Pith. sign in

REVIEW 5 major objections 4 minor 5 cited by

Chain-of-thought reasoning degrades semantic-ID recommender accuracy; the paper diagnoses the cause and fixes it without retraining.

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

2026-08-02 22:27 UTC pith:FITETHI2

load-bearing objection A useful, training-free fix for a real failure mode in semantic-ID recommenders, but the causal diagnosis is shakier than the method. the 5 major comments →

arxiv 2602.16587 v2 pith:FITETHI2 submitted 2026-02-18 cs.IR

Why Thinking Hurts: Diagnosing and Rectifying Linguistic Inertia in Large Language Models for Recommendation

Luankang Zhang , Yonghao Huang , Hang Lv , Xuyang Zhi , Mingjia Yin , Yuyang Ye , Wei Guo , Hao Wang
show 1 more author
Enhong Chen
This is my paper
classification cs.IR
keywords generative recommendationchain-of-thought reasoningsemantic IDlinguistic inertiacontrastive decodinginference-time alignmentLLM recommendationOpenOneRec
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 claims that adding Chain-of-Thought reasoning before item prediction in semantic-ID-based recommender models backfires: thinking mode can cut accuracy by up to 25%. The cause, the authors argue, is linguistic inertia—when free-form reasoning text is inserted, the model's decoding shifts toward general language patterns and pays less attention to the item-history evidence encoded in semantic IDs. The paper decomposes the prediction score into a useful history-grounded term and a noisy 'general subspace prior,' and proposes a training-free correction called LICD: compress the reasoning chain into a one-sentence preference summary, then rerank candidates by subtracting only the CoT-induced bias that disagrees with the history-only baseline. The result, if true, is that recommendation models can keep the benefits of reasoning without sacrificing ID-grounded accuracy and without retraining.

Core claim

The central finding is that the performance drop from 'thinking mode' in OpenOneRec is not a failure of reasoning itself but a distributional shift: the model's latent space contains a general-text subspace and a Semantic-ID subspace that overlap but are not aligned, and a long CoT chain makes decoding drift toward the general-text prior. Formally, the paper writes the thinking-mode score as S(y|x,c) = CPMI(y;x|c) + S(y|c), identifying the conditional pointwise mutual information term as the useful Semantic-ID-consistent evidence and the CoT-only score as the harmful bias. Empirically, thinking mode raises the Space Dominance Index (attention on general tokens relative to ID tokens) and lowe

What carries the argument

The load-bearing object is the contrastive scoring identity S(y) = (1+α) z̃_E(y) − α(z̃_A(y)−z̃_B(y)), built from three normalized context scores: Expert (history plus compressed preference summary), Amateur (null history plus raw CoT chain), and Baseline (history only). Its work is to estimate 'ungrounded drift' as the gap between CoT-only and history-only scores and subtract exactly that gap while keeping the expert score. Supporting machinery: the CPMI decomposition of Eqs. (1)–(2), the compression operator ĉ = T(c) that maps free-form reasoning to a short fixed-template preference sentence, and the diagnostic indices SDI and AEI that quantify attention dominance.

Load-bearing premise

The method assumes that the difference between the CoT-only score (with a null history) and the history-only score isolates 'ungrounded textual drift'—but the CoT chain was generated from the real history, so the subtraction may also remove genuine history-grounded evidence; the scalar α is then papering over that leakage.

What would settle it

Run LICD while replacing the Amateur context's reasoning chain with a chain sampled for a different user (or a random chain) while keeping everything else fixed. If the method still yields the same gains, the penalty term is not removing user-specific grounded evidence, and the claimed mechanism—subtracting only ungrounded drift—is not what is doing the work. A second check: if the gains persist when the compression step is removed, the bias subtraction alone is the active ingredient, which would contradict the paper's two-component story.

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

If this is right

  • If LICD works as claimed, thinking mode can be switched on for semantic-ID recommenders without the 25% accuracy penalty, at inference time only.
  • Reasoning-augmented recommendation becomes viable without retraining or re-aligning the backbone.
  • The diagnosis predicts that any verbose inserted text—not just CoT—can bias decoding toward the general subspace; compression is the corrective.
  • The bias-subtraction scheme offers a template for removing model-specific 'excess' conditioning drift while retaining grounded signal.
  • The approach transfers across backbone scales (1.7B and 8B), suggesting it is a calibration layer rather than a model-specific patch.

