Pith. sign in

REVIEW 4 major objections 6 minor 65 references

EvoReason claims that self-evolving primitive-guided on-policy distillation transfers explicit chain-of-thought supervision into compact latent tokens more effectively than static distillation, improving recommendation accuracy while keepin

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

2026-08-03 15:25 UTC pith:LZALSCOO

load-bearing objection Plausible empirical system paper with moderate novelty, but the self-evolving distillation mechanism is not cleanly isolated and the statistical support is thin; worth a serious referee. the 4 major comments →

arxiv 2607.29010 v1 pith:LZALSCOO submitted 2026-07-31 cs.IR

EvoReason: Self-Evolving Reasoning Primitive-Guided On-Policy Distillation for Latent Reasoning in Generative Recommendation

Zhuang Zhuang , Zhipeng Wei , Rongfeng Guo , Shijie Li , Peng Zhao , Jie Chen , Fei Pan This is my paper
classification cs.IR
keywords generative recommendationlatent reasoningchain-of-thought distillationon-policy distillationreasoning primitivesself-evolving supervisionsequential recommendationcold-start 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.

The paper argues that existing latent-reasoning recommenders distill raw chain-of-thought text into hidden tokens, which is inefficient because those traces entangle reusable reasoning behaviors with redundant, unstable linguistic variation. It proposes EvoReason, which first extracts reusable function-like reasoning primitives from agentic trajectories, then uses them to generate structured teacher reasoning, and finally runs a self-evolving loop: the student's latent outputs are refined by a stop-gradient teacher pass, distilled back under a confidence gate, and aligned through KV states while the primitive library updates. The authors report that this loop outperforms prior generative and latent-reasoning methods, with the largest gains over the strongest latent baseline around +17.9% Recall@5 and +27.4% NDCG@5 on Beauty, plus positive one-week online A/B gains. If the mechanism holds, it shows that the content and timing of reasoning supervision, not just its presence, is what lets compact latent tokens inherit the benefits of explicit chain-of-thought.

Core claim

On its own terms, the paper's central claim is that raw CoT trajectories are the wrong unit of supervision for latent reasoning. EvoReason replaces them with a tripartite mechanism: a library of reasoning primitives acts as pseudo-tools that constrain teacher generation; the teacher refines the student's own on-policy latent-conditioned trajectory through a stop-gradient forward pass, generating a corrected trajectory that is distilled back with a confidence gate and aligned via compressed KV representations; and the primitive library evolves with the student's outcomes. The result is a closed loop in which supervision and latent policy co-adapt, yielding better alignment between explicit re

What carries the argument

The load-bearing object is the self-evolving reasoning-primitive library: function-like pseudo-tools induced from agentic trajectories, each with input parameters, a structured reasoning procedure, and output format constraints. The teacher invokes these primitives to generate structured CoT supervision; the student distills it through confidence-gated on-policy distillation (a weighted objective with a teacher-induced confidence gain) and KV-state alignment. The loop closes when the primitive library updates based on latent reasoning outcomes, so supervision tracks the student's evolving capability.

Load-bearing premise

The teacher that corrects the student's latent trajectory is the student's own parameters with a stop-gradient and extra context (Eq. 17–18), so the whole self-evolving loop rests on the assumption that this augmented self-pass produces genuinely corrected reasoning rather than just re-predicting the student's existing biases.

What would settle it

Check whether the stop-gradient teacher refinement actually changes the student's final predictions on examples where the initial trajectory is wrong. If an ablation that replaces the refined context with the original context (or with an independent teacher's CoT) performs the same or better, or if the confidence-gated teacher gain is uncorrelated with whether the refinement fixes the top-1 error, the claim that self-evolution drives the transfer is falsified.

