Pith. sign in

REVIEW 4 major objections 4 minor 55 references

SharpRec claims that the ceiling on merged LLM recommenders comes from the merge itself—averaging erases the salient weights—and that sharpness-aware training plus a nonlinear salience-recovery step lifts it.

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

2026-08-01 02:37 UTC pith:PSDZVHTC

load-bearing objection Well-executed empirical recipe in a real bottleneck, but the central PSA theorem is undermined by the paper's own math: for γ=0.98 the derived distribution is lighter-tailed than Gaussian, not heavy-tailed, and the coverage proof assumes the mass transport it is supposed to establish. the 4 major comments →

arxiv 2607.25366 v1 pith:PSDZVHTC submitted 2026-07-28 cs.IR cs.LG

Sharpness-aware Model Merging with Salience Recovery for LLM-based Cross-Domain Sequential Recommendation

Huwei Ji , Jiajie Su , Yuyuan Li , Xiaohua Feng , Chaochao Chen This is my paper
classification cs.IR cs.LG
keywords Cross-domain sequential recommendationModel mergingLarge language modelsLoRASharpness-aware minimizationParameter salienceStatistical homogenizationRecommendation scalability
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 merging domain-specific LLM recommenders by averaging their LoRA adapters in parameter space fails for two structural reasons: independently fine-tuned adapters land in sharp, geometrically conflicting minima that cause negative transfer, and the arithmetic mean acts as a low-pass filter that turns the weight distribution into a narrow Gaussian, wiping out the large-magnitude coefficients that encode domain-specific preferences. SharpRec counters the first with sharpness-aware geometric alignment (SGA), which biases each domain's fine-tuning toward flat minima so merged parameters stay in connected low-loss basins; it counters the second with preference salience activation (PSA), a post-merge nonlinear map that amplifies large weights relative to small ones to restore heavy-tailed structure. The paper claims these two modules together break the saturation that appears after merging three or four source domains, and reports consistent gains over single-domain, LLM-based, and prior merging baselines in both dual- and multi-domain experiments.

Core claim

On the paper's own terms, the central discovery is that two observed failure modes of LLM-based cross-domain sequential recommendation—negative transfer on heterogeneous domain pairs and rapid saturation as source domains accumulate—are both parameter-space artifacts of linear model merging. SGA resolves the first by converting each domain adapter's optimum into a flat minimum, with a proven bound on merging interference in terms of Hessian sharpness times domain divergence. PSA resolves the second by reparameterizing the averaged weights as sign(θ̃)|θ̃|^γ(1+αe^{-β|θ̃|}) after subtracting a tiny Gaussian noise, which the paper argues restores heavy-tailed weight distributions and reactivates

What carries the argument

Two coupled mechanisms. SGA: a sharpness-aware min-max objective min_θ max_{||ε||≤ρ} L(θ+ε) run during LoRA fine-tuning of each domain, which flattens the loss landscape so that the merged parameter vector lies in every domain's low-loss basin. PSA: an element-wise nonlinearity T(θ)=sign(θ)|θ|^γ(1+αe^{-β|θ|}) applied to the merged weights after a small Gaussian 'disentanglement' step, with γ∈(0,1) intended to re-heavy-tail the distribution and re-amplify outlier weights. Theorems 4.1 and 4.2 are the formal load-bearing results: an interference bound (sharpness × divergence) and a coverage-expansion claim.

Load-bearing premise

The argument depends on the idea that pushing parameter weights away from their averaged values lands them in regions where the model's outputs are more sensitive to those weights; with the paper's chosen γ=0.98, the transform makes the tails lighter rather than heavier, so that mass-transport premise is not satisfied as stated.

What would settle it

On a trained SharpRec checkpoint, compute the fraction of LoRA update weights beyond ±3 standard deviations before and after the PSA transform at γ=0.98. If tail mass does not increase—or if applying PSA to a merged model that is already heavy-tailed gives no NDCG gain—then the salience-reactivation mechanism is not the cause of the reported lift. Separately, estimate the Jacobian sensitivity of the backbone's LoRA parameters in the weight tails; if it is not larger there, Theorem 4.2's coverage inequality loses its justification.

