Pith. sign in

REVIEW 3 major objections 6 minor 62 references

HiLaR organizes an LLM's hidden reasoning states from broad preferences to fine-grained intent, and rewards each state by its marginal contribution, to improve recommendation ranking beyond homogeneous latent reasoning.

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 01:44 UTC pith:7W3X6ODR

load-bearing objection Solid, incremental work on latent reasoning for LLM recommenders; the hierarchy idea is clean, but the gain reward is under-tested and Table 1 lacks error bars. the 3 major comments →

arxiv 2607.27760 v1 pith:7W3X6ODR submitted 2026-07-30 cs.IR cs.AI

Hierarchical Latent Reasoning for LLM-based Recommendation

Peiyu Hu , Siying Gu , Weihai Lu , Zhuodong Liu , Yuntian Tang , Jiahao Liang , Yiying Xie , Jiang Rong
show 3 more authors
Zhaokai Luo Zhiyong Wang Jia Wang
This is my paper
classification cs.IR cs.AI
keywords LLM-based recommendationlatent reasoninghierarchical preference modelingresidual quantizationprocess rewardcredit assignmentreinforcement learninggenerative recommendation
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.

What this paper is trying to establish is that the hidden reasoning states an LLM produces before recommending an item should not be treated as interchangeable refinement steps. The authors claim that when those states are organized into a temporal coarse-to-fine hierarchy—mirroring how user preferences move from stable long-term interests to recent intent—and when each state is given its own layer-aware reward, the resulting recommender ranks items more accurately than existing latent-reasoning methods. A sympathetic reader would care because latent reasoning promises the benefits of chain-of-thought without the token cost, and this work points to a principled way to give each latent step a role and a credit signal. On four public benchmark datasets, HiLaR generally outperforms strong sequential, generative, and LLM-based baselines, with ablations confirming that the hierarchy, the alignment, and the process rewards each contribute.

Core claim

The central claim is that latent reasoning states in an LLM recommender can be assigned distinct, temporally grounded preference roles: early states encode broad historical preferences, later states encode increasingly recent and fine-grained intent, and the final state encodes the current target intent. HiLaR realizes this by first learning a residual quantizer that decomposes a time-aware user representation into cumulative levels supervised by progressively narrower temporal windows, then fine-tuning the LLM so its K-th latent reasoning state aligns with the K-th quantized preference level, and finally applying a group-relative policy-gradient update whose rewards include, for each state,

What carries the argument

The load-bearing mechanism is the pairing of a temporal-guided residual quantizer with a marginal-gain process reward. The quantizer takes a user's history, encodes it with a frozen LLM, and repeatedly subtracts the previously explained component to produce K cumulative preference embeddings, each supervised by a different temporal window (overall history to the target item). The LLM's K latent reasoning states are aligned to those embeddings, so the hierarchy is transferred into the reasoning process. Then, during reinforcement optimization, each sampled latent state contributes to the reward through the clipped improvement it yields in the target title's log-likelihood, which gives dense p

Load-bearing premise

The central premise that may give way is that the marginal increase in the model's own likelihood of the target item caused by a latent state is a trustworthy proxy for that state's true contribution, so optimizing it will not produce trivial states that inflate likelihood without improving ranking.

What would settle it

Concretely, one could train HiLaR with the process reward replaced by a permutation of the layer-wise gains (or by a random reward of the same scale) and check whether ranking accuracy still improves; or inspect the learned latent states and show that later states do not encode finer-grained preference information than earlier ones after the alignment and reward stages.

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

If this is right

  • If the central claim holds, latent reasoning no longer needs to be a flat sequence; temporal preference granularity is a usable organizing principle for hidden states.
  • Layer-wise marginal-likelihood gains provide a denser training signal than final-only outcomes, which should stabilize trajectory optimization in sparse-reward settings.
  • The hierarchy's benefit grows with history length, suggesting the approach pays off most where user behavior is rich.
  • Because latent states add few decode steps, the method keeps inference efficient compared with explicit chain-of-thought reasoning.
  • The three components—temporal quantization, latent alignment, and process rewards—each contribute, and removing any one lowers ranking accuracy.