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

If this is right

  • Latent reasoning can absorb explicit chain-of-thought benefits without paying inference-time text generation: reported per-sample latency stays near a plain generative retrieval model instead of the order-of-magnitude jump of explicit CoT methods.
  • Self-evolving supervision, not just on-policy training, is the driver: ablations show removing primitive structure, removing evolution, or replacing OPD with static distillation each lowers performance.
  • The gains concentrate where semantic supervision matters most: cold-start users show a larger relative improvement than very active users.
  • The learned reasoning is more transferable: cross-dataset generalization degrades less than the static-distillation baseline.
  • In a production recall deployment, replacing the baseline recall model with EvoReason lifts downstream business metrics, implying the improved candidate set is the cause.

Where Pith is reading between the lines

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

  • The same self-evolving primitive-guided distillation loop is task-agnostic in principle: any expensive teacher reasoning (math, planning, tool use) could be compressed into latent tokens for a cheap student without an external teacher at inference, provided the on-policy refinement genuinely corrects errors.
  • Because teacher and student share parameters, the loop is vulnerable to self-confirmation; a natural extension would insert an independent verifier or reward model to decide when the refined trajectory is actually better, instead of trusting the confidence gate.
  • The primitive library size K is presented as a fixed hyperparameter, but the concave sensitivity pattern suggests the right granularity of reasoning decomposition is data-dependent and could be learned or adapted per task.
  • One could make the mechanism measurable: track whether the confidence gate's accepted refinements correlate with eventual label accuracy; if not, the 'refinement' is cosmetic rather than corrective.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes EvoReason, a generative recommendation framework that trains latent reasoning tokens under supervision derived from reasoning primitives extracted from agentic recommendation trajectories. Three stages are presented: primitive discovery from ReAct-style trajectories, primitive-aware teacher SFT, and self-evolving on-policy latent reasoning distillation. The final stage combines GRPO-style reinforcement learning on the student's own rollouts with a confidence-gated distillation loss that uses the same underlying parameters as the teacher via a stop-gradient forward pass, plus a KV-alignment regularizer. Experiments span two public Amazon datasets and one proprietary industrial dataset, with offline metrics, ablations, an online A/B test, and additional analyses of cold-start performance, cross-domain transfer, latency, and hyperparameter sensitivity. The main claimed contribution is that self-evolving primitive-guided on-policy distillation transfers explicit reasoning into latent tokens more effectively than static CoT distillation, yielding consistent gains over latent-reasoning baselines such as LASAR.

Significance. If the claimed effects hold, EvoReason would be a meaningful advance in latent reasoning for generative recommendation: it addresses a real limitation of prior work (raw CoT trajectories are redundant and static) and provides a practical mechanism for aligning explicit reasoning supervision with the student's latent reasoning process. The paper's strengths include a clearly specified training pipeline, a thorough set of comparisons across multiple baseline families, online A/B evidence, and a useful sensitivity analysis of the primitive-library size. The central empirical claim, however, rests on comparisons without any reported variance or significance testing, and the key novelty — self-evolving on-policy distillation with a shared-parameter teacher — is not isolated in the ablations in a way that rules out confounds from GRPO or from primitive prompting alone. The significance is therefore conditional on resolving these evidential gaps.