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

If this is right

  • Because the interference bound is controlled by sharpness times divergence, SGA can be added to any multi-domain LoRA merging pipeline before averaging, with no change to inference or deployment.
  • If PSA's salience recovery is the active ingredient, the saturation plateau in multi-domain LLM merging is not a data ceiling but an operator ceiling, so other nonlinear or sparse merge operators deserve attention.
  • The scalability experiments imply that the marginal value of additional source domains stays positive much longer than in linear merging, changing the cost-benefit calculus for adding auxiliary domains.
  • The overlap-robustness result implies that merging-based CDSR can work in hybrid-user settings with as little as 20% user overlap, where traditional overlap-dependent CDSR methods degrade.
  • Ablations show SGA contributes the larger share of the gain in the difficult Sport→Toy transfer, while PSA's contribution grows with domain count, suggesting the two fixes address distinct failure modes.

Where Pith is reading between the lines

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

  • The paper's own optimal γ regime (0.94–0.98) yields a Generalized Error Distribution with shape p=2/γ>2, which is lighter-tailed than a Gaussian, not heavier. So the empirical gains may come from contrast enhancement or shrinkage of small weights rather than from actual tail heaviness—a testable alternative explanation the paper does not distinguish.
  • If the distributional argument is repaired, the same salience-recovery trick could apply to any setting where linear averaging of fine-tuned models washes out sparse expert knowledge—e.g., multi-task LLM adapters, continual learning checkpoints, or ensembles of specialized rank-1 updates.
  • The coverage theorem assumes tail parameters have larger Jacobian sensitivity; a direct measurement of per-weight output sensitivity on the actual backbone would let future work replace that assumption with an empirical salience map and potentially target recovery to specific layers rather than all weights.

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

4 major / 4 minor

Summary. The paper proposes SharpRec, a framework for LLM-based cross-domain sequential recommendation that combines two modules: Sharpness-aware Geometric Alignment (SGA) and Preference Salience Activation (PSA). SGA applies sharpness-aware minimization during domain-specific LoRA fine-tuning to align models in flat minima and reduce merging interference. PSA is a post-fusion nonlinear transformation applied to the merged LoRA parameters, intended to restore a heavy-tailed parameter distribution and thereby reactivate salient preference signals that are diluted by linear averaging. The paper reports strong empirical gains over baselines on dual-domain and multi-domain Amazon datasets, along with ablations and sensitivity analyses. The central theoretical support is Theorem 4.2, which claims that PSA expands the model's functional coverage by moving probability mass from low-sensitivity regions to high-sensitivity tail regions.

Significance. If the theoretical claims were sound, the paper would address a real bottleneck in LLM-based CDSR model merging: performance saturation under multi-domain fusion. The empirical study is extensive, with code released, multiple baselines, ablations, and sensitivity analyses. However, the core mathematical justification is internally inconsistent: for the chosen hyperparameters, the PSA transformation produces lighter-tailed, not heavier-tailed, distributions, and the proof of coverage expansion assumes the conclusion. Since the 'heavy-tailed' mechanism is the paper's central novelty and is invoked in the abstract, introduction, method, and experimental discussion, the contribution as written is not established. The reported experimental gains are large and suggestive, but without a valid theoretical foundation the paper cannot be accepted in its current form.