Where Pith is reading between the lines

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

  • One consequence the paper leaves implicit: the same coarse-to-fine latent hierarchy could be applied to other sequence-generation tasks (e.g., dialogue or code) where temporal or granular structure is known.
  • The marginal-gain reward is computed on the model's own likelihood of the training target; a stronger test would be to compute gains on a held-out target to see whether the signal is true improvement or overfitting.
  • If the hierarchy claim is right, targeted perturbations—zeroing, swapping, or removing late reasoning states—should degrade ranking more than early states; this is a direct falsifiable prediction the paper did not report.
  • A practical extension would be to combine the marginal-gain reward with verification-based or collaborative-consistency signals to guard against states that merely inflate likelihood.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes HiLaR, a three-stage framework for LLM-based top-N recommendation. It first learns K-level coarse-to-fine user preference embeddings via temporal-guided residual quantization of a frozen LLM's history representation; it then fine-tunes a latent reasoning module whose K latent states are aligned with these embeddings while predicting the target item title; finally it applies GRPO with a reward combining task, collaborative-preference, effectiveness, alignment, and marginal target-likelihood-gain terms. Experiments on four Amazon datasets compare HiLaR against sequential, generative, and LLM-based baselines, reporting general improvements, with ablation and sensitivity analyses supporting the main components.

Significance. If the reported results are stable, HiLaR makes a useful contribution to efficient LLM-based recommendation by showing that temporally grounded coarse-to-fine organization of latent reasoning states can outperform homogeneous latent refinement, and by introducing dense process rewards for latent trajectories. Strengths include a clear problem formulation, a concrete three-stage architecture, a released code repository, and a broad set of baselines and ablations. The use of the target item as supervised training signal is standard and not circular. However, the distinctive process-reward component is not isolated in the ablations, and the statistical reporting is insufficient to verify the small gains over strong baselines; both issues are load-bearing for the paper's central empirical claim.

major comments (3)
  1. [§4.3, Eq. (12)–(16); Fig. 3] The layer-aware gain reward R_g^gain = Σ_k ρ_k clip(Δ_{g,k}, -b, b) is the distinctive component of the proposed process-level optimization, yet it is never ablated in isolation. The ablation variants in Figure 3 remove the entire GRPO, the effectiveness reward, the alignment reward, or the preference reward, but all keep R_g^gain through R_proc. Thus the contribution of the marginal-gain signal is confounded with the other rewards and with GRPO as a whole. Moreover, Δ_{g,k} is the marginal increase in the target title's own log-likelihood under the current policy, making the reward self-referential and susceptible to reward hacking (e.g., concentrating all predictive information in τ_K and rendering earlier states uninformative). To support the claim that process-level optimization contributes, please add a variant with λ_gain = 0 (or equivalent) that retains R_align, R_eff, and the fin
  2. [§5.2, Table 1; §5.1] Main results are averaged over five random seeds, and Table 1 states that improvements over the strongest baseline are evaluated with paired t-tests at p<0.05, but the table reports no standard deviations, confidence intervals, or per-metric significance markers. Several improvements are small — e.g., CDs N@5 is 0.0912 for HiLaR vs 0.0924 for VRec, and Toys N@10 is 0.0801 vs 0.0771 for FLR — so without variance information the reader cannot judge whether the claimed gains are stable. Please report mean ± std (or bootstrap CIs), indicate which differences are significant, and address multiple comparisons across the four datasets and four metrics.
  3. [§5.5, Fig. 6(a)] The evidence that HiLaR's latent states are more differentiated than LatentR3's or VRec's uses the same Δ_{g,k} metric that HiLaR is explicitly trained to maximize. This comparison is partly by construction: the baselines were not optimized for layer-wise target-likelihood gains. To make RQ4's conclusion non-circular, please supplement with an independent analysis — for example, probing early vs late states against the temporal-window labels, or a behavioral study showing that early states encode coarse preferences and late states encode recent intents. As written, the layer-wise gain figure mainly reflects the training objective.
minor comments (6)
  1. [Abstract; §5.2] The abstract says HiLaR 'generally outperforms' baselines, which is accurate, but on CDs N@5 VRec achieves 0.0924 while HiLaR achieves 0.0912. Please state the number of metric-dataset cells where HiLaR is best, or otherwise make the exception explicit.
  2. [§4.1, Eq. (6)] The temporal-window split into 'equal-count chronological windows' is a key inductive bias but is not precisely defined. Please specify how ties and variable history lengths are handled, and consider adding sensitivity analysis for the window boundaries, since Figure 4 does not vary this hyperparameter.
  3. [§4.3, text after Eq. (12)] The statement that 'ℓ_{g,0} is computed without latent states' is ambiguous. Clarify what input the model receives in that case — an empty latent sequence, a zero vector, or something else — and ensure the definition is consistent across SFT, reward computation, and evaluation.
  4. [§4.2 vs §4.3] Alignment uses squared L2 distance (L_align) during SFT but cosine similarity (R_align) in GRPO. Please justify this metric mismatch or use a consistent similarity measure.
  5. [References; §5.1] The text repeatedly refers to Qwen2.5-1.5B, but reference [48] is the Qwen2 Technical Report. Please cite the Qwen2.5 model report or adjust the reference to match the model actually used.
  6. [§5.6, Fig. 7] Figure 7 shows efficiency comparisons only as bars without numeric values or error bars. Please include exact numbers, ideally in a table or as labels, and state whether the measurements are averaged over multiple runs.

