 12881–12889

  53. [53]

    Baogui Xu, Chengjin Xu, and Bing Su. 2023. Cross-modal graph attention net- work for entity alignment. InProceedings of ACM International Conference on Multimedia. ACM, 3715–3723

    2023

  54. [54]

    Linyao Yang, Hongyang Chen, Xiao Wang, Jing Yang, Fei-Yue Wang, and Han Liu. 2024. Two heads are better than one: Integrating knowledge from knowledge graphs and large language models for entity alignment. arXiv:2401.16960 https: //arxiv.org/abs/2401.16960

    Pith/arXiv arXiv 2024

  55. [55]

    Linyan Yang, Jingwei Cheng, and Fu Zhang. 2024. Advancing Cross-Lingual En- tity Alignment with Large Language Models: Tailored Sample Segmentation and Zero-Shot Prompts. InFindings of the Association for Computational Linguistics: EMNLP. ACL, 8122–8138

    2024

  56. [56]

    Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. 2023. The dawn of LMMs: Preliminary explorations with GPT-4V(ision). arXiv:2309.17421 https://arxiv.org/abs/2309.17421

    Pith/arXiv arXiv 2023

  57. [57]

    Shiji Zang, Chenyang Bu, Yunpeng Hong, He Ren, and Weiping Ding. 2026. Med- ical multimodal entity linking under modality missingness.Artificial Intelligence in Health(2026), 026070015. doi:10.36922/AIH026070015

    work page doi:10.36922/aih026070015 2026

  58. [58]

    Yan Zhang, Xiangyu Luo, Jing Hu, Miao Zhang, Kui Xiao, and Zhifei Li. 2025. Graph structure prefix injection transformer for multi-modal entity alignment. Information Processing & Management62, 3 (2025), 104048

    2025

  59. [59]

    Yu Zhao, Yike Wu, Xiangrui Cai, Ying Zhang, Haiwei Zhang, and Xiaojie Yuan

  60. [60]

    InFindings of the Association for Computational Linguistics

    From alignment to entailment: A Uuified textual entailment framework for entity alignment. InFindings of the Association for Computational Linguistics. ACL, 8795–8806

  61. [61]

    Yu Zhao, Ying Zhang, Xuhui Sui, and Xiangrui Cai. 2025. ME3A: A Multimodal Entity Entailment framework for multimodal Entity Alignment.Information Processing & Management62, 1 (2025), 103951

    2025

  62. [62]

    Liangli Zhen, Peng Hu, Xu Wang, and Dezhong Peng. 2019. Deep supervised cross-modal retrieval. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 10394–10403

    2019

  63. [63]

    Ziyue Zhong, Meihui Zhang, Ju Fan, and Chenxiao Dou. 2022. Semantics driven embedding learning for effective entity alignment. InProceedings of IEEE Interna- tional Conference on Data Engineering. IEEE, 2127–2140

    2022

  64. [64]

    Yunfeng Zhou, Cui Zhu, Wenjun Zhu, and Hongyang Li. 2024. SCMEA: A stacked co-enhanced model for entity alignment based on multi-aspect information fusion and bidirectional contrastive learning.Neural Networks173 (2024), 106178

    2024

  65. [65]

    Bin Zhu, Meng Wu, Yunpeng Hong, Yi Chen, Bo Xie, Fei Liu, Chenyang Bu, and Weiping Ding. 2023. MMIEA: Multi-modal interaction entity alignment model for knowledge graphs.Information Fusion100 (2023), 101935

    2023

  66. [66]

    Xingrui Zhuo, Shirui Pan, Jiapu Wang, Gongqing Wu, Zan Zhang, Rui Li, Zizhong Wei, and Xindong Wu. 2025. Progressive Prefix-Memory Tuning for Complex Logical Query Answering on Knowledge Graphs. InProceedings of International Joint Conference on Artificial Intelligence. 3716–3724

    2025

  67. [67]

    Xingrui Zhuo, Jiapu Wang, Gongqing Wu, Shirui Pan, and Xindong Wu. 2025. Effective Instruction Parsing Plugin for Complex Logical Query Answering on Knowledge Graphs. InProceedings of the ACM on Web Conference. ACM, 4780– 4792

    2025

  68. [68]

    Xingrui Zhuo, Jiapu Wang, Gongqing Wu, Zhongyuan Wang, Jichen Zhang, Shirui Pan, and Xindong Wu. 2026. Knowledge Reasoning Language Model: Unifying Knowledge and Language for Inductive Knowledge Graph Reasoning. InProceedings of International Conference on Learning Representations. https: //openreview.net/forum?id=2g8EmFwNTB A Complementary Theoretical An...

    2026

This paper was first reviewed by grok-4.5 on July 14, 2026.