Where Pith is reading between the lines

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

  • The Amateur context's chain c is generated from the real history x, so z_A may carry genuine history-grounded evidence; if so, the subtraction could remove real signal rather than only textual drift. A clean test: generate c from another user's history and see whether gains vanish.
  • The same contrastive-correction pattern might apply beyond recommendation—any task where a language model conditions on both structured IDs and free text and drifts toward linguistic fluency.
  • The abstract promises three benchmarks, but the results table reports two (AD and Product); Video appears only in the motivating figure, so the 'consistently outperforms' claim is currently evidenced on two domains.
  • The manuscript's cross-references to 'Section 4.2' for the empirical analysis point to the wrong section (the analysis is in Section 3); this does not affect the method but should be corrected.

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

5 major / 4 minor

Summary. The paper investigates why enabling chain-of-thought reasoning in the OpenOneRec semantic-ID recommender degrades recommendation quality. It attributes the degradation to 'General Subspace Prior' / 'Linguistic Inertia': free-form CoT text shifts attention away from historical Semantic-ID evidence. The authors propose a training-free framework, LICD, that compresses CoT into a structured preference summary and applies bias-subtracted contrastive decoding with three scoring contexts. Experiments on AD and Product report consistent improvements over no-thinking and original-thinking baselines. A code repository is provided.

Significance. If the diagnosis were established, the paper would offer a practical training-free remedy for CoT-enabled generative recommender systems and a clear information-theoretic perspective on why verbose reasoning can hurt. The CPMI decomposition in Eq. (1) is exact and the idea of penalizing only 'excess' CoT support is attractive. The claim is, however, strongly mechanism-dependent: the main evidence for 'ungrounded drift' is correlational and the correction relies on a tuned scalar. With additional controls and reporting, the contribution could be a useful, reproducible decoding-time calibration layer.

major comments (5)
  1. [§3.1, Eq. (2)] The equation is an identity, but the labels are an interpretation. Calling S(y|c) a 'Harmful Bias (General Subspace Prior)' assumes that the CoT-only distribution contributes no useful evidence about y. Since c is generated from x, S(y|c) can carry history-grounded information. Please provide a direct test, e.g., estimate I(y; c | x) or compare decoding with a chain generated from a null history, and then label Eq. (2) as a possible decomposition rather than an established causal one.
  2. [§4.2, Eq. (5) and Eq. (8)] The 'Amateur (CoT-only)' context scores P(y | x∅, c) with c generated from the actual history x. Subtracting z_A - z_B therefore removes not only ungrounded drift but also any genuine evidence carried by c. This is acknowledged in the paragraph before Eq. (4), but the method proceeds as if the separation is clean. A necessary control is to also evaluate using a chain generated from x∅, or to report the conditional mutual information I(y; c | x); otherwise the observed gains may simply reflect down-weighting the whole reasoning chain.
  3. [§4.2, Eq. (8); §5.1] The correction strength α is a free parameter whose value and selection protocol are not reported. Because the main experimental claim is that Eq. (8) 'validates our core mechanism', the paper must state how α was chosen for each backbone/dataset and show sensitivity. Without this, the improvement could be a tuned linear interpolation between the expert score and the drift-corrected score rather than evidence for the mechanism.
  4. [§5.1, Table 2, Fig. 1] The abstract and introduction promise 'three large-scale benchmarks', but Table 2 reports results only for AD and Product. Figure 1 mentions Video, but no Video results or experimental setup appear. Either provide the third benchmark in the main text or appendix, or correct the abstract. The 'consistently outperforms' claim is currently supported by only two datasets.
  5. [§3.2, Table 1] The SDI and AEI metrics are not defined precisely enough to be reproducible: it is unclear over which layers, heads, and token positions the attention weights are aggregated, and no error bars or statistical tests are reported. Since Table 1 is the central evidence for the attention-dominance mechanism, please include definitions and variability measures (e.g., over multiple random chain samples or test instances).
minor comments (4)
  1. [§4, §4.2] The text refers to 'the empirical analysis in Section 4.2' and 'As shown in Section 4.2' where the analysis is in Section 3.2. Please correct the cross-references.
  2. [Fig. 2] The caption contains the typo 'Visiualization'. Also, please report the explained variance of the PCA and the number of token samples used; the figure alone does not show how distinct the subspaces are.
  3. [Table 2] The caption states '(p-value<0.05)' but does not identify the test or what comparison it applies to. Please provide the test procedure and report error bars or confidence intervals.
  4. [§4.1, Fig. 3] The compression operator is said to use a 'strict length budget' and a fixed template, but no concrete budget or template string is given. Please specify these in an appendix for reproducibility.