major comments (4)
  1. [Section 4.3 / Appendix C, Eq. (33)] The central claim that PSA 'restores heavy-tailed characteristics' is contradicted by the paper's own derivation. Eq. (33) gives p_PSA(y) ∝ |y|^{1/γ-1} exp(-|y|^{2/γ}/(2σ^2)), a GED with shape parameter p = 2/γ. For the reported γ = 0.98 < 1, p ≈ 2.04, so the tails decay faster than a Gaussian; this is a lighter-tailed distribution, not a heavy-tailed one. The polynomial prefactor cannot reverse an exponential tail order with p > 2. Since Section 5.5 reports optimal γ in [0.94, 0.98], the inconsistency is operative in the actual experiments. The bounded factor (1 + α e^{-β|x|}) in Eq. (12) is asymptotically constant and cannot change the tail order. Thus the mechanism does not do what the paper claims.
  2. [Appendix D / Theorem 4.2] The proof of coverage expansion assumes its own conclusion. Eq. (37) requires p_PSA - p_Gauss > 0 on W_T and < 0 on W_C; the text asserts this follows from 'mass transport' (Appendix D, first paragraph of the proof), but no independent transport theorem is proved. In fact, for γ < 1, Eq. (33) concentrates probability mass near the center more than a Gaussian does, so the sign pattern is reversed. Eq. (35), which asserts that the Jacobian determinant is larger in the tail region, is another unverified assumption. Consequently C2 > C1 is not established; the theorem is circular.
  3. [Appendices C and D] The proofs are 'adapted from' prior work [18, 41] and are not self-contained. In particular, the claim that the PSA transformation induces a heavy-tailed distribution is not actually proved in Appendix C; it is merely stated that Eq. (33) belongs to the GED family. The statement in Section 4.3 that 'The proof that this transformation induces a heavy-tailed distribution is provided in Appendix C' is therefore unsupported. Moreover, Appendix D explicitly says 'As derived in Appendix C, the PSA transformation functions as a mass transport mechanism' -- a circular step, since mass transport is exactly what needs to be established to prove coverage expansion.
  4. [Section 5.1.4 / Section 5.5] The hyperparameters γ, ρ, σ_g, α, β, and λ_k are selected based on validation performance on the same datasets used for evaluation. Because the theoretical claim is supposed to explain why these particular values (e.g., γ = 0.98) are effective, the selection procedure does not provide independent evidence for the mechanism. No held-out/dataset-level verification is provided, so the reported 'lifting of the performance ceiling' may reflect tuning to the evaluation settings rather than the proposed salience-recovery principle.
minor comments (4)
  1. [Figure 6] The x-axis labels for γ are inconsistent: subfigures (c) and (d) use '0.84 0.88 0.9 0.92 0.94 0.96 0.98 1.0' while the text describes range [0.94, 0.98]; also the tick spacing is irregular. Please harmonize.
  2. [Section 5.2 / Table 2] The reported improvements over the strongest baselines are very large (e.g., HR@3 80.13 vs. 62.34 on Book→Movie). No statistical significance tests (e.g., paired tests) are reported, despite repeated runs with standard deviations. Please add significance testing or at least confidence intervals for the main comparisons.
  3. [Section 5.1.2 / Prompt template] The prompt template in Figure 3 is described only for SharpRec; it is unclear whether baselines like LLM4CDSR and WeaveRec use the same prompt and candidate-pool setup. A brief statement on prompt/candidate uniformity would strengthen the fairness of the comparison.
  4. [References] Several references are to 2026 publications, which may be appropriate for a 2026 KDD venue, but some cited works seem unrelated to the content (e.g., [4], [5], [7], [10], [35], [36]). Please verify that each reference is cited in the correct context.

Circularity Check

1 steps flagged

Theorem 4.2's coverage proof assumes the mass-transport property it sets out to prove; Appendix C's own density contradicts the heavy-tail premise at γ=0.98.

specific steps
  1. self definitional [Appendix D (Proof of Theorem 4.2), Eqs. (35)-(37); Sec. 4.3 Theorem 4.2]
    "the PSA transformation functions as a mass transport mechanism: it shifts probability mass from the low-sensitivity center W_C to the high-sensitivity tail W_T. ... because the weighting factor |det(J_Φ)| is strictly larger in the tail region (W_T) than in the central region (W_C), the positive contribution from the first integral strictly dominates the negative contribution from the second."

    Theorem 4.2 asserts C2 > C1 on the grounds that PSA 'shifts probability mass' from the low-sensitivity center to the high-sensitivity tail and that |det J| is 'strictly larger' there (Eq. 35). These two premises are precisely the content of the conclusion: if mass moves to regions with larger Jacobian weight, the coverage integral increases by construction. No independent argument establishes the mass-transfer premise, and Appendix C's own Eq. (33), with the deployed γ=0.98, has GED shape p=2/γ≈2.04>2, i.e., lighter-than-Gaussian tails, so the assumed shift to W_T is false for the hyperparameters actually used (Sec. 5.1.4). ΔC>0 is therefore assumed, not derived.

