Pith. sign in

REVIEW 4 major objections 5 minor 39 references

SGCL: Unifying Self-Supervised and Supervised Learning for Graph Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single supervised graph contrastive loss replaces both the BPR recommendation loss and the InfoNCE self-supervised loss in graph recommendation, improving accuracy and cutting training time.

desk verdict An efficiency-focused graph recommender whose central loss claim collapses to an alignment-uniformity objective; the paper needs major revision before its unification story is credible. read the letter →

arxiv 2507.13336 v1 pith:JEBLRI6P submitted 2025-07-17 cs.IR

Weizhi Zhang , Liangwei Yang , Zihe Song , Henrry Peng Zou , Ke Xu , Yuanjie Zhu , Philip S. Yu This is my paper
classification cs.IR
keywords graphrecommendationsupervisedcontrastivelearningself-supervisedcollaborativefilteringbipartitetrainingefficiencyrecommendersystems
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

SGCL is a graph recommender built around a single supervised graph contrastive loss. Observed user-item interactions act as positive pairs, and pairwise similarities among users and among items in the batch fill the contrastive denominator, so the one objective is meant to supply both the ranking signal of BPR and the representation-shaping signal of InfoNCE. The paper argues this removes the gradient inconsistency of multi-task SSL recommenders, which typically optimize $\mathcal{L}_{\mathrm{rec}} + \lambda \mathcal{L}_{\mathrm{gcl}}$ with separate graph encoders and augmentation branches. If the claim holds, graph recommendation training becomes simpler, faster, and more accurate: on Amazon Beauty and Toys-and-Games, SGCL reports higher Recall and NDCG than the compared baselines while using far fewer epochs and less total training time.

What carries the argument

The load-bearing object is the supervised graph contrastive loss of Eq. 5: a cross-entropy-style contrastive objective in which the positive pair is a user and an item linked by an observed interaction, and the negatives are the other users and items in the same batch through their pairwise similarity terms $u_i^\top u_{i'}/\tau$ and $v_j^\top v_{j'}/\tau$. It does the work of both task losses at once: the positive term supplies the ranking supervision of BPR, while the denominator supplies the repulsion that the InfoNCE loss would normally provide, without any graph augmentation or explicit negative sampling. The encoder is the light graph convolution of LightGCN, with layerwise embeddings combined as $u_i=\sum_k \alpha_k u_i^k$ and $v_j=\sum_k \alpha_k v_j^k$; SGCL keeps this encoder but strips away the second encoder and the perturbation branches used by multi-task SSL recommenders.

What would settle it

Run SGCL twice on one dataset, once with an explicit mask that excludes $i'=i$ and $j'=j$ from the denominator and once with the formula exactly as printed; then compare Recall@20 and the average norm of user and item embeddings. If the two versions behave almost identically, the masking question is irrelevant to the reported gains; if the unmasked version performs better but has much smaller embedding norms, the improvement may come from norm shrinkage rather than from the unified contrastive objective.

Watch

Extended reading notes

Core claim

The central claim is that recommendation and self-supervised graph learning can be one task. SGCL replaces the usual objective $\mathcal{L}_{\mathrm{sslrec}} = \mathcal{L}_{\mathrm{rec}} + \lambda \mathcal{L}_{\mathrm{gcl}}$ with a single supervised graph contrastive loss, $$\mathcal{L}_{\mathrm{sgcl}} = \sum_{(i,j)\in\mathcal{B}} -\log \frac{\exp(u_i^\top v_j/\tau)}{\sum_{(i',j')\in\mathcal{B}}\left(\exp(u_i^\top u_{i'}/\tau) + \exp(v_j^\top v_{j'}/\tau)\right)},$$ where each observed interaction $(i,j)$ is a positive pair and the denominator compares that user with other users and that item with other items in the batch. The paper argues that this one loss both teaches the model which items a user prefers and shapes the embedding space through contrast, so the separate BPR loss, the InfoNCE loss, the graph augmentation, and the negative-sampling step are no longer needed. On the reported datasets, SGCL reaches higher Recall@20, Recall@50, NDCG@20, and NDCG@50 than the baselines while converging in fewer epochs.

Load-bearing premise

The load-bearing premise is that the denominator of the loss behaves as proper negative examples; the paper never says whether an embedding is compared with itself in that denominator, and if it is, the loss may be rewarding small embedding sizes rather than good recommendations.

