arxiv: 2604.26760 · v1 · submitted 2026-04-29 · 💻 cs.IR

Recognition: unknown

Factorized Latent Reasoning for LLM-based Recommendation

Tianqi Gao , Chengkai Huang , Zihan Wang , Cao Liu , Ke Zeng , Lina Yao

Authors on Pith no claims yet

Pith reviewed 2026-05-07 11:13 UTC · model grok-4.3

classification 💻 cs.IR

keywords LLM-based recommendationfactorized latent reasoningsequential recommendationmulti-factor attentionuser preference modelingdisentangled representationsreinforcement learning alignment

0 comments

The pith

Decomposing user preferences into multiple latent factors improves LLM-based sequential recommendations over single-vector methods.

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

The paper argues that current LLM approaches to recommendation compress user intent into one latent vector, which cannot reflect the varied aspects of what people like across different interactions. It introduces Factorized Latent Reasoning to split the process into several specialized factors, each refined by attention on distinct parts of the history. Orthogonality, diversity, and sparsity regularizations keep the factors from collapsing together, while dynamic aggregation and group-relative policy optimization handle the final output and alignment. Experiments across benchmarks show gains in accuracy plus better robustness and the ability to inspect which factors drove a result. Readers should care because recommendation engines shape everyday online experiences, and handling nuanced preferences more explicitly could raise satisfaction without major added cost.

Core claim

Factorized Latent Reasoning decomposes the latent reasoning process into multiple disentangled preference factors via a lightweight multi-factor attention module that iteratively refines a shared thought representation, applies orthogonality, attention diversity, and sparsity regularizations to encourage specialization, dynamically aggregates the factors for prediction, and integrates group-relative policy optimization for direct latent-space alignment, yielding consistent gains over single-vector baselines in accuracy, robustness, and interpretability.

What carries the argument

Lightweight multi-factor attention module that iteratively refines a latent thought representation by letting each factor attend to distinct aspects of the user's interaction history.

If this is right

FLR produces higher recommendation accuracy than strong single-vector latent baselines across multiple datasets.
The model becomes more robust to changes in user interaction patterns.
Individual factors can be examined to explain why a particular item is recommended.
Reinforcement learning alignment occurs stably inside the latent space without separate fine-tuning stages.

Where Pith is reading between the lines

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

The same factorization idea could be tested in other LLM tasks that involve layered user goals, such as conversational agents or personalized content generation.
Varying the number of factors on different datasets might reveal how preference complexity differs by domain or user group.
Because the module is lightweight, the approach may combine easily with larger base LLMs without proportional slowdown.

Load-bearing premise

User preferences consist of multiple independent aspects that can be separated and specialized through attention and regularization rather than represented as one combined vector.

What would settle it

A side-by-side test on the same benchmarks where a single latent vector model, given equivalent compute, matches or exceeds FLR accuracy, robustness, and interpretability would show the factorization step is not required.

Figures

Figures reproduced from arXiv: 2604.26760 by Cao Liu, Chengkai Huang, Ke Zeng, Lina Yao, Tianqi Gao, Zihan Wang.

**Figure 1.** Figure 1: Schematic comparison of three reasoning paradigms: (a) Explicit CoT Reasoning, (b) Latent Reasoning, and (c) our view at source ↗

**Figure 2.** Figure 2: Illustration of the FLR architecture and its Two-Stage training paradigm. view at source ↗

**Figure 3.** Figure 3: Illustration of the Factorized Latent Reasoning view at source ↗

**Figure 4.** Figure 4: Visualization of factor disentanglement on Toys and view at source ↗

**Figure 5.** Figure 5: Relative performance improvement of LR-GRPO view at source ↗

**Figure 6.** Figure 6: Visualization of attention patterns on Amazon Video Games. The x-axis shows recommended items, and the y-axis view at source ↗