full rationale

The only load-bearing circular step is the theoretical justification of PSA. Appendix D proves Theorem 4.2 by assuming that PSA transports mass from the low-sensitivity center to the high-sensitivity tail and that the Jacobian determinant is larger in the tail; these assumptions entail the coverage-expansion conclusion immediately, making the theorem true by construction rather than by derivation from the PSA transform's actual statistics. Moreover, the paper's own change-of-variables result, Eq. (33), yields a GED with shape 2/γ, and with the reported γ=0.98 this shape is 2.04 > 2, so the 'heavy-tailed' premise is internally contradicted. The SGA analysis is adapted from external prior work ([18]) and is a standard Taylor bound, not self-citation; the experimental comparisons against baselines are independent and not fitted to the theory. The circularity is confined to the central theoretical claim that PSA 'lifts the performance upper bound,' but because that claim is a headline contribution, the overall circularity score is elevated to 7. The empirical results could still be valid, but the claimed first-principles guarantee is not.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The free parameters (γ, ρ, σ_g, α, β) are fitted to validation performance on the same datasets used for evaluation; the theoretical mechanism depends on γ>1 while the implementation uses γ=0.98, and the proof of Theorem 4.2 relies on an asserted Jacobian property and a false mass-transport assumption. No new physical entities are introduced.

free parameters (5)
  • γ (salience activation factor) = 0.98
    Controls the exponent in Eq. (12); tuned via sensitivity analysis in Fig. 6; the 'heavy-tailed' claim requires γ>1, but the paper sets γ<1.
  • ρ (perturbation radius) = 0.01
    SAM perturbation radius in Eq. (8), tuned manually; sensitivity shown in Fig. 6a-b.
  • σ_g (noise variance) = 0.0001
    Gaussian noise injected in Eq. (11); tuned manually; sensitivity shown in Fig. 6e-f.
  • α, β (PSA smoothness coefficients) = α=0.1, β=10
    Shape the (1+α e^{-β|x|}) factor in Eq. (12); tuned manually in Section 5.1.4.
  • merging coefficients λ_k = not specified
    Eq. (5) weights per-domain adapters; the paper does not report how λ_k is set in the experiments, but it is a free choice affecting all results.
axioms (4)
  • standard math The loss L is twice differentiable and the Hessian spectral norm σ(θ) is finite (used for Theorem 4.1).
    Required for the Taylor-expansion bound in Appendix B; standard for smooth LLM losses but not explicitly verified on the actual loss landscape.
  • ad hoc to paper |det(J_Φ(w))| is strictly larger in the tail region W_T than in the central region W_C (Eq. 35).
    Empirical assertion with no evidence; it is a premise of Theorem 4.2 and effectively assumes the conclusion that tail parameters contribute more to functional coverage.
  • ad hoc to paper PSA transports probability mass from the low-sensitivity center to the high-sensitivity tail (Appendix D).
    False for γ<1: the transformation with γ=0.98 compresses large |x| (since |x|^γ < |x| for |x|>1) and expands small |x|, moving tail probability toward the center. The 'mass transport' assumption is contradicted by the paper's own Eq. (33).
  • domain assumption LoRA adapters trained with SAM lie in connected low-loss basins so that linear interpolation stays in low-loss region.
    Borrowed from [18]; not verified on CDSR data beyond indirect performance metrics. The paper assumes flat minima transfer to the CDSR setting without direct evidence of basin connectivity.

pith-pipeline@v1.3.0-alltime-deepseek · 23229 in / 16050 out tokens · 152249 ms · 2026-08-01T02:37:40.640687+00:00 · methodology

0 comments
read the original abstract

LLM-based Cross-Domain Sequential Recommendation (CDSR) leverages LLMs to enhance target performance via deep semantic reasoning, alleviating the dependency on overlapping users. Among LLM-based paradigms, model merging is particularly promising for multi-domain scenarios due to its superior scalability and flexibility in integrating diverse knowledge sources. However, our empirical investigations reveal two critical bottlenecks: (1) cross-domain knowledge conflict; and (2) performance saturation in multi-domain fusion. Our analysis attributes these phenomena to parameter-level misalignment and statistical homogenization during the merging process. To address these bottlenecks, we propose SharpRec, Sharpness-aware Model Merging with Salience Recovery for LLM-based CDSR, a framework designed to lift the performance upper bound of merged models. SharpRec incorporates two synergistic modules: Sharpness-aware Geometric Alignment to establish a stable geometric foundation for interference-free fusion; and Preference Salience Activation to effectively recover the distinctive features essential for bolstering target domain performance. Extensive experiments in both dual-domain and multi-domain scenarios demonstrate that SharpRec consistently outperforms state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2607.25366 by Chaochao Chen, Huwei Ji, Jiajie Su, Xiaohua Feng, Yuyuan Li.