Editorial extensions

If this is right

  • Recommendation training reduces to a single optimization objective, so the manual weight $\lambda$ that balances BPR and contrastive losses disappears.
  • Graph augmentation, negative sampling, and redundant graph convolution are removed from the training pipeline, which is what makes each SGCL epoch cheaper than SGL-style baselines.
  • SGCL converges faster than LightGCN and SGL: on the reported datasets it reaches stable performance within the first 30 epochs, while the baselines are still far from convergence.
  • On Beauty and Toys-and-Games, SGCL reports the highest Recall@20, Recall@50, NDCG@20, and NDCG@50 among the compared methods, including the SSL baselines NCL, SGL, LightGCL, and SimGCL.
  • Total training time is lower than LightGCN and SGL on both datasets, despite similar per-epoch cost on Toys-and-Games, because SGCL needs far fewer epochs.

Reading between the lines

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

  • Beyond the paper, a masked variant that removes $u_i^\top u_i$ and $v_j^\top v_j$ from the denominator would isolate whether the self-similarity terms affect embedding norms; the manuscript does not report this ablation.
  • Beyond the paper, the unification idea transfers to any two-stage loss with a ranking term and a contrastive term, so it could be tested in sequential or session-based recommenders that currently use a similar multi-task setup.
  • Beyond the paper, an ablation that re-adds explicit negative sampling to SGCL would show how much of the speed gain comes from the loss design versus from dropping sampler overhead.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SGCL, a graph recommender that replaces the standard two-loss pipeline (BPR ranking loss plus InfoNCE self-supervised loss) with a single supervised graph contrastive loss (Eq. 5). The encoder is a LightGCN-style graph convolution with layer-wise averaging, and the loss treats observed user-item interactions as positive pairs while the denominator sums user-user and item-item similarities in the batch. Experiments on Amazon Beauty and Toys-and-Games report higher Recall/NDCG than baselines and lower training time.

Significance. If the proposed loss genuinely combined ranking supervision with contrastive self-supervision, SGCL would be a valuable simplification: it removes augmentation, negative sampling, and the trade-off weight λ, and the reported convergence speed is practically attractive. The authors provide an open-source implementation and a time-complexity table, which are useful for reproducibility. However, the loss in Eq. 5 as written does not contain any non-interacted user-item pair, so the 'unification' claim is not supported; the objective is essentially an alignment-plus-uniformity loss in the spirit of DirectAU, which is not cited or compared. The experimental section also lacks statistical significance testing and lists only two datasets despite stating three. These issues are central to the paper's claims.

major comments (4)
  1. [§3.1, Eq. (5)] The denominator sums over (i',j') in B of exp(u_i^T u_i'/τ) + exp(v_j^T v_j'/τ). There is no term exp(u_i^T v_j'/τ) or exp(u_i'^T v_j/τ), so the loss never directly contrasts a positive interaction against a non-interacted user-item pair. Consequently the claim that Eq. (5) 'unifies' or 'replaces' the BPR ranking loss is not supported; BPR enforces u_i^T v_j > u_i^T v_j' for sampled negative items, while Eq. (5) only maximizes alignment of positive pairs and pushes users apart from users and items apart from items. The authors should either modify the denominator to include user-item negative pairs or reframe the contribution as an alignment/uniformity objective and compare against DirectAU.
  2. [§3.1, Eq. (5) implementation] The denominator includes the self-similarity terms exp(||u_i||^2/τ) and exp(||v_j||^2/τ) whenever (i,j) is in the batch. Unless these terms are masked or embeddings are normalized, the loss will contain a norm-shrinking gradient that can lead to trivial solutions. The paper does not specify how this is handled in practice. Please state the masking scheme (or normalization) and report whether the final embeddings are L2-normalized for ranking.
  3. [§4.1, Abstract, §4.3.3] The paper claims experiments on 'three real-world datasets' (Abstract; §4.3.3) but Table 2 and Table 3 report results on only Beauty and Toys-and-Games. Either add the third dataset or correct the text. Also, no standard deviations or significance tests are reported; several improvements over the best baseline are small (e.g., Beauty N@50: 0.0807 vs 0.0799 for SimGCL; Toys R@50: 0.1965 vs 0.1869 for SGL). Report variance over multiple seeds and test significance.
  4. [§3.2, Table 1] The complexity analysis is unclear. For Eq. (5), the denominator for each positive pair computes 2B dot products (B user-user and B item-item), so the total is O(2B^2 d) plus O(B d) for numerators, not obviously smaller than SGL's combined user- and item-side InfoNCE, which also has O(2B^2 d). Please clarify how the user-user and item-item similarity matrices are computed and why SGCL's GCL loss is reported as O(Bd+2B^2d) versus SGL's O(2Bd+2B^2d).
minor comments (5)
  1. [§3.2] The text 'While SGLC obviates the necessity of BPR loss computation' contains a typo: SGLC should be SGCL.
  2. [Figure 2] The axis label 'Average Trainig Time per Epoch' contains a typo; it should be 'Training'.
  3. [§4.1] The section says 'we conduct the experiments on two public real-world datasets' while the abstract says three; please make the dataset count consistent throughout.
  4. [§4.3.3] The sentence 'with the LightGCN and the SGL on three datasets' should be corrected to 'two datasets' unless a third dataset is added.
  5. [§3.1, Eq. (5) notation] The symbol B is used both for the batch of user-item interactions and for the number of items in negative sampling; please define it explicitly and use distinct notation if needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SGCL's loss is proposed and empirically tested; no fitted parameter or self-citation chain drives the results.