**Figure 7.** Figure 7: Performance improvement of FLR over LatentR view at source ↗

**Figure 8.** Figure 8: Performance of FLR w.r.t. the number of latent view at source ↗

read the original abstract

Large language models (LLMs) have recently been adopted for recommendation by framing user preference modeling as a language generation problem. However, existing latent reasoning approaches typically represent user intent with a single latent vector, which struggles to capture the inherently multi-faceted nature of user preferences. We propose Factorized Latent Reasoning (FLR), a novel framework for LLM-based sequential recommendation that decomposes latent reasoning into multiple disentangled preference factors. FLR introduces a lightweight multi-factor attention module that iteratively refines a latent thought representation, where each factor attends to distinct aspects of the user's interaction history. To encourage diversity and specialization, we design orthogonality, attention diversity, and sparsity regularization objectives, and dynamically aggregate factor contributions for the final prediction. We further integrate FLR with an efficient reinforcement learning strategy based on group-relative policy optimization, enabling stable alignment directly in the latent reasoning space. Experiments on multiple benchmarks show that FLR consistently outperforms strong baselines while improving robustness and interpretability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

FLR offers a factorized latent approach for multi-faceted preferences in LLM recs but falls short on proving the factors are disentangled.

read the letter

The main takeaway is that this paper proposes Factorized Latent Reasoning (FLR) to handle multi-faceted user preferences in LLM-based recommendation by decomposing the latent space into multiple factors. The approach is a practical extension of existing latent reasoning methods, but it does not provide direct evidence that the factors end up independent. What is new is the combination of a multi-factor attention module that iteratively refines latent representations with each factor focusing on different parts of the interaction history, plus the three regularization terms for orthogonality, attention diversity, and sparsity. They also add an efficient reinforcement learning step using group-relative policy optimization directly in the latent space. The experiments report consistent improvements over strong baselines on several benchmarks, with claims of better robustness and interpretability. The paper does a decent job laying out a lightweight way to add factorization without major changes to the LLM pipeline. The RL integration looks efficient and could be a reusable piece for alignment. The soft spots center on the disentanglement claim. The stress-test note hits the mark: there are no quantitative measures shown for how independent the factors are, such as correlations between them or attention overlap statistics. Without those, it is hard to know if the regularizers are doing the work or if the gains are simply from extra parameters and the policy optimization. The abstract is short on math, so the full paper needs to include the exact formulations and ablation studies to support the interpretability angle. If the factors are not truly specialized, the robustness benefit stays unproven. This work is aimed at the community working on sequential recommendation using large language models. A reader there could pick up the architecture and try the regularizers in their own setups. It deserves peer review because it presents a clear, testable idea with empirical results that could move the subfield forward, provided the missing validation is addressed in revisions. I recommend sending it to referees with a note to check the factor independence metrics.

Referee Report

3 major / 2 minor

Summary. The paper proposes Factorized Latent Reasoning (FLR), a framework for LLM-based sequential recommendation that decomposes user preference modeling from a single latent vector into multiple disentangled factors. It introduces a multi-factor attention module that iteratively refines latent thought representations with each factor attending to distinct aspects of interaction history. Orthogonality, attention diversity, and sparsity regularization objectives are added to promote specialization, factors are dynamically aggregated for prediction, and the approach is combined with group-relative policy optimization for stable RL alignment in latent space. Experiments on multiple benchmarks are reported to show consistent outperformance over strong baselines along with gains in robustness and interpretability.

Significance. If the core premise holds, the work would offer a meaningful advance in LLM-based recommendation by addressing the multi-faceted nature of user preferences through explicit factorization rather than monolithic latent representations. The combination of attention-based factorization with targeted regularizations and latent-space RL is a constructive direction that could improve both performance and downstream interpretability in sequential recommendation tasks.

major comments (3)