Figure 1
Figure 1. Figure 1: Empirical evidence of negative transfer and scalability bottlenecks in CDSR. (a) & (b) show that merging heterogeneous [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of parameter geometric compatibility. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The prompt template for our SharpRec framework, illustrating the integration of historical sequences and can￾didate pools within a conversational LLM structure [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance comparison w.r.t. the number of [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Impact of hyperparameters 𝜌, 𝛾, and 𝜎𝑔 on recom￾mendation performance (HR@5 and NDCG@5). once 𝜎𝑔 surpasses a critical threshold (e.g., 𝜎𝑔 > 0.01), the exces￾sive variance destabilizes the feature alignment between domains, resulting in a consistent drop in accuracy. 6 Conclusion In this work, we explore the potential of model merging for CDSR. Our empirical studies identify two critical bottlenecks in exis… view at source ↗
Figure 5
Figure 5. Figure 5: Impact of user overlap ratios (ranging from 20% to [PITH_FULL_IMAGE:figures/full_fig_p009_5.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

55 extracted references · 11 linked inside Pith

  1. [1]

    Jiangxia Cao, Xin Cong, Jiawei Sheng, Tingwen Liu, and Bin Wang. 2022. Con- trastive cross-domain sequential recommendation. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 138–147

    2022

  2. [2]

    Shu Chen, Zitao Xu, Weike Pan, Qiang Yang, and Zhong Ming. 2024. A survey on cross-domain sequential recommendation.arXiv preprint arXiv:2401.04971 (2024)

    Pith/arXiv arXiv 2024

  3. [3]

    Zeyu Cui, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. M6-rec: Generative pretrained language models are open-ended recommender systems.arXiv preprint arXiv:2205.08084(2022)

    Pith/arXiv arXiv 2022

  4. [4]

    Jiabao Dong, Lingyuan Yang, Pengji Fang, Shixiang Li, Yusheng Kong, and Lei Ren. 2026. IEI-TIA: Industrial Embodied Intelligence Trustworthy Interpretable Agent for Robotic Long-Horizon and Repetitive Tasks.IEEE Transactions on Automation Science and Engineering23 (2026), 9211–9222

    2026

  5. [5]

    Tianyu Du, Shouling Ji, Lujia Shen, Yao Zhang, Jinfeng Li, Jie Shi, Chengfang Fang, Jianwei Yin, Raheem Beyah, and Ting Wang. 2021. Cert-rnn: Towards certifying the robustness of recurrent neural networks. InProceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security

    2021

  6. [6]

    Wei Feng and Zongyuan Ge. 2025. Generalized Category Discovery under Do- main Shift: A Frequency Domain Perspective.Advances in Neural Information Processing Systems38 (2025), 111721–111749

    2025

  7. [7]

    Yuqian Fu, Runze Wang, Bin Ren, Guolei Sun, Biao Gong, Yanwei Fu, Danda Pani Paudel, Xuanjing Huang, and Luc Van Gool. 2025. Objectrelator: Enabling cross- view object relation understanding across ego-centric and exo-centric perspec- tives. InProceedings of the IEEE/CVF International Conference on Computer Vision. 6530–6540

    2025

  8. [8]

    Jingtong Gao, Xiangyu Zhao, Bo Chen, Fan Yan, Huifeng Guo, and Ruiming Tang. 2023. AutoTransfer: Instance transfer for cross-domain recommendations. InProceedings of the 46th international ACM SIGIR conference on research and development in information retrieval. 1478–1487

    2023

  9. [9]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). InProceedings of the 16th ACM conference on recommender systems. 299–315

    2022

  10. [10]

    Biao Gong, Chenggang Yan, Junjie Bai, Changqing Zou, and Yue Gao. 2020. Ham- ming embedding sensitivity guided fusion network for 3D shape representation. IEEE Transactions on Image Processing29 (2020), 8381–8390

    2020

  11. [11]

    Lei Guo, Li Tang, Tong Chen, Lei Zhu, Quoc Viet Hung Nguyen, and Hongzhi Yin. 2021. DA-GCN: A domain-aware attentive graph convolution network for shared-account cross-domain sequential recommendation.arXiv preprint arXiv:2105.03300(2021)

    Pith/arXiv arXiv 2021

  12. [12]