major comments (4)
  1. [§4.4.2, Eqs. (17)–(22)] The teacher in the OPD stage is a stop-gradient forward pass of the same student parameters G_theta, and y_r is generated by that same G_theta conditioned on the student's own latent states and trajectory. The paper provides no evidence that y_r actually corrects errors in y_o — no success-rate comparison of refined vs. original trajectories, no external verifier, and no ablation with an independently trained teacher. The confidence gate g_t in Eq. (21) is computed from two log-likelihoods of the same parameter vector under different contexts, so it measures self-consistency, not teacher correctness. As written, the OPD loss (Eq. 22) is self-conditioned likelihood maximization, and the claimed 'reasoning transfer' is not established unless the authors show that y_r is more accurate or otherwise contains information not already in y_o. Please add a direct empirical test, e.g., compare y_r
  2. [§5.2, Table 2; §5.4, Table 4] The abstract and Observation 3 claim that EvoReason 'consistently outperforms' baselines, but no confidence intervals, standard deviations, or significance tests are reported for any offline metric. For a system paper whose central assertion is empirical superiority, this is a load-bearing gap. The online A/B test (Table 4) reports only point estimates (+8.11% ADVV, +6.23% revenue) with no confidence intervals, p-values, or description of bucketing variance. Please provide at least three seeds with standard deviations for offline results and statistical significance for the online metrics; otherwise the phrase 'consistently' is unsupported.
  3. [§5.3, Table 3] The ablation study is run only on Beauty. The paper's main claim spans three datasets, including a proprietary industrial dataset, but the contribution of each component is tested on a single public benchmark. Moreover, the 'w/o OPD' variant replaces OPD with 'conventional knowledge distillation' — it is not specified whether this variant still includes the GRPO loss (Eq. 16) and the KV alignment. If GRPO is still active, the comparison does not isolate on-policy distillation from RL; if GRPO is also removed, the ablation conflates three changes. Please run ablations on at least one additional dataset and clearly state which loss terms are active in each ablation.
  4. [§5.1, Table 1 and Section 5.4] The 'Industrial' dataset is listed in Table 2 but never described in Table 1, and the online A/B test does not describe the dataset size, user population, candidate space, or how the recall model interfaces with downstream ranking. Since the industrial and online results are part of the paper's support for 'consistently outperforms', the absence of this information makes those results unverifiable. Please add dataset statistics and a more detailed description of the production setup.
minor comments (6)
  1. [§5.1, Baselines] The baseline list spells 'LASAR' as 'LASAE' in the final sentence of the Baselines paragraph; the reference [4] is to LASAR. Please correct the name.
  2. [Appendix A.2, Table 7] The text says 'OneReason introduces only a marginal increase in latency', but the method being evaluated is EvoReason. The table's 'Tiger' should be 'TIGER'. Please align terminology.
  3. [Eq. (2) vs. Eq. (22)] The OPD objective in Eq. (2) is a KL divergence with the teacher, but Eq. (22) is a weighted MLE using sg(l^T_t). This is not necessarily wrong, but the relationship between the two should be explained, especially why the target is the teacher's log-likelihood rather than the teacher's distribution.
  4. [Eqs. (13)–(14)] The notation x_eot is used without definition; it appears to denote a special token and should be introduced explicitly.
  5. [§5.5, Figure 3] The success-rate curves in Figure 3(a) show values around 0.15–0.21, but the offline Recall@5 values in Table 2 are much lower (0.07). Please clarify what 'success rate' means and why it differs from the reported metrics.
  6. [§4.2.3, Eq. (7)] The Update function for the primitive library is described only at a high level. Since self-evolving primitives are a core contribution, please provide more detail on how new primitives are induced, how they are merged with existing ones, and how redundancy is controlled.

Circularity Check

0 steps flagged

No significant circularity: the paper's claims are empirical benchmark comparisons with ablations, and the shared-parameter teacher is an explicitly acknowledged self-distillation design rather than a hidden definitional reduction.

full rationale

EvoReason is an empirical systems paper. The load-bearing assertions are benchmark comparisons (Tables 2, 3, 4, 5) evaluated on held-out data or online traffic; they are not derived from an equation that contains the answer. Stage 1 seeds the primitive library from externally generated QwQ-32B agentic trajectories, so the teacher's extra conditioning is not exclusively the student's own output. The shared-backbone teacher in Section 4.4.2 is explicitly described: "The teacher shares the same backbone with the student and performs a stop-gradient forward pass" and "y_r = G_theta(x_T, P)" (Eqs. 17-18). This is an acknowledged self-distillation/self-refinement loop, not a hidden prediction: Eq. 22 optimizes a confidence-gated log-likelihood of the student's own tokens, and y_r is generated under a different, primitive-augmented context. The concern that y_r is not independently verified—no external verifier or success-rate comparison between y_r and y_o—bears on robustness and attribution (e.g., GRPO vs. OPD), not on definitional circularity. Self-citations such as [34] and [55] are contextual and not load-bearing. There is no uniqueness theorem or prior-work ansatz imported to force the architecture. Thus no step reduces a claimed result to its own inputs by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