Circularity Check

3 steps flagged

Partially self-definitional diagnosis: Eq. (2) labels the CoT-only score term 'Harmful Bias' as an interpretive assertion, and Eq. (5)'s Amateur context conditions on a chain generated from x, so LICD's subtraction may remove history-grounded evidence; the correction weight α is also unreported.

specific steps
  1. self definitional [Section 3.1, Eq. (2)]
    "Consequently, the prediction score decomposes into: S(y|x,c)=CPMI(y;x|c) Useful Bias (Semantic ID Consistency) + S(y|c) Harmful Bias (General Subspace Prior)"

    The decomposition is an exact identity, but the label 'Harmful Bias' is an interpretive claim, not a derived result. The entire diagnosis—that CoT degrades performance via the 'General Subspace Prior'—rests on calling S(y|c) harmful. The proposed method then subtracts an estimate of exactly this term, so any improvement after removal is partly a definitional consequence of the labeling rather than independent evidence that the term is harmful.

  2. self definitional [Section 4.2, Eqs. (5)-(8)]
    "However, this creates a key tension: the chain c is generated from the interaction history x, and thus may carry history-grounded intermediate deductions that are genuinely predictive of the target semantic ID."

    The 'Amateur (CoT-only)' context in Eq. (5) replaces x with a null prompt x∅ but retains c, which was sampled from P(c|x). Thus z_A(y)=log P(y|x∅,c) is not a pure general-subspace prior; it carries user-specific evidence mediated by c. The paper defines ungrounded drift as z_A−z_B and then subtracts this quantity in Eq. (8). The definition of 'excess CoT influence' is therefore chosen to match the operation performed, and LICD's gains may simply reflect down-weighting the entire reasoning chain rather than removing only ungrounded textual drift.

  3. fitted input called prediction [Section 4.2, Eq. (8); Section 5]
    "S(y)=(1+α) z̃_E(y) − α(z̃_A(y)−z̃_B(y)), where α≥0 controls the correction strength."

    The framework is described as 'training-free,' but Eq. (8) contains a free scalar α, and no value or selection protocol is reported. The experimental validation is therefore not a parameter-free confirmation of the mechanism; it demonstrates that some tuned linear combination of expert, CoT-only, and history-only scores improves metrics. This leaves the central claim underdetermined: the improvement does not establish that the subtracted component is harmful ungrounded drift rather than useful history-grounded reasoning.

full rationale

The paper does not rely on self-citation in a load-bearing way, and the CPMI identity in Eq. (1) is mathematically exact. However, the central diagnostic step is partially self-definitional: Eq. (2) names the CoT-only score S(y|c) the 'Harmful Bias (General Subspace Prior),' and the method proceeds to subtract an estimate of that term. The Amateur context used to estimate it is contaminated because c was generated from x, as the paper itself concedes in §4.2. Moreover, the free α in Eq. (8) is never specified, so the reported gains could arise from re-weighting rather than from selectively removing ungrounded drift. These issues make the causal mechanism underdetermined, but they do not make the entire derivation equivalent to its inputs: the empirical comparison against OpenOneRec-think is a genuine benchmark and the proposed score is a concrete, testable transformation. Overall circularity is moderate.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The central claim rests on a conceptual split between subspaces, an asserted harmful-bias labeling, and a tuned scalar α; none of these are independently validated. The only standard steps are the CPMI identity and z-score normalization.