    Guy Hadad, Haggai Roitman, Yotam Eshel, Bracha Shapira, and Lior Rokach. 2025. X-Cross: Dynamic Integration of Language Models for Cross-Domain Sequential Recommendation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1497–1507

    2025

  13. [13]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  14. [14]

    Min Hou, Xin Liu, Le Wu, Chenyi He, Hao Liu, Zhi Li, Xin Li, and Si Wei

  15. [15]

    Ruijie Hou, Zhaoyang Yang, Yu Ming, Hongyu Lu, Zhuobin Zheng, Yu Chen, Qinsong Zeng, and Ming Chen. 2024. Cross-Domain LifeLong Sequential Model- ing for Online Click-Through Rate Prediction. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5116–5125

    2024

  16. [16]

    Siteng Huang, Biao Gong, Yutong Feng, Min Zhang, Yiliang Lv, and Donglin Wang. 2024. Troika: Multi-path cross-modal traction for compositional zero- shot learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 24005–24014

    2024

  17. [17]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

    2018

  18. [18]

    Yeoreum Lee, Jinwook Jung, and Sungyong Baik. 2025. Mitigating parameter interference in model merging via sharpness-aware fine-tuning.arXiv preprint arXiv:2504.14662(2025)

    Pith/arXiv arXiv 2025

  19. [19]

    Chengxi Li, Yejing Wang, Qidong Liu, Xiangyu Zhao, Wanyu Wang, Yiqi Wang, Lixin Zou, Wenqi Fan, and Qing Li. 2023. STRec: Sparse transformer for sequential recommendations. InProceedings of the 17th ACM conference on recommender systems. 101–111

    2023

  20. [20]

    Haipeng Li, Jiangxia Cao, Yiwen Gao, Yunhuai Liu, and Shuchao Pang. 2024. Cross-Domain Sequential Recommendation via Neural Process.arXiv preprint arXiv:2410.13588(2024)

    Pith/arXiv arXiv 2024

  21. [21]

    Pan Li, Zhichao Jiang, Maofei Que, Yao Hu, and Alexander Tuzhilin. 2021. Dual attentive sequential learning for cross-domain click-through rate prediction. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 3172–3180

    2021

  22. [22]

    Zixu Li, Yupeng Hu, Zhiwei Chen, Shiqi Zhang, Qinlei Huang, Zhiheng Fu, and Yinwei Wei. 2026. HABIT: Chrono-Synergia Robust Progressive Learning Framework for Composed Image Retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 6762–6770

    2026

  23. [23]

    Zixu Li, Yupeng Hu, Zhiheng Fu, Zhiwei Chen, Yongqi Li, and Liqiang Nie. 2026. TEMA: Anchor the Image, Follow the Text for Multi-Modification Composed Image Retrieval. arXiv:2604.21806 [cs.CV]

    Pith/arXiv arXiv 2026

  24. [24]

    Xinyu Lin, Wenjie Wang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua

  25. [25]

    Qidong Liu, Fan Yan, Xiangyu Zhao, Zhaocheng Du, Huifeng Guo, Ruiming Tang, and Feng Tian. 2023. Diffusion augmentation for sequential recommenda- tion. InProceedings of the 32nd ACM International conference on information and knowledge management. 1576–1586

    2023

  26. [26]

    Qidong Liu, Xiangyu Zhao, Yejing Wang, Zijian Zhang, Howard Zhong, Chong Chen, Xiang Li, Wei Huang, and Feng Tian. 2025. Bridge the domains: Large language models enhanced cross-domain sequential recommendation. InProceed- ings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1582–1592

    2025

  27. [27]

    Xinyi Liu, Ruijie Wang, Dachun Sun, Dilek Hakkani Tur, and Tarek Abdelzaher

  28. [28]

    Haokai Ma, Ruobing Xie, Lei Meng, Xin Chen, Xu Zhang, Leyu Lin, and Jie Zhou. 2024. Triple sequence learning for cross-domain recommendation.ACM Transactions on Information Systems42, 4 (2024), 1–29

    2024

  29. [29]

    Muyang Ma, Pengjie Ren, Zhumin Chen, Zhaochun Ren, Lifan Zhao, Peiyu Liu, Jun Ma, and Maarten de Rijke. 2022. Mixed information flow for cross-domain sequential recommendations.ACM Transactions on Knowledge Discovery from Data (TKDD)16, 4 (2022), 1–32