full rationale

The paper's central contribution is the SGCL loss in Eq. (5), and its central claims (accuracy, speed, convergence) are empirical results measured against a held-out test split after supervised training. The only tuned hyperparameter is the temperature tau = 0.2, chosen by validation; no parameter is fitted to the reported test metrics and then renamed a prediction. The loss is proposed, not derived from a theorem, so there is no derivation chain whose conclusion is identical to its premise. References to LightGCN, BPR, InfoNCE, and SSL baselines are external and standard; the authors' self-citations ([23], [24], [33], [34]) appear only as contextual RecSys work and are not load-bearing for the SGCL objective. The claim that a single loss removes multi-task gradient conflict is a design assertion, not a fitted result; whether the denominator of Eq. (5) actually supplies a BPR-style ranking signal (or degenerates to an alignment-plus-uniformity objective similar to DirectAU) is a correctness and novelty concern about the loss's interpretation, not circularity. No equation is defined in terms of the results it is used to predict, and no self-citation is invoked to force the choice of Eq. (5).

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the design of the SGCL loss in Eq. 5, which assumes that user-user and item-item similarity sums are effective contrastive terms. The temperature τ is a free parameter. No new entities are postulated.

free parameters (3)
  • temperature τ = 0.2
    Set to 0.2 after being found optimal; governs the sharpness of the contrastive distribution in Eq. 5.
  • layer weight α_k = 1/(K+1)
    Assumed uniform across layers in Eq. 4; not tuned in experiments but stated as an assumption.
  • number of layers K
    Not specified in the paper; a key architectural choice that affects graph convolution depth.
assumptions (3)
  • domain assumption The light graph convolution (Eq. 3) with symmetric normalization produces useful embeddings for recommendation.
    Adopted from LightGCN without re-derivation; assumes the propagation rule is appropriate.
  • ad hoc to paper The denominator in Eq. 5, which sums user-user and item-item similarities, acts as an effective contrastive set for learning user-item rankings.
    This is the core design choice of SGCL; the paper provides no theoretical justification for why these terms serve as proper negatives, and self-similarity handling is unspecified.
  • standard math InfoNCE-style contrastive loss (Eq. 2) is a valid self-supervised objective for graph recommendation.
    Borrowed from prior work; used as the basis for the SGCL loss.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SGCL: Unifying Self-Supervised and Supervised Learning for Graph Recommendation." pith.science (2026). https://pith.science/paper/JEBLRI6P

@misc{pith2026250713336,
  author       = {Pith},
  title        = {Pith review of: SGCL: Unifying Self-Supervised and Supervised Learning for Graph Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEBLRI6P}},
  note         = {Machine review of arXiv:2507.13336}
}
read the original abstract

Recommender systems (RecSys) are essential for online platforms, providing personalized suggestions to users within a vast sea of information. Self-supervised graph learning seeks to harness high-order collaborative filtering signals through unsupervised augmentation on the user-item bipartite graph, primarily leveraging a multi-task learning framework that includes both supervised recommendation loss and self-supervised contrastive loss. However, this separate design introduces additional graph convolution processes and creates inconsistencies in gradient directions due to disparate losses, resulting in prolonged training times and sub-optimal performance. In this study, we introduce a unified framework of Supervised Graph Contrastive Learning for recommendation (SGCL) to address these issues. SGCL uniquely combines the training of recommendation and unsupervised contrastive losses into a cohesive supervised contrastive learning loss, aligning both tasks within a single optimization direction for exceptionally fast training. Extensive experiments on three real-world datasets show that SGCL outperforms state-of-the-art methods, achieving superior accuracy and efficiency.