free parameters (2)
  • α (correction strength) = not reported (tuned)
    Controls the amount of drift subtracted in Eq. (8); no value or selection procedure is given, so the method's gains depend on this fitted scalar.
  • compression length budget / template = not reported
    The compressed preference statement is constrained to be 'short' and 'one sentence' (§4.1/Fig. 3), a hand-chosen inductive bias; its exact budget is not specified.
axioms (4)
  • domain assumption The token space splits into a 'General Subspace' and a 'Semantic ID Subspace' that are partially overlapping but distinct (§3.1, Fig. 2).
    The PCA visualization is offered as evidence, but no formal separation or metric is defined.
  • ad hoc to paper S(y|c) in Eq. (2) is a 'Harmful Bias' rather than evidence-bearing reasoning.
    This labeling is asserted; since c is sampled from x, S(y|c) can carry history-grounded information.
  • standard math z-score normalization makes the three context scores commensurable for subtraction in Eq. (8).
    Standard normalization, but its validity across distributions of different entropy is assumed.
  • ad hoc to paper The compressor T preserves preference-relevant signal while removing 'high-entropy linguistic surface forms' (§4.1).
    No evaluation of compression fidelity is provided.
invented entities (1)
  • General Subspace Prior / Linguistic Inertia no independent evidence
    purpose: Explains why CoT degrades SID prediction by dominating inference and diluting Semantic ID evidence.
    It is inferred from PCA and attention statistics on the same model and validated only through the proposed correction; it has no falsifiable handle outside the paper (e.g., a measurable prediction independent of the method).

pith-pipeline@v1.3.0-alltime-deepseek · 9966 in / 13237 out tokens · 117241 ms · 2026-08-02T22:27:19.164033+00:00 · methodology

0 comments
read the original abstract

Chain-of-Thought (CoT) reasoning is widely used to improve LLM performance, and recent foundation recommender models adopt it by generating textual reasoning before predicting target items represented by Semantic IDs (SIDs). However, we observe that enabling thinking mode in models such as OpenOneRec can degrade recommendation quality by up to 25%. We investigate this failure and identify Linguistic Inertia: when a textual CoT segment is inserted before SID generation, the model relies more on natural-language context and less on historical SID evidence. Further analyses show that this effect is amplified by reduced access to historical information and longer CoT lengths. To mitigate it, we propose Linguistic-Inertia-Calibrated Decoding (LICD), a training-free framework that combines Reasoning-Chain Compression and Bias-Subtracted Contrastive Inference. Experiments on three large-scale benchmarks show that LICD consistently outperforms both no-thinking and original-thinking baselines. Our code is available at https://anonymous.4open.science/r/LICD-4573.

Figures

Figures reproduced from arXiv: 2602.16587 by Enhong Chen, Hang Lv, Hao Wang, Luankang Zhang, Mingjia Yin, Wei Guo, Xuyang Zhi, Yonghao Huang, Yuyang Ye.

Figure 1
Figure 1. Figure 1: Impact of Thinking Mode on OpenOneRec. However, a critical anomaly emerges when integrating explicit Chain-of-Thought (CoT) reasoning into these ID-based founda￾tion models. While reasoning capabilities are pivotal for capturing complex user preferences, we observe that enabling the “thinking mode” in foundation recommender models paradoxically leads to performance degradation ( [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 2
Figure 2. Figure 2: Visiualization of Two Token Subspaces. Consider a generative recommender model where the input 𝑥 comprises tokens from two partially overlapping subspaces: the Semantic ID Subspace (recommendation items) and the General Subspace (textual instructions). The target output 𝑦 is the ground￾truth semantic ID. In the thinking mode, the model first generates a chain of thought 𝑐 typically within the general subsp… view at source ↗
Figure 3
Figure 3. Figure 3: Prompt Structure for Section 4.1. Compression operator. We define a deterministic transformation 𝑐ˆ = T (𝑐), 𝑐ˆ ∈ Cpref, (3) where Cpref denotes a restricted space of preference statements. Concretely, 𝑐ˆ is constrained to a fixed template and a strict length budget, which prevents the accumulation of general-text inertia while preserving the key preference cues that are useful for pre￾dicting the target s… view at source ↗

discussion (0)

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

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. IE as Cache: Information Extraction Enhanced Agentic Reasoning

    cs.CL 2026-04 unverdicted novelty 7.0

    IE-as-Cache framework repurposes information extraction as a dynamic cognitive cache to improve agentic reasoning accuracy in LLMs on challenging benchmarks.