No new physical entities are postulated; the method rests on software constructs (primitive library, latent slots). Free hyperparameters K, M, T, beta, and loss weights are either tuned or unreported, so the reported numbers are conditional on unstated settings. The key domain assumptions concern SID quality, primitive quality, self-distillation validity, and KV-space alignment.

free parameters (7)
  • Primitive library size K = 5 (default; best of {3, 5, 10})
    Chosen via sensitivity sweep in Table 8: Beauty R@5 is 0.0703 for K=3, 0.0724 for K=5, 0.0718 for K=10.
  • Number of latent tokens M
    Introduced in Eq 1 and Eq 12 but never specified in the paper; affects latent reasoning capacity.
  • Latent refinement iterations T
    Iterations in Eq 13; never reported.
  • Confidence-gate sharpness beta
    Eq 21 controls the gating of teacher signal; not reported.
  • OPD loss weight lambda_OPD
    Eq 24 combines RL, OPD, and KV losses; weights are not reported.
  • KV alignment weight lambda_KV
    Eq 24; not reported.
  • GRPO rollout group size N
    Used to compute group-relative advantage in Eq 15; not reported.
axioms (5)
  • domain assumption Semantic ID item encoding is a sufficient target space for generative recommendation.
    Section 3.1 adopts OneRec-Think's four-level SID hierarchy; if SIDs lose information, all downstream reasoning cannot recover it.
  • domain assumption QwQ-32B agentic trajectories are high-quality and contain reusable reasoning primitives.
    Section 4.2.1-4.2.2; the paper does not independently validate primitive quality, and the induction step is unspecified.
  • domain assumption Teacher/student parameter sharing with stop-gradient provides a valid supervision signal.
    Section 4.4.2, Eqs 17-22; if the self-generated refinement is no better than the student's own output, the loop is self-confirming.
  • domain assumption KV-space L1 alignment between compressed teacher reasoning and student latent states is a meaningful semantic alignment objective.
    Eq 23; relies on R-KV compression retaining reasoning information.
  • standard math GRPO clipped objective is a valid policy-gradient estimator for SID token-level rewards.
    Eqs 15-16; standard RL estimator imported from the literature.
invented entities (2)
  • Reasoning primitive library (pseudo-tools) no independent evidence
    purpose: Structures teacher CoT generation into reusable reasoning behaviors; evolves during training.
    No external or formal validation; measured only through downstream task performance in this paper.
  • Primitive-aware latent slots no independent evidence
    purpose: Initializes latent reasoning states from primitive slots in Eq 12.
    Architectural invention; no independent evidence that it captures primitive semantics.

pith-pipeline@v1.3.0-daily-deepseek · 16267 in / 14945 out tokens · 146513 ms · 2026-08-03T15:25:25.149776+00:00 · methodology

0 comments
read the original abstract

Generative recommendation benefits from reasoning-enhanced inference, and latent reasoning offers an efficient paradigm by encoding intermediate reasoning processes into compact continuous representations for latency-sensitive deployment. Despite its efficiency, existing latent reasoning approaches typically rely on directly distilling raw chain-of-thought (CoT) trajectories into latent representations, assuming that textual reasoning traces provide sufficient supervision. However, recommendation reasoning trajectories contain diverse reasoning processes with redundant expressions and unstable reasoning paths, making raw CoT supervision suboptimal for learning transferable latent reasoning representations. To address this challenge, we propose EvoReason, a self-evolving latent reasoning framework that adaptively aligns explicit reasoning supervision with the student's latent reasoning space through primitive-guided on-policy distillation. First, EvoReason extracts reusable reasoning primitives from high-quality agentic recommendation trajectories, where each primitive captures an essential reasoning behavior and serves as a pseudo-tool for structured teacher reasoning. Then, based on these primitives, we equip the teacher with primitive-aware reasoning capabilities, enabling it to generate structured CoT supervision with reduced redundancy and improved consistency. Finally, during latent reasoning optimization, EvoReason introduces a self-evolving on-policy distillation mechanism, where the primitive-guided reasoning process evolves according to the student's latent reasoning outcomes. Through this closed-loop co-evolution, policy updates continuously improve latent reasoning behaviors is refined according to the resulting latent reasoning outcomes, enabling progressively better-aligned CoT supervision and more effective reasoning transfer.