Figures

Figures reproduced from arXiv: 2507.13336 by the authors.

Figure 1
Figure 1. The left figure illustrates the pipeline of traditional [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. This approach necessitates the redundant graph encoders [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Performance of NDCG in the first 30 epochs. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 19 canonical work pages

  1. [1]

    Xuheng Cai, Chao Huang, Lianghao Xia, and Xubin Ren. 2023. LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation. In The Eleventh International Conference on Learning Representations

    2023

  2. [2]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Interna- tional conference on machine learning . PMLR, 1597–1607

    2020

  3. [3]

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin

  4. [4]

    Kaveh Hassani and Amir Hosein Khasahmadi. 2020. Contrastive multi-view representation learning on graphs. InInternational conference on machine learning. PMLR, 4116–4126

    work page 2020

  5. [5]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval . 639–648

    2020

  6. [6]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web . 173–182

    2017

  7. [7]

    Hyunwoo Hwangbo, Yang Sok Kim, and Kyung Jin Cha. 2018. Recommendation system development for fashion retail e-commerce.Electronic Commerce Research and Applications 28 (2018), 94–101

    work page 2018

  8. [8]

    Mohsen Jamali and Martin Ester. 2010. A matrix factorization technique with trust propagation for recommendation in social networks. In Proceedings of the fourth ACM conference on Recommender systems . 135–142

    work page 2010

Show all 39 references
  1. [9]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization tech- niques for recommender systems. Computer 42, 8 (2009), 30–37

    2009

  2. [10]

    Zihan Lin, Changxin Tian, Yupeng Hou, and Wayne Xin Zhao. 2022. Improving graph collaborative filtering with neighborhood-enriched contrastive learning. In Proceedings of the ACM Web Conference 2022 . 2320–2329

    2022

  3. [11]

    Qi Liu, Zhilong Zhou, Gangwei Jiang, Tiezheng Ge, and Defu Lian. 2023. Deep Task-specific Bottom Representation Network for Multi-Task Recommendation. arXiv preprint arXiv:2308.05996 (2023)

    2023 arXiv

  4. [12]

    Jing Ma, Liangwei Yang, Qiong Feng, Weizhi Zhang, and Philip S Yu. 2023. Graph- based village level poverty identification. InProceedings of the ACM web conference

    2023

  5. [13]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  6. [14]

    Zhen Peng, Wenbing Huang, Minnan Luo, Qinghua Zheng, Yu Rong, Tingyang Xu, and Junzhou Huang. 2020. Graph representation learning via graphical mutual information maximization. In Proceedings of The Web Conference 2020 . 259–270

    2020

  7. [15]

    Ajit Rajwade, Anand Rangarajan, and Arunava Banerjee. 2012. Image denoising using the higher order singular value decomposition.IEEE Transactions on Pattern Analysis and Machine Intelligence 35, 4 (2012), 849–862

    2012

  8. [16]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  9. [17]

    Jiaxi Tang, Yoel Drori, Daryl Chang, Maheswaran Sathiamoorthy, Justin Gilmer, Li Wei, Xinyang Yi, Lichan Hong, and Ed H Chi. 2023. Improving Training Stability for Multitask Ranking Models in Recommender Systems. arXiv preprint arXiv:2302.09178 (2023)

    2023 arXiv

  10. [18]

    Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. 2020. What makes for good views for contrastive learning? Advances in neural information processing systems 33 (2020), 6827–6839

    2020

  11. [19]

    Jianling Wang, Raphael Louca, Diane Hu, Caitlin Cellier, James Caverlee, and Liangjie Hong. 2020. Time to Shop for Valentine’s Day: Shopping Occasions and Sequential Recommendation in E-commerce. In Proceedings of the 13th Interna- tional Conference on Web Search and Data Mini...