  2. WhisperRec: Latent Reasoning for Efficient Foundation Recommendation Models

    cs.IR 2026-07 conditional novelty 6.0

    WhisperRec distills multi-view chain-of-thought rationale into three latent tokens, beating explicit-reasoning recommenders at about ten times the inference throughput.

  3. Beyond Fixed Depths and Widths: Optimizing Textual Decoding Tries in LLM-based Generative Recommendation

    cs.IR 2026-07 conditional novelty 6.0

    BONSAI constructs variable-depth, low-branching decoding tries for LLM-based generative recommendation and reports 16–22% relative gains over state-of-the-art baselines.

  4. SAPO: Step-Aligned Policy Optimization for Reasoning-Based Generative Recommendation

    cs.AI 2026-05 unverdicted novelty 6.0

    SAPO computes per-reasoning-step group-relative advantages in RL to improve credit assignment for structured generation of semantic identifiers in recommendation systems.

  5. Rethinking the Necessity of Adaptive Retrieval-Augmented Generation through the Lens of Adaptive Listwise Ranking

    cs.IR 2026-04 unverdicted novelty 5.0

    AdaRankLLM shows adaptive listwise reranking outperforms fixed-depth retrieval for most LLMs by acting as a noise filter for weak models and an efficiency optimizer for strong ones, with lower context use.

Reference graph

Works this paper leans on

59 extracted references · 20 linked inside Pith · cited by 5 Pith papers

  1. [1]

    Hervé Abdi and Lynne J Williams. 2010. Principal component analysis.Wiley interdisciplinary reviews: computational statistics2, 4 (2010), 433–459

    2010

  2. [2]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  3. [3]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM conference on recommender systems. 1007–1014

    2023

  4. [4]

    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 Linguistics: NAACL 2025. 8132–8148

    2025

  5. [5]

    1999.Elements of information theory

    Thomas M Cover. 1999.Elements of information theory. John Wiley & Sons

    1999

  6. [6]

    Yi Fang, Wenjie Wang, Yang Zhang, Fengbin Zhu, Qifan Wang, Fuli Feng, and Xiangnan He. 2025. Reason4Rec: Large Language Models for Recommendation with Deliberative User Preference Alignment.arXiv preprint arXiv:2502.02061 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-rec: Towards interactive and explainable llms-augmented recommender system.arXiv preprint arXiv:2303.14524(2023)

    Pith/arXiv arXiv 2023

  8. [8]

    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

  9. [9]

    Wei Guo, Hao Wang, Luankang Zhang, Jin Yao Chin, Zhongzhou Liu, Kai Cheng, Qiushi Pan, Yi Quan Lee, Wanqi Xue, Tingjia Shen, et al. 2024. Scaling new fron- tiers: Insights into large recommendation models.arXiv preprint arXiv:2412.00714 (2024)

    Pith/arXiv arXiv 2024

  10. [10]

    Ari Holtzman, Peter West, Vered Shwartz, Yejin Choi, and Luke Zettlemoyer

  11. [11]

    Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2023. Learn- ing Vector-Quantized Item Representation for Transferable Sequential Recom- menders. arXiv:2210.12316 [cs.IR] https://arxiv.org/abs/2210.12316

    Pith/arXiv arXiv 2023

  12. [12]

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

    2018

  13. [13]

    Jieyong Kim, Hyunseo Kim, Hyunjin Cho, SeongKu Kang, Buru Chang, Jinyoung Yeo, and Dongha Lee. [n. d.]. Review-driven personalized preference reasoning with large language models for recommendation. CoRR, abs/2408.06276, 2024. doi: 10.48550.arXiv preprint ARXIV.2408.06276([n. d.])