Figures

Figures reproduced from arXiv: 2607.29010 by Fei Pan, Jie Chen, Peng Zhao, Rongfeng Guo, Shijie Li, Zhipeng Wei, Zhuang Zhuang.

Figure 1
Figure 1. Figure 1: Framework comparison between (1) conventional [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of EvoReason. Stage 1 discovers reusable reasoning primitives from high-quality trajectories and constructs a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training Dynamics of Self-Evolution on the Beauty [PITH_FULL_IMAGE:figures/full_fig_p008_3.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

65 extracted references · 28 linked inside Pith

  1. [1]

    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

  2. [2]

    Zefan Cai, Wen Xiao, Hanshi Sun, Yikai Zhang, Ke Wan, Yucheng Li, Yeyang Zhou, Li-Wen Chang, Jiuxiang Gu, Zhen Dong, et al. 2026. R-kv: Redundancy-aware kv cache compression for reasoning models.Advances in neural information processing systems38 (2026), 60980–61005

    2026

  3. [3]

    Weixin Chen, Yuhan Zhao, Jingyuan Huang, Zihe Ye, Mingxuan Ju, Tong Zhao, Neil Shah, Li Chen, and Yongfeng Zhang. 2026. Memrec: Collaborative memory- augmented agentic recommender system. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 44515–44544

    2026

  4. [4]

    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

  5. [5]

    Yuntian Deng, Yejin Choi, and Stuart Shieber. 2024. From explicit cot to implicit cot: Learning to internalize cot step by step.arXiv preprint arXiv:2405.14838 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    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

  7. [7]

    Hao Fan, Mengyi Zhu, Yanrong Hu, Hailin Feng, Zhijie He, Hongjiu Liu, and Qingyang Liu. 2025. TiM4Rec: An efficient sequential recommendation model based on time-aware structured state space duality model.Neurocomputing (2025), 131270

    2025

  8. [8]

    Zhen Fang, Wenxuan Huang, Yu Zeng, Yiming Zhao, Shuang Chen, Kaituo Feng, Yunlong Lin, Lin Chen, Zehui Chen, Shaosheng Cao, et al. 2026. Flow-opd: On- policy distillation for flow matching models.arXiv preprint arXiv:2605.08063 (2026)

    Pith/arXiv arXiv 2026

  9. [9]

    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

  10. [10]

    Hao Gu, Rui Zhong, Yu Xia, Wei Yang, Chi Lu, Peng Jiang, and Kun Gai. 2025. R 4ec: A reasoning, reflection, and refinement framework for recommendation sys- tems. InProceedings of the Nineteenth ACM Conference on Recommender Systems. 411–421

    2025

  11. [11]

    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 in- centivizes reasoning in LLMs through reinforcement learning.Nature645, 8081 (2025), 633–638

    2025

  12. [12]

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

    Pith/arXiv arXiv 2024

  13. [13]

    Guoxiong He, Zhuang Zhuang, Chen Li, Hongliang Qiao, Shanshan Feng, Hao Wang, Fan Li, and Gao Cong. [n. d.]. A Comprehensive Survey of Generative AI for POI Recommendation. ([n. d.])