[Method (regularization objectives) and Experiments] The central claim that the multi-factor attention module plus orthogonality/attention-diversity/sparsity objectives produce genuinely disentangled preference factors (and that these factors drive the reported gains) is load-bearing but unsupported by direct evidence. No quantitative diagnostics—such as pairwise factor correlations, mutual information between factors, or attention overlap statistics—are provided to show that the factors are less redundant than those in a standard multi-head attention baseline. Without such checks, the robustness and interpretability benefits cannot be attributed to disentanglement rather than increased capacity or the RL component.
[Experiments] The experimental section reports consistent outperformance on multiple benchmarks but supplies no error bars, statistical significance tests, or ablation studies isolating the contribution of the factorization regularizations versus the added parameters or the group-relative policy optimization. This makes it impossible to verify whether the headline gains stem from the claimed disentanglement mechanism.
[§3 (FLR framework)] The abstract and method description state that the regularizations encourage diversity and specialization, yet the paper does not demonstrate that these objectives actually achieve the intended effect beyond the loss terms themselves (e.g., no post-training analysis of factor independence or specialization on held-out data).

minor comments (2)

[Method] Notation for the multi-factor attention module and the dynamic aggregation step could be clarified with an explicit equation or diagram showing how factor contributions are combined for the final token prediction.
[Introduction] The paper would benefit from a short related-work paragraph explicitly contrasting FLR with prior multi-interest or disentangled recommendation models (e.g., those using capsule networks or variational approaches) to highlight the novelty of the LLM-latent-reasoning integration.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback and positive evaluation of FLR's potential contribution. We agree that stronger quantitative evidence is needed to support the disentanglement claims and will revise the manuscript accordingly by adding the requested diagnostics, statistical reporting, and analyses.

read point-by-point responses

Referee: [Method (regularization objectives) and Experiments] The central claim that the multi-factor attention module plus orthogonality/attention-diversity/sparsity objectives produce genuinely disentangled preference factors (and that these factors drive the reported gains) is load-bearing but unsupported by direct evidence. No quantitative diagnostics—such as pairwise factor correlations, mutual information between factors, or attention overlap statistics—are provided to show that the factors are less redundant than those in a standard multi-head attention baseline. Without such checks, the robustness and interpretability benefits cannot be attributed to disentanglement rather than increased capacity or the RL component.

Authors: We agree that direct quantitative diagnostics are necessary to substantiate the disentanglement claims and to distinguish the contribution of the proposed regularizations from capacity or RL effects. In the revised manuscript we will add: (1) average pairwise cosine similarities and correlation matrices between the learned factor representations, (2) variational estimates of mutual information between factors, and (3) attention overlap statistics (e.g., Jaccard index over top-k attended items) computed against a multi-head attention ablation without the orthogonality/attention-diversity/sparsity terms. These metrics will be reported on the benchmark datasets to demonstrate reduced redundancy. revision: yes
Referee: [Experiments] The experimental section reports consistent outperformance on multiple benchmarks but supplies no error bars, statistical significance tests, or ablation studies isolating the contribution of the factorization regularizations versus the added parameters or the group-relative policy optimization. This makes it impossible to verify whether the headline gains stem from the claimed disentanglement mechanism.

Authors: We acknowledge the absence of error bars, significance testing, and targeted ablations. In the revision we will: (1) report means and standard deviations over at least five random seeds with error bars in all tables, (2) include paired statistical tests (t-test or Wilcoxon signed-rank) to establish significance of improvements, and (3) add ablation studies that separately remove the multi-factor attention module, each regularization objective, and the group-relative policy optimization component while keeping parameter count comparable. This will isolate the contribution of the factorization mechanism. revision: yes
Referee: [§3 (FLR framework)] The abstract and method description state that the regularizations encourage diversity and specialization, yet the paper does not demonstrate that these objectives actually achieve the intended effect beyond the loss terms themselves (e.g., no post-training analysis of factor independence or specialization on held-out data).