Circularity Check

1 steps flagged

Figure 6(a)'s layer-wise gain analysis is partly self-confirmatory; the gain reward is never ablated alone.

specific steps
  1. fitted input called prediction [Section 5.5 (Figure 6(a)) vs Section 4.3 (Eq. for R_g^gain)]
    "As shown in Figure 6(a), HiLaR produces more differentiated layer-wise gains than LatentR3 and VRec, suggesting that its latent states capture complementary recommendation signals under the proposed coarse-to-fine supervision. ... We use the clipped layer-wise gains to construct the process reward: R_g^gain = Σ_k ρ_k clip(Δ_{g,k}, −b, b)."

    The plotted quantity in Fig. 6(a) is the layer-wise target log-probability gain Δ_{g,k}, the same quantity that the GRPO objective explicitly maximizes through R_g^gain. Thus comparing HiLaR's gains with those of LatentR3/VRec mostly verifies that HiLaR was optimized to increase this exact reward; it is not an independent behavioral finding. The 'complementary recommendation signals' are defined by the target-likelihood gain that is the training signal, so the analysis is partly self-confirmatory. It does not by itself establish that the states encode coarse-to-fine preference structure beyond the objective itself. (The headline ranking improvement over external baselines is separate and not circular.)

full rationale

The paper's central claim is an empirical accuracy improvement on held-out test splits of four Amazon datasets against external baselines (Table 1). That comparison is self-contained: HiLaR's scores are computed with the same catalog-constrained decoding used for the baselines, and the target item is used as supervision in quantization, SFT, and reward computation, which is standard supervised training rather than circular reasoning. I find no construction-level circularity in the headline result. The one partially circular element is the latent-state analysis in Section 5.5: Figure 6(a) reports 'layer-wise target log-probability gain,' which is exactly Δ_{g,k} that the GRPO reward R_g^gain maximizes in Section 4.3; observing that HiLaR has higher/more differentiated gains than models not trained on that reward is to a significant extent a check that the optimizer optimized its own reward, not independent evidence of 'complementary recommendation signals.' This is a partial self-confirmatory analysis, but it is not the basis of the main benchmark result. In addition, Figure 3 never ablates R_g^gain alone, so the specific contribution of the marginal-gain reward is not isolated; this is a missing control rather than a circular reduction. Score 3 reflects the partial self-confirmation without treating the external benchmark claim as circular.