    2020

  12. [20]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval . 165–174

    2019

  13. [21]

    Wei Wei, Chao Huang, Lianghao Xia, and Chuxu Zhang. 2023. Multi-Modal Self-Supervised Learning for Recommendation. In Proceedings of the ACM Web Conference 2023. 790–800

    2023

  14. [22]

    Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-supervised graph learning for recommendation. In Proceed- ings of the 44th international ACM SIGIR conference on research and development in information retrieval. 726–735

    2021

  15. [23]

    Ke Xu, Weizhi Zhang, Zihe Song, Yuanjie Zhu, and Philip S Yu. 2025. Graph Neural Controlled Differential Equations For Collaborative Filtering. In Companion Proceedings of the ACM on Web Conference 2025 . 1446–1449

    2025

  16. [24]

    Ke Xu, Yuanjie Zhu, Weizhi Zhang, and S Yu Philip. 2023. Graph Neural Ordinary Differential Equations-based method for Collaborative Filtering. In 2023 IEEE International Conference on Data Mining (ICDM) . IEEE, 1445–1450

    2023

  17. [25]

    Wooseong Yang, Chen Wang, Zihe Song, Weizhi Zhang, and Philip S Yu. 2024. Item Cluster-aware Prompt Learning for Session-based Recommendation. arXiv preprint arXiv:2410.04756 (2024)

    2024 arXiv

  18. [26]

    Wooseong Yang, Weizhi Zhang, Yuqing Liu, Yuwei Han, Yu Wang, Junhyun Lee, and Philip S Yu. 2025. Cold-Start Recommendation with Knowledge-Guided Retrieval-Augmented Generation. arXiv preprint arXiv:2505.20773 (2025)

    2025

  19. [27]

    Yonghui Yang, Zhengwei Wu, Le Wu, Kun Zhang, Richang Hong, Zhiqiang Zhang, Jun Zhou, and Meng Wang. 2023. Generative-contrastive graph learning for recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieva...

    2023

  20. [28]

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 974–983

    2018

  21. [29]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations. Advances in neural information processing systems 33 (2020), 5812–5823

    2020

  22. [30]

    Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. 2022. Are graph augmentations necessary? simple graph contrastive learning for recommendation. In Proceedings of the 45th international ACM SIGIR conference on research and development in info...

    2022

  23. [31]

    Weizhi Zhang, Yuanchen Bei, Liangwei Yang, Henry Peng Zou, Peilin Zhou, Aiwei Liu, Yinghui Li, Hao Chen, Jianling Wang, Yu Wang, et al. 2025. Cold-start rec- ommendation towards the era of large language models (llms): A comprehensive survey and roadmap. arXiv preprint arXiv:2...

    2025 arXiv

  24. [32]

    Weizhi Zhang, Liangwei Yang, Yuwei Cao, Ke Xu, Yuanjie Zhu, and S Yu Philip

  25. [33]

    Weizhi Zhang, Liangwei Yang, Zihe Song, Henry Peng Zou, Ke Xu, Liancheng Fang, and Philip S Yu. 2024. Do we really need graph convolution during training? light post-training graph-ode for efficient recommendation. In Proceedings of the 33rd ACM International Conference on Inf...

    2024

  26. [34]

    Weizhi Zhang, Liangwei Yang, Wooseong Yang, Henry Peng Zou, Yuqing Liu, Ke Xu, Sourav Medya, and Philip S Yu. 2025. Llminit: A free lunch from large language models for selective initialization of recommendation. arXiv preprint arXiv:2503.01814 (2025)

    2025

  27. [35]

    Weizhi Zhang, Xinyang Zhang, Chenwei Zhang, Liangwei Yang, Jingbo Shang, Zhepei Wei, Henry Peng Zou, Zijie Huang, Zhengyang Wang, Yifan Gao, et al

  28. [2012]

    arXiv preprint arXiv:1205.2618 (2012)

    BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618 (2012)

    2012 arXiv

  29. [2019]

    In The world wide web conference

    Graph neural networks for social recommendation. In The world wide web conference. 417–426

  30. [2023]

    In 2023 IEEE International Conference on Big Data (BigData)

    Dual-Teacher Knowledge Distillation for Strict Cold-Start Recommendation. In 2023 IEEE International Conference on Big Data (BigData) . IEEE, 483–492

    2023

  31. [2025]

    arXiv preprint arXiv:2506.06254 (2025)

    Personaagent: When large language model agents meet personalization at test time. arXiv preprint arXiv:2506.06254 (2025)

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.