Authors: We agree that post-training verification of the regularizations' effects is required. We will add to the revised manuscript: (1) factor independence metrics (pairwise correlations and mutual information) evaluated on held-out test data, (2) quantitative attention diversity and sparsity statistics achieved after training, and (3) qualitative case studies illustrating factor specialization (e.g., distinct factors focusing on different preference aspects such as temporal recency versus item category). These analyses will be placed in Section 4 or a dedicated subsection. revision: yes

Circularity Check

0 steps flagged

Novel framework with independent design choices; no reduction to inputs by construction

full rationale

The paper introduces FLR as a new architecture: a multi-factor attention module plus orthogonality/attention-diversity/sparsity regularizers and group-relative policy optimization. These are presented as modeling decisions to address multi-faceted preferences, not as quantities derived from or equivalent to prior fitted parameters or self-cited results. No equations or claims reduce one element to another by definition (e.g., no 'prediction' that is the regularization term itself). Experiments on benchmarks are cited as support rather than a closed self-referential loop. The derivation chain is therefore self-contained and additive rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

The abstract introduces the FLR framework and its components without referencing external axioms or fitted constants; the main additions are the new attention module and regularization terms.

invented entities (2)

multi-factor attention module no independent evidence
purpose: iteratively refines a latent thought representation where each factor attends to distinct aspects of interaction history
Core new component introduced to decompose reasoning.
orthogonality, attention diversity, and sparsity regularization objectives no independent evidence
purpose: encourage diversity and specialization among factors
New training objectives added to the framework.

pith-pipeline@v0.9.0 · 5467 in / 1253 out tokens · 98529 ms · 2026-05-07T11:13:08.919843+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

46 extracted references · 20 canonical work pages · 4 internal anchors

[1]

Keqin Bao, Jizhi Zhang, Wenjie Wang, et al. 2025. A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems.Trans. Recomm. Syst.3, 4 (2025), 53:1–53:27

2025
[2]

Millennium Bismay, Xiangjue Dong, and James Caverlee. 2025. ReasoningRec: Bridging Personalized Recommendations and Human-Interpretable Explanations through LLM Reasoning. InFindings of the Association for Computational Lin- guistics: NAACL 2025, Albuquerque, New Mexico, USA, April 29 - May 4, 2025. Association for Computational Linguistics, 8132–8148

2025
[3]

Yuntian Deng, Kiran Prasad, Roland Fernandez, et al . 2023. Implicit Chain of Thought Reasoning via Knowledge Distillation.CoRRabs/2311.01460 (2023). arXiv:2311.01460

work page arXiv 2023
[4]

Yi Fang, Wenjie Wang, Yang Zhang, et al. 2025. Reason4Rec: Large Language Models for Recommendation with Deliberative User Preference Alignment.CoRR abs/2502.02061 (2025). arXiv:2502.02061

work page arXiv 2025
[5]

Jonas Geiping, Sean McLeish, Neel Jain, et al. 2025. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach.CoRRabs/2502.05171 (2025). arXiv:2502.05171

work page internal anchor Pith review arXiv 2025
[6]

Hao Gu, Rui Zhong, Yu Xia, et al . 2025. R4ec: A Reasoning, Reflection, and Refinement Framework for Recommendation Systems. InProceedings of the Nineteenth ACM Conference on Recommender Systems (RecSys ’25). Association for Computing Machinery, New York, NY, USA, 411–421

2025
[7]

Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, et al. 2024. Training Large Language Models to Reason in a Continuous Latent Space.CoRRabs/2412.06769 (2024). arXiv:2412.06769

work page internal anchor Pith review arXiv 2024
[8]

Balázs Hidasi and Alexandros Karatzoglou. 2018. Recurrent Neural Networks with Top-k Gains for Session-based Recommendations. InProceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM 2018, Torino, Italy, October 22-26, 2018. ACM, 843–852

2018
[9]

Chengkai Huang, Xiaodi Chen, Hongtao Huang, Quan Z Sheng, and Lina Yao
[10]

Generative Chain of Behavior for User Trajectory Prediction.arXiv preprint arXiv:2601.18213(2026)

work page arXiv 2026
[11]

Chengkai Huang, Hongtao Huang, Tong Yu, Kaige Xie, Junda Wu, Shuai Zhang, Julian Mcauley, Dietmar Jannach, and Lina Yao. 2025. A Survey of Founda- tion Model-Powered Recommender Systems: From Feature-Based, Generative to Agentic Paradigms.arXiv preprint arXiv:2504.16420(2025)

work page arXiv 2025
[12]