Axiom & Free-Parameter Ledger

11 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities or forces; it is a new architecture and training scheme. The main assumptions are the informativeness of frozen LLM hidden states, the suitability of target-item-supervised quantization, and the safety of self-likelihood-based process rewards. Free parameters include several reward weights and hyperparameters, some of which are not reported in the main text.

free parameters (11)
  • Number of latent levels K = 4
    Chosen via sensitivity analysis (Section 5.3); represents the number of hierarchical preference granularities.
  • Codebook size = 256
    Fixed for quantizer in Section 4.1; not tuned in experiments.
  • Alignment weight λ_align = 0.10
    Selected via sensitivity analysis (Section 5.3).
  • Process reward weight λ_proc = 0.20
    Selected via sensitivity analysis (Section 5.3).
  • Rollout group size G = 6
    Selected via sensitivity analysis (Section 5.3).
  • Exploration noise per layer σ_k = (0.12, 0.08, 0.05, 0.025)
    Set to control exploration in GRPO (Section 4.3).
  • KL coefficient β = 0.05
    Set in Section 5.1.
  • Reward weights λ_hit, λ_prefix, λ_F1, λ_task, λ_pref, λ_eff
    Mentioned in Section 4.3 as tunable; details deferred to supplementary material, not included in the paper.
  • Layer gain weights ρ_k
    Control relative importance of each reasoning level in process reward; not specified in main text.
  • Gain clipping bound b
    Used in clip(Δ_{g,k}, -b, b); value not reported.
  • Temporal window split = equal-count
    Assumption on how to divide history into K windows; no alternative tested.
axioms (5)
  • domain assumption Frozen LLM hidden states encode sufficient user preference information to be quantized into meaningful coarse-to-fine codes.
    Section 4.1 uses a frozen LLM to encode the user history and then trains a quantizer on these embeddings; if the hidden states are not informative, the whole hierarchy collapses.
  • domain assumption The target item can serve as the finest-grained temporal supervision during quantizer training.
    Equation 6 sets Y_{u,K} = {i+} for the final quantization level, assuming the target item is the correct 'current intent' to supervise the last residual code.
  • domain assumption Marginal target-likelihood gain is a valid and safe process reward for credit assignment.
    Section 4.3 defines Δ_{g,k} as the difference in target token log-probability between latent prefixes; this assumes that improving target likelihood per layer is a good training signal and that optimizing it will not lead to reward hacking.
  • domain assumption A frozen collaborative model (SASRec) provides a useful preference signal for RL.
    Section 4.3 uses a frozen SASRec to compute preference reward R^g_pref; this assumes that collaborative filtering scores align with user preferences and can guide LLM generation.
  • standard math GRPO with group-relative advantage is applicable to continuous latent reasoning actions.
    The paper applies the GRPO objective (Eq. 16) to Gaussian policies over continuous states, a standard extension of DeepSeek-R1's method.

pith-pipeline@v1.3.0-daily-deepseek · 14048 in / 10149 out tokens · 106067 ms · 2026-08-01T01:44:44.364755+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) have shown strong potential for recommendation by leveraging their semantic understanding and contextual modeling capabilities. Recent studies further introduce reasoning mechanisms to improve user preference modeling. However, explicit natural-language reasoning incurs substantial inference overhead, whereas existing latent reasoning methods mainly focus on generating or verifying intermediate states, leaving their layer-wise preference roles and contributions insufficiently characterized. We propose HiLaR, a Hierarchical Latent Reasoning framework with layer-aware reinforcement optimization for LLM-based recommendation. HiLaR constructs temporal-guided hierarchical user preference representations, aligns them with multiple LLM latent reasoning states, and organizes the reasoning process from broad preferences to fine-grained current intents. To further optimize the reasoning trajectory, HiLaR combines final recommendation feedback with layer-aware process rewards derived from the marginal target-likelihood gain of each state. Experiments on four Amazon benchmark datasets show that HiLaR generally outperforms strong sequential, generative, and LLM-based recommendation baselines. Ablation and sensitivity analyses further verify the contribution of hierarchical representation learning, latent alignment, and process-level optimization. Our code is available in https://github.com/hupeiyu21/HiLaR.

Figures

Figures reproduced from arXiv: 2607.27760 by Jiahao Liang, Jiang Rong, Jia Wang, Peiyu Hu, Siying Gu, Weihai Lu, Yiying Xie, Yuntian Tang, ZhaoKai Luo, Zhiyong Wang, Zhuodong Liu.