  14. [14]

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

  15. [15]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531(2015)

    Pith/arXiv arXiv 2015

  16. [16]

    Wenjin Hou, Shangpin Peng, Weinong Wang, Zheng Ruan, Yue Zhang, Zhenglin Zhou, Mingqi Gao, Yifei Chen, Kaiqi Wang, Hongming Yang, et al. 2026. Uni-opd: Unifying on-policy distillation with a dual-perspective recipe.arXiv preprint arXiv:2605.03677(2026)

    Pith/arXiv arXiv 2026

  17. [17]

    Dengyang Jiang, Xin Jin, Dongyang Liu, Zanyi Wang, Mingzhe Zheng, Ruoyi Du, Xiangpeng Yang, Qilong Wu, Zhen Li, Peng Gao, et al. 2026. D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models.arXiv preprint arXiv:2605.05204(2026)

    Pith/arXiv arXiv 2026

  18. [18]

    Li Jiang, Haoran Xu, Yichuan Ding, and Amy Zhang. 2026. Trajectory-Refined Distillation.arXiv preprint arXiv:2606.08432(2026)

    Pith/arXiv arXiv 2026

  19. [19]

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

    2018

  20. [20]

    Anna Kuzina, Maciej Pioro, Paul N Whatmough, and Babak Ehteshami Bejnordi

  21. [21]

    Zhihan Lei, Jiarui Yan, Joshua Momo, and William W Cohen. 2026. Inducing Reasoning Primitives from Agent Traces.arXiv preprint arXiv:2606.02994(2026)

    Pith/arXiv arXiv 2026

  22. [22]

    Baohao Liao, Hanze Dong, Christof Monz, Xinxing Xu, Li Dong, and Furu Wei

  23. [23]

    Zizhuo Lin, Quanling Liu, Jinsheng Quan, Chao Zhang, Yifan Zhu, Xing Shi, Jingtao Xu, Zhihui Li, and Yawei Luo. 2026. Same Evidence, Different Answers: Conference’17, July 2017, Washington, DC, USA Zhuang Zhuang, Zhipeng Wei, Rongfeng Guo, Shijie Li, Peng Zhao, Jie Chen, and Fei Pan Canonical-Context On-Policy Distillation for Multi-Turn Language Models.a...

    Pith/arXiv arXiv 2026

  24. [24]

    Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee

  25. [25]

    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

  26. [26]

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

  27. [27]

    Bryan D Payne, Martim Carbone, Monirul Sharif, and Wenke Lee. 2008. Lares: An architecture for secure active monitoring using virtualization. In2008 IEEE Symposium on Security and Privacy (sp 2008). IEEE, 233–247

    2008

  28. [28]

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

  29. [29]

    Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. 2025. Codi: Compressing chain-of-thought into continuous space via self-distillation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 677–693

    2025

  30. [30]

    Mingyang Song and Mao Zheng. 2026. A survey of on-policy distillation for large language models.arXiv preprint arXiv:2604.00626(2026)

    Pith/arXiv arXiv 2026

  31. [31]

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

    arXiv 2025

  32. [32]

    Wenhui Tan, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Ruihua Song, and Jian Luan

  33. [33]

    Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Jian Wu, and Yuning Jiang. 2026. Think before recommend: Unleashing the latent reasoning power for sequential recommendation.IEEE Transactions on Knowledge and Data Engineering(2026)

    2026

  34. [34]

    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

  35. [35]

    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

  36. [36]

    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 reasoning in large language models.Advances in neural information processing systems35 (2022), 24824–24837

    2022

  37. [37]

    Lena Strobl, William Merrill, Gail Weiss, David Chiang, and Dana Angluin. 2024. What formal languages can transformers express? a survey.Transactions of the Association for Computational Linguistics12 (2024), 543–561

    2024

  38. [38]

    Haibo Xing, Hao Deng, Yucheng Mao, Lingyu Mu, Jinxin Hu, Yi Xu, Hao Zhang, Jiahao Wang, Shizhun Wang, Yu Zhang, et al. 2025. Reg4rec: Reasoning-enhanced generative model for large-scale recommendation systems.arXiv preprint arXiv:2508.15308(2025)

    arXiv 2025

  39. [39]

    Think silently, think fast: Dynamic latent compression of llm reasoning chains.Advances in Neural Information Processing Systems38 (2026), 4646–4668

    2026

  40. [40]