Chengkai Huang, Shoujin Wang, Xianzhi Wang, and Lina Yao. 2023. Dual con- trastive transformer for hierarchical preference modeling in sequential recom- mendation. InProceedings of the 46th international acm sigir conference on research and development in information retrieval. 99–109

2023
[13]

Chengkai Huang, Shoujin Wang, Xianzhi Wang, and Lina Yao. 2023. Modeling temporal positive and negative excitation for sequential recommendation. In Proceedings of the ACM Web Conference 2023. 1252–1263. Factorized Latent Reasoning for LLM-based Recommendation Conference’17, July 2017, Washington, DC, USA

2023
[14]

Chengkai Huang, Junda Wu, Yu Xia, Zixu Yu, Ruhan Wang, Tong Yu, Ruiyi Zhang, Ryan A Rossi, Branislav Kveton, Dongruo Zhou, et al . 2025. Towards agentic recommender systems in the era of multimodal large language models.arXiv preprint arXiv:2503.16734(2025)

work page arXiv 2025
[15]

Chengkai Huang, Yu Xia, Rui Wang, Kaige Xie, Tong Yu, Julian McAuley, and Lina Yao. 2025. Embedding-informed adaptive retrieval-augmented generation of large language models. InProceedings of the 31st International Conference on Computational Linguistics. 1403–1412

2025
[16]

Hongtao Huang, Chengkai Huang, Junda Wu, Tong Yu, Julian McAuley, and Lina Yao. 2025. Listwise Preference Diffusion Optimization for User Behavior Trajectories Prediction.arXiv preprint arXiv:2511.00530(2025)

work page arXiv 2025
[17]

Hongtao Huang, Chengkai Huang, Tong Yu, Xiaojun Chang, Wen Hu, Julian McAuley, and Lina Yao. 2026. Dual Conditional Diffusion for Sequential Recom- mendation. InProceedings of the Nineteenth ACM International Conference on Web Search and Data Mining. 206–216

2026
[18]

Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Rec- ommendation. InIEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018. IEEE Computer Society, 197–206

2018
[19]

Alex Kendall, Yarin Gal, and Roberto Cipolla. 2018. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. InProceedings of the IEEE conference on computer vision and pattern recognition. 7482–7491

2018
[20]

Jieyong Kim, Hyunseo Kim, Hyunjin Cho, et al. 2025. Review-driven Personalized Preference Reasoning with Large Language Models for Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2025, Padua, Italy, July 13-18, 2025. ACM, 1697–1706

2025
[21]

Xiaoyu Kong, Junguang Jiang, Bin Liu, et al. 2025. Think before Recommendation: Autonomous Reasoning-enhanced Recommender.CoRRabs/2510.23077 (2025). arXiv:2510.23077

work page arXiv 2025
[22]

Koren, R

Y. Koren, R. Bell, and C. Volinsky. 2009. Matrix Factorization Techniques for Recommender Systems.Computer42, 8 (2009), 30–37

2009
[23]

Enze Liu, Bowen Zheng, Xiaolei Wang, et al. 2025. LARES: Latent Reasoning for Sequential Recommendation.CoRRabs/2505.16865 (2025)

work page arXiv 2025
[24]

Jiahao Liu, Xueshuo Yan, Dongsheng Li, et al. 2025. Improving LLM-powered Recommendations with Personalized Information. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2025, Padua, Italy, July 13-18, 2025. ACM, 2560–2565

2025
[25]

Zhiqiang Liu, Yanxia Liu, and Chengkai Huang. 2021. Semi-online knowledge distillation.arXiv preprint arXiv:2111.11747(2021)

work page arXiv 2021
[26]

Zhanyu Liu, Shiyao Wang, Xingmei Wang, et al . 2025. OneRec-Think: In- Text Reasoning for Generative Recommendation.CoRRabs/2510.11639 (2025). arXiv:2510.11639

work page arXiv 2025
[27]

Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

work page internal anchor Pith review arXiv 2017
[28]

Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 188–197

2019
[29]

Le Pan, Yuanjiang Cao, Chengkai Huang, Wenjie Zhang, and Lina Yao. 2025. Counterfactual inference for eliminating sentiment bias in recommender systems. arXiv preprint arXiv:2505.03655(2025)

work page arXiv 2025
[30]

Liwei Pan, Weike Pan, Meiyan Wei, et al. 2026. A survey on sequential recom- mendation.Frontiers Comput. Sci.20, 3 (2026), 2003606

2026
[31]

A Yang Qwen, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengpeng Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2. 5 technical report.arXiv preprint(2024)

2024
[32]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

work page internal anchor Pith review arXiv 2024
[33]

Leheng Sheng, An Zhang, Yi Zhang, et al . 2025. Language Representations Can be What Recommenders Need: Findings and Potentials. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net

2025
[34]

Jiakai Tang, Sunhao Dai, Teng Shi, et al. 2025. Think Before Recommend: Un- leashing the Latent Reasoning Power for Sequential Recommendation.CoRR abs/2503.22675 (2025). arXiv:2503.22675

work page arXiv 2025
[35]

Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. InProceedings of the Eleventh ACM International Conference on Web Search and Data Mining, WSDM 2018, Marina Del Rey, CA, USA, February 5-9, 2018. ACM, 565–573

2018
[36]

Alicia Tsai, Adam Kraft, Long Jin, et al. 2024. Leveraging LLM Reasoning En- hances Personalized Recommender Systems. InFindings of the Association for Computational Linguistics: ACL 2024. Association for Computational Linguistics, Bangkok, Thailand, 13176–13188

2024
[37]

Wenjie Wang, Xinyu Lin, Fuli Feng, et al. 2023. Generative recommendation: Towards next-generation recommender paradigm.arXiv preprint arXiv:2304.03516 (2023)

work page arXiv 2023
[38]

Xin Wang, Hong Chen, Yuwei Zhou, et al. 2023. Disentangled Representation Learning for Recommendation.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 1 (2023), 408–424

2023
[39]

Yuling Wang, Changxin Tian, Binbin Hu, et al. 2024. Can Small Language Models be Good Reasoners for Sequential Recommendation?. InProceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024. ACM, 3876– 3887

2024
[40]

Jason Wei, Xuezhi Wang, Dale Schuurmans, et al. 2022. Chain-of-thought prompt- ing elicits reasoning in large language models.Advances in neural information processing systems35 (2022), 24824–24837

2022
[41]

Xiaoxin Ye, Chengkai Huang, Hongtao Huang, and Lina Yao. 2025. Beyond negative transfer: Disentangled preference-guided diffusion for cross-domain sequential recommendation.arXiv preprint arXiv:2509.00389(2025)

work page arXiv 2025
[42]

Xiaoxin Ye, Chengkai Huang, Hongtao Huang, and Lina Yao. 2026. Gaussian Mixture Flow Matching with Domain Alignment for Multi-Domain Sequential Recommendation. InProceedings of the ACM Web Conference 2026. 6159–6170

2026
[43]

Runyang You, Yongqi Li, Xinyu Lin, et al. 2025. Rˆ2ec: Towards Large Recom- mender Models with Reasoning. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

2025
[44]

Weiqi Yue, Yuyu Yin, Xin Zhang, et al. 2025. CoT4Rec: Revealing User Preferences Through Chain of Thought for Recommender Systems. InAAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA. AAAI Press, 13142–13151

2025
[45]

Yang Zhang, Wenxin Xu, Xiaoyan Zhao, et al. 2025. Reinforced Latent Reasoning for LLM-based Recommendation.arXiv preprint arXiv:2505.19092(2025)

work page arXiv 2025
[46]

Keyu Zhao, Fengli Xu, and Yong Li. 2025. Reason-to-Recommend: Using Interaction-of-Thought Reasoning to Enhance LLM Recommendation.CoRR abs/2506.05069 (2025). arXiv:2506.05069

work page arXiv 2025