Figure 1
Figure 1. Figure 1: Motivation for HiLaR. (a) Different latent states [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of HiLaR. (a) Hierarchical Latent Construction and Alignment encodes time-aware user histories into hierarchical preference representations and aligns them with the latent reasoning states of the LLM. (b) Hierarchical Reward￾Guided GRPO samples multiple latent reasoning trajectories and optimizes them with process-level and final recommendation rewards. The representation available after level 𝑘 i… view at source ↗
Figure 3
Figure 3. Figure 3: Ablation results on the CDs and Toys datasets using [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hyperparameter sensitivity of HiLaR on the CDs [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training dynamics of GRPO with final-only and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Case study of latent reasoning behavior. (a) Layer [PITH_FULL_IMAGE:figures/full_fig_p007_6.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

62 extracted references · 19 linked inside Pith

  1. [1]

    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

  2. [2]

    Keqin Bao, Jizhi Zhang, Wenjie Wang, Yang Zhang, Zhengyi Yang, Yanchen Luo, Chong Chen, Fuli Feng, and Qi Tian. 2025. A bi-step grounding paradigm for large language models in recommendation systems.ACM Transactions on Recommender Systems3, 4 (2025), 1–27

    2025

  3. [3]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Xinyue Huo, Chong Chen, and Fuli Feng

  4. [4]

    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

  5. [5]

    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

  6. [6]

    Yiwen Chen, Fuwei Zhang, Zehao Chen, Deqing Wang, Hehan Li, Peizhi Xu, Hanmeng Liu, Shuanglong Li, Xin Pei, Fuzhen Zhuang, et al . 2026. LASAR: Latent Adaptive Semantic Aligned Reasoning for Generative Recommendation. arXiv preprint arXiv:2605.10207(2026)

    Pith/arXiv arXiv 2026

  7. [7]

    Huixuan Chi, Hao Xu, Hao Fu, Mengya Liu, Mengdi Zhang, Yuji Yang, Qinfen Hao, and Wei Wu. 2022. Long short-term preference modeling for continuous- time sequential recommendation.arXiv preprint arXiv:2208.00593(2022)

    Pith/arXiv arXiv 2022

  8. [8]

    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

  9. [9]

    Yuntian Deng, Kiran Prasad, Roland Fernandez, Paul Smolensky, Vishrav Chaud- hary, and Stuart Shieber. 2023. Implicit chain of thought reasoning via knowledge distillation.arXiv preprint arXiv:2311.01460(2023)

    Pith/arXiv arXiv 2023

  10. [10]

    Tianqi Gao, Chengkai Huang, Zihan Wang, Cao Liu, Ke Zeng, and Lina Yao. 2026. Factorized Latent Reasoning for LLM-based Recommendation.arXiv preprint arXiv:2604.26760(2026)

    Pith/arXiv arXiv 2026

  11. [11]

    Yunfan Gao, Tao Sheng, You-Lin 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

  12. [12]

    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

  13. [13]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025)

    Pith/arXiv arXiv 2025

  14. [14]

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a contin- uous latent space.arXiv preprint arXiv:2412.06769(2024). Hu et al

    Pith/arXiv arXiv 2024

  15. [15]

    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. 843–852

    2018

  16. [16]

    Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating long semantic ids in parallel for recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 956–966

    2025

  17. [17]

    Peiyu Hu, Weihai Lu, Siying Gu, Elliott Wen, Changyu Zeng, Senzhang Wang, and Jia Wang. 2026. A Modular Survey for Semantic ID-Based Generative Recommendation. (2026)

    2026

  18. [18]

    Peiyu Hu, Wayne Lu, and Jia Wang. 2026. From ids to semantics: A generative framework for cross-domain recommendation with adaptive semantic tokeniza- tion. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 14874–14882

    2026

  19. [19]

    Jianchao Ji, Zelong Li, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Juntao Tan, and Yongfeng Zhang. 2024. Genrec: Large language model for generative recommen- dation. InEuropean Conference on Information Retrieval. Springer, 494–502

    2024

  20. [20]

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

    2018

  21. [21]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners.Advances in neural information processing systems35 (2022), 22199–22213

    2022

  22. [22]