    2022

  30. [30]

    2019.𝜋-net: A parallel information-sharing network for shared-account cross- domain sequential recommendations

    Muyang Ma, Pengjie Ren, Yujie Lin, Zhumin Chen, Jun Ma, and Maarten de Rijke. 2019.𝜋-net: A parallel information-sharing network for shared-account cross- domain sequential recommendations. InProceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval. 685– 694

    2019

  31. [31]

    InCompanion Proceedings of the ACM on Web Conference 2025

    Uncovering cross-domain recommendation ability of large language models. InCompanion Proceedings of the ACM on Web Conference 2025. 2736–2743

    2025

  32. [32]

    Bo Peng, Xinyi Ling, Ziru Chen, Huan Sun, and Xia Ning. 2024. ecellm: Gen- eralizing large language models for e-commerce from large-scale, high-quality instruction data.arXiv preprint arXiv:2402.08831(2024)

    Pith/arXiv arXiv 2024

  33. [33]

    Tingjia Shen, Hao Wang, Jiaqing Zhang, Sirui Zhao, Liangyue Li, Zulong Chen, Defu Lian, and Enhong Chen. 2024. Exploring user retrieval integration towards large language models for cross-domain sequential recommendation.arXiv preprint arXiv:2406.03085(2024)

    Pith/arXiv arXiv 2024

  34. [34]

    Wenchao Sun, Muyang Ma, Pengjie Ren, Yujie Lin, Zhumin Chen, Zhaochun Ren, Jun Ma, and Maarten De Rijke. 2021. Parallel split-join networks for shared account cross-domain sequential recommendations.IEEE Transactions on Knowl- edge and Data Engineering35, 4 (2021), 4106–4123

    2021

  35. [35]

    Chung Park, Taesan Kim, Hyungjun Yoon, Junui Hong, Yelim Yu, Mincheol Cho, Minsung Choi, and Jaegul Choo. 2024. Pacer and runner: Cooperative learning framework between single-and cross-domain sequential recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2071–2080

    2024

  36. [36]

    Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, DanDan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. 2025. Animate-X: Universal Character Image Animation with Enhanced Motion Representation. In The Thirteenth International Conference on Learning Representations

    2025

  37. [37]

    Songbai Tan, Xuerui Qiu, Yao Shu, Gang Xu, Linrui Xu, Xiangyu Xu, Huiping Zhuang, Ming Li, and Fei Yu. 2025. WMarkGPT: Watermarked Image Under- standing via Multimodal Large Language Models. InProceedings of the 42nd International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 267). PMLR, 58621–58636

    2025

  38. [38]

    Zuoli Tang, Zhaoxin Huan, Zihao Li, Xiaolu Zhang, Jun Hu, Chilin Fu, Jun Zhou, Lixin Zou, and Chenliang Li. 2025. One model for all: Large language models are domain-agnostic recommendation systems.ACM Transactions on Information Systems43, 5 (2025), 1–27

    2025

  39. [39]

    Shuai Tan, Biao Gong, Yutong Feng, Kecheng Zheng, Dandan Zheng, Shuwei Shi, Yujun Shen, Jingdong Chen, and Ming Yang. 2025. Mimir: Improving video diffusion models for precise text understanding. InProceedings of the Computer Vision and Pattern Recognition Conference. 23978–23988

    2025

  40. [40]

    Yuhao Wang, Xiangyu Zhao, Bo Chen, Qidong Liu, Huifeng Guo, Huanshuo Liu, Yichao Wang, Rui Zhang, and Ruiming Tang. 2023. PLATE: A prompt-enhanced paradigm for multi-scenario recommendations. InProceedings of the 46th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval. 1498–1507

    2023

  41. [41]

    Zijing Wang, Xingle Xu, Yongkang Liu, Yiqun Zhang, Peiqin Lin, Shi Feng, Xiaocui Yang, Daling Wang, and Hinrich Schütze. 2025. Why do more experts fail? a theoretical analysis of model merging.arXiv preprint arXiv:2505.21226(2025). Sharpness-aware Model Merging with Salience Recovery for LLM-based Cross-Domain Sequential Recommendation KDD ’26, August 09–...