    Pith/arXiv arXiv 2024

  14. [14]

    Minjin Kim, Minju Kim, Hana Kim, Beong-woo Kwak, Soyeon Chun, Hyunseo Kim, SeongKu Kang, Youngjae Yu, Jinyoung Yeo, and Dongha Lee. 2024. Pearl: A review-driven persona-knowledge grounded conversational recommendation dataset.arXiv preprint arXiv:2403.04460(2024)

    Pith/arXiv arXiv 2024

  15. [15]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. Large Language Models are Zero-Shot Reasoners. arXiv:2205.11916 [cs.CL] https://arxiv.org/abs/2205.11916

    Pith/arXiv arXiv 2023

  16. [16]

    Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. 2024. Large Language Models for Generative Recommendation: A Survey and Visionary Discus- sions. InProceedings of the 2024 Joint International Conference on Computa- tional Linguistics, Language Resources and Evaluation (LREC-COLING 2024), Nicoletta Calzolari, Min-Yen Kan, Veronique Hoste, Alessandro Lenci,...

    2024

  17. [17]

    Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang

  18. [18]

    Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al

  19. [19]

    Andrzej Maćkiewicz and Waldemar Ratajczak. 1993. Principal components analysis (PCA).Computers & Geosciences19, 3 (1993), 303–342

    1993

  20. [20]

    Yatin Nandwani, Vineet Kumar, Dinesh Raghu, Sachindra Joshi, and Luis Lastras

  21. [21]

    Qiushi Pan, Hao Wang, Guoyuan An, Luankang Zhang, Wei Guo, and Yong Liu

  22. [22]

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al

  23. [23]

    Milad Sabouri, Masoud Mansoury, Kun Lin, and Bamshad Mobasher. 2025. To- wards Explainable Temporal User Profiling with LLMs. InAdjunct Proceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization. 219–227

    2025

  24. [24]

    InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.)

    Pointwise Mutual Information Based Metric and Decoding Strategy for Faithful Generation in Document Grounded Dialogs. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 10335–10347. doi:10.18653/v1/2023.emnlp-main.639

    doi:10.18653/v1/2023.emnlp-main.639 2023

  25. [25]

    Zhu Sun, Hongyang Liu, Xinghua Qu, Kaidong Feng, Yan Wang, and Yew Soon Ong. 2024. Large language models for intent-driven session recommendations. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 324–334

    2024

  26. [26]

    Revisiting scalable sequential recommendation with Multi-Embedding Approach and Mixture-of-Experts.arXiv preprint arXiv:2510.25285(2025)

    arXiv 2025

  27. [27]

    Alicia Tsai, Adam Kraft, Long Jin, Chenwei Cai, Anahita Hosseini, Taibai Xu, Zemin Zhang, Lichan Hong, Ed H Chi, and Xinyang Yi. 2024. Leveraging llm reasoning enhances personalized recommender systems. InFindings of the Asso- ciation for Computational Linguistics: ACL 2024. 13176–13188. Luankang Zhang et al

    2024

  28. [28]

    Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

    2023

  29. [29]

    Hao Wang, Mingjia Yin, Luankang Zhang, Sirui Zhao, and Enhong Chen. 2025. Mf-gslae: A multi-factor user representation pre-training framework for dual- target cross-domain recommendation.ACM Transactions on Information Systems 43, 2 (2025), 1–28

    2025

  30. [30]

    Chi, and Xinyang Yi

    Anima Singh, Trung Vu, Nikhil Mehta, Raghunandan Keshavan, Maheswaran Sathiamoorthy, Yilin Zheng, Lichan Hong, Lukasz Heldt, Li Wei, Devansh Tandon, Ed H. Chi, and Xinyang Yi. 2024. Better Generalization with Semantic IDs: A Case Study in Ranking for Recommendations. arXiv:2306.08121 [cs.IR] https: //arxiv.org/abs/2306.08121

    Pith/arXiv arXiv 2024

  31. [31]

    Kefan Wang, Hao Wang, Kenan Song, Wei Guo, Kai Cheng, Zhi Li, Yong Liu, Defu Lian, and Enhong Chen. 2025. A universal framework for compressing embeddings in ctr prediction. InInternational Conference on Database Systems for Advanced Applications. Springer, 84–100

    2025

  32. [32]

    Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Jian Wu, and Yuning Jiang. 2025. Think before recommend: Unleashing the latent reasoning power for sequential recommendation.arXiv preprint arXiv:2503.22675(2025)

    Pith/arXiv arXiv 2025

  33. [33]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903 [cs.CL] https: //arxiv.org/abs/2201.11903

    Pith/arXiv arXiv 2023

  34. [34]

    Hao Wang, Wei Guo, Luankang Zhang, Jin Yao Chin, Yufei Ye, Huifeng Guo, Yong Liu, Defu Lian, Ruiming Tang, and Enhong Chen. 2025. Generative large recom- mendation models: emerging trends in llms for recommendation. InCompanion Proceedings of the ACM on Web Conference 2025. 49–52

    2025

  35. [35]

    Wenjia Xie, Hao Wang, Minghao Fang, Ruize Yu, Wei Guo, Yong Liu, Defu Lian, and Enhong Chen. 2025. Breaking the Bottleneck: User-Specific Optimization and Real-Time Inference Integration for Sequential Recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    2025

  36. [36]

    Kefan Wang, Hao Wang, Wei Guo, Yong Liu, Jianghao Lin, Defu Lian, and Enhong Chen. 2025. DLF: Enhancing explicit-implicit interaction via dynamic low-order- aware fusion for CTR prediction. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2213– 2223

    2025

  37. [37]

    Xiang Xu, Hao Wang, Wei Guo, Luankang Zhang, Wanshan Yang, Runlong Yu, Yong Liu, Defu Lian, and Enhong Chen. 2025. Multi-granularity interest retrieval and refinement network for long-term user behavior modeling in ctr prediction. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2745–2755

    2025

  38. [38]

    Qi Wang, Jindong Li, Shiqi Wang, Qianli Xing, Runliang Niu, He Kong, Rui Li, Guodong Long, Yi Chang, and Chengqi Zhang. 2024. Towards Next-Generation LLM-based Recommender Systems: A Survey and Beyond. arXiv:2410.19744 [cs.IR] https://arxiv.org/abs/2410.19744

    Pith/arXiv arXiv 2024

  39. [39]

    Yufei Ye, Wei Guo, Hao Wang, Hong Zhu, Yuyang Ye, Yong Liu, Huifeng Guo, Ruiming Tang, Defu Lian, and Enhong Chen. 2025. Fuxi-\beta: Towards a light- weight and fast large-scale generative recommendation model.arXiv preprint arXiv:2508.10615(2025)

    Pith/arXiv arXiv 2025

  40. [40]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2024. A Survey on Large Language Models for Recommendation. arXiv:2305.19860 [cs.IR] https://arxiv.org/abs/2305.19860

    Pith/arXiv arXiv 2024

  41. [41]

    Haocheng Yu, Yaxiong Wu, Hao Wang, Wei Guo, Yong Liu, Yawen Li, Yuyang Ye, Junping Du, and Enhong Chen. 2025. Thought-augmented planning for llm-powered interactive recommender agent.arXiv preprint arXiv:2506.23485 (2025)

    Pith/arXiv arXiv 2025

  42. [42]

    Wenjia Xie, Hao Wang, Luankang Zhang, Rui Zhou, Defu Lian, and Enhong Chen

  43. [43]

    Junjie Zhang, Beichen Zhang, Wenqi Sun, Hongyu Lu, Wayne Xin Zhao, Yu Chen, and Ji-Rong Wen. 2025. Slow Thinking for Sequential Recommendation. arXiv preprint arXiv:2504.09627(2025)

    Pith/arXiv arXiv 2025

  44. [44]

    Luankang Zhang, Hang Lv, Qiushi Pan, Kefen Wang, Yonghao Huang, Xinrui Miao, Yin Xu, Wei Guo, Yong Liu, Hao Wang, and Enhong Chen. 2026. The Next Paradigm Is User-Centric Agent, Not Platform-Centric Service.arXiv preprint arXiv:2602.15682(2026)

    arXiv 2026

  45. [45]

    Yufei Ye, Wei Guo, Jin Yao Chin, Hao Wang, Hong Zhu, Xi Lin, Yuyang Ye, Yong Liu, Ruiming Tang, Defu Lian, et al . 2025. Fuxi-𝛼: Scaling recommendation model with feature interaction enhanced transformer. InCompanion Proceedings of the ACM on Web Conference 2025. 557–566