    Xiaoyu Kong, Junguang Jiang, Bin Liu, Ziru Xu, Han Zhu, Jian Xu, Bo Zheng, Jiancan Wu, and Xiang Wang. 2026. Think before Recommendation: Autonomous Reasoning-enhanced Recommender.Advances in Neural Information Processing Systems38 (2026), 141209–141232

    2026

  23. [23]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive image generation using residual quantization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11523–11532

    2022

  24. [24]

    Chenhui Li and Weihai Lu. 2026. Decoding the Market’s Pulse: Context-Enriched Agentic Retrieval Augmented Generation for Predicting Post-Earnings Price Shocks. InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 3055–3073

    2026

  25. [25]

    Siqi Liang and Yudi Zhang. 2025. Generative Recommendation: A Survey of Models, Systems, and Industrial Advances. (2025)

    2025

  26. [26]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. Llara: Large language-recommendation assistant. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1785–1795

    2024

  27. [27]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s verify step by step. InInternational Conference on Learning Representations, Vol. 2024. 39578–39601

    2024

  28. [28]

    Xinyu Lin, Hanqing Zeng, Hanchao Yu, Yinglong Xia, Jiang Zhang, Aashu Singh, Fei Liu, Wenjie Wang, Fuli Feng, Tat-Seng Chua, et al . 2026. Verifiable Rea- soning for LLM-based Generative Recommendation. InProceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1118–1129

    2026

  29. [29]

    Enze Liu, Bowen Zheng, Xiaolei Wang, Wayne Xin Zhao, Jinpeng Wang, Sheng Chen, and Ji-Rong Wen. 2025. Lares: Latent reasoning for sequential recommen- dation.arXiv preprint arXiv:2505.16865(2025)

    Pith/arXiv arXiv 2025

  30. [30]

    Shijie Liu, Ruixin Ding, Weihai Lu, Jun Wang, Mo Yu, Xiaoming Shi, and Wei Zhang. 2025. Coherency improved explainable recommendation via large lan- guage model. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 12201–12209

    2025

  31. [31]

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

  32. [32]

    Wayne Lu and Xiaoxi Cui. 2026. DEALT: LLM-driven Diversity-Enhanced Data Augmentation for Long-Tail Text Classification. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 32338–32346

    2026

  33. [33]

    Weihai Lu, Zhejun Zhao, Yanshu Li, and Huan He. 2026. MM-StanceDet: Retrieval-Augmented Multi-modal Multi-agent Stance Detection.arXiv preprint arXiv:2604.27934(2026)

    Pith/arXiv arXiv 2026

  34. [34]

    Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Meiqi Guo, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, et al. 2024. Improve mathematical reasoning in language models by automated process supervision.arXiv preprint arXiv:2406.06592(2024)

    Pith/arXiv arXiv 2024

  35. [35]

    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

  36. [36]

    Maxwell Nye, Anders Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, Charles Sutton, and Augustus Odena. 2021. Show Your Work: Scratch- pads for Intermediate Computation with Language Models.arXiv preprint arXiv:2112.00114(2021)

    Pith/arXiv arXiv 2021

  37. [37]

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

  38. [38]

    Leheng Sheng, An Zhang, Yi Zhang, Yuxin Chen, Xiang Wang, and Tat-Seng Chua. 2025. Language representations can be what recommenders need: Findings and potentials. Ininternational Conference on Learning Representations, Vol. 2025. 91632–91658

    2025

  39. [39]

    Jiakai Tang, Xu Chen, Wen Chen, Jian Wu, Yuning Jiang, and Bo Zheng. 2026. Parallel Latent Reasoning for Sequential Recommendation.arXiv preprint arXiv:2601.03153(2026)

    arXiv 2026

  40. [40]

    OneRec Team, Biao Yang, Boyang Ding, Chenglong Chu, Dunju Zang, Fei Pan, Han Li, Hao Jiang, Honghui Bao, Huanjie Wang, et al. 2026. OneReason Technical Report.arXiv preprint arXiv:2606.06260(2026)

    Pith/arXiv arXiv 2026

  41. [41]

    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

    2024

  42. [42]

    Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning.Advances in neural information processing systems30 (2017)

    2017

  43. [43]

    Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. 2024. Math-shepherd: Verify and reinforce llms step-by- step without human annotations. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 9426–9439

    2024

  44. [44]

    Xinfeng Wang, Jin Cui, Yoshimi Suzuki, and Fumiyo Fukumoto. 2024. Rdrec: Rationale distillation for llm-based recommendation. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 65–74

    2024

  45. [45]

    Zhang, Qing Cui, Longfei Li, Jun Zhou, and Sheng Li

    Yan Wang, Zhixuan Chu, Xin Ouyang, Simeng Wang, Hongyan Hao, Yue Shen, Jinjie Gu, Siqiao Xue, James Y. Zhang, Qing Cui, Longfei Li, Jun Zhou, and Sheng Li. 2023. Enhancing Recommender Systems with Large Language Model Reasoning Graphs.arXiv preprint arXiv:2308.10835(2023)

    Pith/arXiv arXiv 2023

  46. [46]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reason- ing in large language models.Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  47. [47]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. Llmrec: Large language models with graph augmentation for recommendation. InProceedings of the 17th ACM international conference on web search and data mining. 806–815

    2024

  48. [48]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Cheng- peng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfen...

    Pith/arXiv arXiv 2024

  49. [49]

    Zhenrui Yue, Sara Rabhi, Gabriel de Souza Pereira Moreira, Dong Wang, and Even Oldridge. 2023. Llamarec: Two-stage recommendation using large language models for ranking.arXiv preprint arXiv:2311.02089(2023)

    Pith/arXiv arXiv 2023

  50. [50]

    Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, and Noah D. Goodman. 2024. Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking.arXiv preprint arXiv:2403.09629(2024)

    Pith/arXiv arXiv 2024

  51. [51]

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji- Rong Wen. 2026. Recommendation as instruction following: A large language model empowered recommendation approach.ACM Transactions on Information Systems43, 5 (2026), 1–37

    2026

  52. [52]

    Yang Zhang, Keqin Bao, Ming Yan, Wenjie Wang, Fuli Feng, and Xiangnan He

  53. [53]

    Yang Zhang, Fuli Feng, Jizhi Zhang, Keqin Bao, Qifan Wang, and Xiangnan He

  54. [54]

    Yang Zhang, Wenxin Xu, Xiaoyan Zhao, Wenjie Wang, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2025. Reinforced latent reasoning for llm-based recommen- dation.arXiv preprint arXiv:2505.19092(2025)

    arXiv 2025

  55. [55]

    Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. 2025. Processbench: Identifying process errors in mathematical reasoning. InProceedings of the 63rd Annual Hierarchical Latent Reasoning for LLM-based Recommendation Meeting of the Association for Computational Linguistics (Volume 1: Lon...

    2025

  56. [56]

    InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Text-like encoding of collaborative information in large language models for recommendation. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 9181–9191

  57. [58]

    Collm: Integrating collaborative embeddings into large language models for recommendation.IEEE Transactions on Knowledge and Data Engineering (2025)

    2025

  58. [61]

    Rui-Jie Zhu, Tianhao Peng, Tianhao Cheng, Xingwei Qu, Jinfa Huang, Dawei Zhu, Hao Wang, Kaiwen Xue, Xuanliang Zhang, Yong Shan, Tianle Cai, Taylor Kergan, Assel Kembay, Andrew Smith, Chenghua Lin, Binh Nguyen, Yuqi Pan, Yuhong Chou, Zefan Cai, Zhenhe Wu, Yongchi Zhao, Tianyu Liu, Jian Yang, Wangchunshu Zhou, Chujie Zheng, Chongxuan Li, Yuyin Zhou, Zhoujun...

  59. [62]

    arXiv:2507.06203 [cs.CL] https://arxiv.org/ abs/2507.06203

    A Survey on Latent Reasoning. arXiv:2507.06203 [cs.CL] https://arxiv.org/ abs/2507.06203

    Pith/arXiv arXiv

  60. [2023]

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

    2023

  61. [2024]

    InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

    Decoding matters: Addressing amplification bias and homogeneity issue in recommendations for large language models. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 10540–10552

    2024

  62. [2025]

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

    arXiv 2025