    Pith/arXiv arXiv 2025

  42. [42]

    Haoran Xin, Ying Sun, Chao Wang, and Hui Xiong. 2025. Llmcdsr: Enhancing cross-domain sequential recommendation with large language models.ACM Transactions on Information Systems(2025)

    2025

  43. [43]

    Shuliang Wang, Jiabao Zhu, Kaibo Wang, and Sijie Ruan. 2025. LeCDSR: Large language model enhanced cross-domain sequential recommendation.Information Fusion(2025), 103762

    2025

  44. [44]

    Wujiang Xu, Xuying Ning, Wenfang Lin, Mingming Ha, Qiongxu Ma, Qianqiao Liang, Xuewen Tao, Linxun Chen, Bing Han, and Minnan Luo. 2024. Towards open-world cross-domain sequential recommendation: A model-agnostic con- trastive denoising approach. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 161–179

    2024

  45. [45]

    Wujiang Xu, Qitian Wu, Runzhong Wang, Mingming Ha, Qiongxu Ma, Linxun Chen, Bing Han, and Junchi Yan. 2024. Rethinking cross-domain sequential recommendation under open-world assumptions. InProceedings of the ACM Web Conference 2024. 3173–3184

    2024

  46. [46]

    Zitao Xu, Shu Chen, Weike Pan, and Zhong Ming. 2025. A multi-view graph contrastive learning framework for cross-domain sequential recommendation. ACM Transactions on Recommender Systems3, 4 (2025), 1–28

    2025

  47. [47]

    Naen Xu, Jinghuai Zhang, Changjiang Li, Zhi Chen, Chunyi Zhou, Qingming Li, Tianyu Du, and Shouling Ji. 2025. VideoEraser: Concept Erasure in Text-to-Video Diffusion Models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 5954–5983

    2025

  48. [48]

    Tianzi Zang, Yanmin Zhu, Ruohan Zhang, Chunyang Wang, Ke Wang, and Jiadi Yu. 2023. Contrastive multi-view interest learning for cross-domain sequential recommendation.ACM Transactions on Information Systems42, 3 (2023), 1–30

    2023

  49. [49]

    Zijian Zhang, Shuchang Liu, Jiaao Yu, Qingpeng Cai, Xiangyu Zhao, Chunxu Zhang, Ziru Liu, Qidong Liu, Hongwei Zhao, Lantao Hu, et al . 2024. M3oe: Multi-domain multi-task mixture-of experts recommendation framework. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 893–902

    2024

  50. [50]

    Jiamu Zheng, Jinghuai Zhang, Tianyu Du, Xuhong Zhang, Jianwei Yin, and Tao Lin. 2025. CollabEdit: Towards Non-destructive Collaborative Knowledge Editing. InInternational Conference on Learning Representations, Vol. 2025. 29154–29174. A Derivation of the SGA Objective Adapted from the proof of Eq. (6) in [ 18] (Appendix B), reformulated for the cross-doma...

    2025

  51. [51]

    Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. [n. d.]. Model merging in llms, mllms, and beyond: Methods, theo- ries, applications and opportunities, 2024.URL https://arxiv. org/abs/2408.07666 2408, 3 ([n. d.])

    Pith/arXiv arXiv 2024

  52. [55]

    B Proof of Theorem 4.1 Following the Taylor-expansion bounding technique of [ 18] (Theo- rem 1, Appendix D), reformulated for the CDSR setting

    is mathematically equivalent to seeking a flat minimum via Sharpness-Aware Minimization. B Proof of Theorem 4.1 Following the Taylor-expansion bounding technique of [ 18] (Theo- rem 1, Appendix D), reformulated for the CDSR setting. Setup and Definitions.We aim to bound the interference error 𝛿: 𝛿=L(𝜆𝜃 𝐴+(1−𝜆)𝜃 𝐵)− [𝜆L(𝜃𝐴)+(1−𝜆)L(𝜃 𝐵)].(18) Let𝜃𝑚𝑒𝑟𝑔𝑒 =𝜆𝜃 ...

    2026

  53. [2015]

    Session-based recommendations with recurrent neural networks.arXiv preprint arXiv:1511.06939(2015)

    Pith/arXiv arXiv 2015

  54. [2024]

    InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Bridging items and language: A transition paradigm for large language model-based recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1816–1826

  55. [2025]

    WeaveRec: An LLM-Based Cross-Domain Sequential Recommendation Framework with Model Merging.arXiv preprint arXiv:2510.26546(2025)

    arXiv 2025