    2025

  46. [46]

    Luankang Zhang, Hao Wang, Zhongzhou Liu, Mingjia Yin, Yonghao Huang, Jiaqi Li, Wei Guo, Yong Liu, Huifeng Guo, Defu Lian, and Enhong Chen. 2026. Can Recommender Systems Teach Themselves? A Recursive Self-Improving Framework with Fidelity Control.arXiv preprint arXiv:2602.15659(2026)

    Pith/arXiv arXiv 2026

  47. [47]

    Runyang You, Yongqi Li, Xinyu Lin, Xin Zhang, Wenjie Wang, Wenjie Li, and Liqiang Nie. 2025. R 2ec: Towards Large Recommender Models with Reasoning. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

    2025

  48. [48]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448

    2024

  49. [49]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommenda- tions.arXiv preprint arXiv:2402.17152(2024)

    Pith/arXiv arXiv 2024

  50. [50]

    Jingyi Zhou, Cheng Chen, Kai Zuo, Manjie Xu, Zhendong Fu, Yibo Chen, Xu Tang, and Yao Hu. 2025. HyMiRec: A Hybrid Multi-interest Learning Framework for LLM-based Sequential Recommendation.arXiv preprint arXiv:2510.13738 (2025)

    arXiv 2025

  51. [51]

    Rui Zhou, Qinglin Jia, Bo Chen, Peng Xu, Yijia Sun, Siyuan Lou, Chaoxin Fu, Mengyuan Fu, Guoming Shen, Zheli Zhou, et al. 2026. A Survey of User Lifelong Behavior Modeling: Perspectives on Efficiency and Effectiveness. (2026)

    2026

  52. [52]

    Luankang Zhang, Kenan Song, Yi Quan Lee, Wei Guo, Hao Wang, Yawen Li, Huifeng Guo, Yong Liu, Defu Lian, and Enhong Chen. 2025. Killing two birds with one stone: Unifying retrieval and ranking with a single generative recom- mendation model. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2224–2234

    2025

  53. [54]

    Luankang Zhang, Hao Wang, Suojuan Zhang, Mingjia Yin, Yongqiang Han, Jiaqing Zhang, Defu Lian, and Enhong Chen. 2024. A unified framework for adaptive representation enhancement and inversed learning in cross-domain recommendation. InInternational Conference on Database Systems for Advanced Applications. Springer, 115–130

    2024

  54. [56]

    Guorui Zhou, Honghui Bao, Jiaming Huang, Jiaxin Deng, Jinghao Zhang, Junda She, Kuo Cai, Lejian Ren, Lu Ren, Qiang Luo, et al. 2025. OpenOneRec Technical Report.arXiv preprint arXiv:2512.24762(2025)

    arXiv 2025

  55. [59]

    Rui Zhou, Hao Wang, Wei Guo, Qinglin Jia, Wenjia Xie, Xiang Xu, Yong Liu, Defu Lian, and Enhong Chen. 2025. MIT: A Multi-Tower Information Transfer Framework Based on Hierarchical Task Relationship Modeling. InCompanion Proceedings of the ACM on Web Conference 2025. 651–660. Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    2025

  56. [2022]

    arXiv:2104.08315 [cs.CL] https://arxiv.org/abs/2104.08315

    Surface Form Competition: Why the Highest Probability Answer Isn’t Always Right. arXiv:2104.08315 [cs.CL] https://arxiv.org/abs/2104.08315

    Pith/arXiv arXiv

  57. [2023]

    Is chatgpt a good recommender? a preliminary study.arXiv preprint arXiv:2304.10149(2023)

    Pith/arXiv arXiv 2023

  58. [2024]

    Breaking determinism: Fuzzy modeling of sequential recommendation using discrete state space diffusion model.Advances in Neural Information Processing Systems37 (2024), 22720–22744

    2024

  59. [2025]

    Onerec-think: In-text reasoning for generative recommendation.arXiv preprint arXiv:2510.11639(2025)

    arXiv 2025