    Fengli Xu, Qianyue Hao, Chenyang Shao, Zefang Zong, Yu Li, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, et al . 2025. Toward large reasoning models: A survey of reinforced reasoning with large language models.Patterns6, 10 (2025)

    2025

  41. [41]

    Shuo Yang, Jinyang Wu, Zhengxi Lu, Yuhao Shen, Fan Zhang, Lang Feng, Shuai Zhang, Haoran Luo, Zheng Lian, Zhengqi Wen, et al. 2026. OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning.arXiv preprint arXiv:2606.26790 (2026)

    Pith/arXiv arXiv 2026

  42. [42]

    Wenkai Yang, Weijie Liu, Ruobing Xie, Kai Yang, Saiyong Yang, and Yankai Lin

  43. [43]

    Runyang You, Yongqi Li, Xinyu Lin, Xin Zhang, Wenjie Wang, Wenjie Li, and Liqiang Nie. 2026. R2ec: Towards Large Recommender Models with Reasoning. Advances in Neural Information Processing Systems38 (2026), 62376–62405

    2026

  44. [44]

    Jinyang Wu, Shuo Yang, Zhengxi Lu, Fan Zhang, Yuhao Shen, Lang Feng, Haoran Luo, Zheng Lian, Shuai Zhang, Zhengqi Wen, et al. 2026. SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning.arXiv preprint arXiv:2607.14777(2026)

    Pith/arXiv arXiv 2026

  45. [45]

    Haobo Zhang, Yutao Zhu, Kelong Mao, Tianhao Li, and Zhicheng Dou. 2026. RecThinker: An Agentic Framework for Tool-Augmented Reasoning in Recom- mendation.arXiv preprint arXiv:2603.09843(2026)

    arXiv 2026

  46. [46]

    Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, et al. 2026. Deepseek- v4: Towards highly efficient million-token context intelligence.arXiv preprint arXiv:2606.19348(2026)

    arXiv 2026

  47. [47]

    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

  48. [48]

    Qingfei Zhao, Huan Song, Shuyu Tian, Jiawei Shao, and Xuelong Li. 2026. Prefix- Guided On-Policy Distillation: Mining Golden Trajectories from Rollouts.arXiv preprint arXiv:2606.21994(2026)

    Pith/arXiv arXiv 2026

  49. [49]

    Zhengyang Zhao, Lu Ma, and Wentao Zhang. 2026. Training with Harnesses: On-Policy Harness Self-Distillation for Complex Reasoning.arXiv preprint arXiv:2605.08741(2026)

    Pith/arXiv arXiv 2026

  50. [50]

    Learning beyond teacher: Generalized on-policy distillation with reward extrapolation.arXiv preprint arXiv:2602.12125(2026)

    Pith/arXiv arXiv 2026

  51. [51]

    Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Shiyao Wang, et al. 2025. Onerec technical report.arXiv preprint arXiv:2506.13695(2025)

    arXiv 2025

  52. [52]

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

    Pith/arXiv arXiv 2024

  53. [53]

    Zhuang Zhuang, Shanshan Feng, Hangwei Qian, Mingqi Yang, Heng Qi, Yanming Shen, and Baocai Yin. 2026. Think2Go: Generative Next POI Recommendation with LLM Reasoning. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2112–2123

    2026

  54. [54]

    Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2024. Agentcf: Collaborative learning with autonomous language agents for recommender systems. InProceedings of the ACM Web Conference 2024. 3679–3689

    2024

  55. [55]

    Zhuang Zhuang, Zhipeng Wei, Ji Dai, Jie Chen, Fei Pan, Peng Jiang, and Kun Gai

  56. [56]

    Zhuang Zhuang, Haitao Yuan, Shanshan Feng, Heng Qi, Yanming Shen, and Baocai Yin. 2025. MGSTDN: Multi-Granularity Spatial-Temporal Diffusion Net- work for Next POI Recommendation. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 4560–4570. EvoReason: Self-Evolving Reasoning Primitive-Guided On-Policy Distilla...

    2025

  57. [58]

    Bowen Zheng, Chao Yi, Dian Chen, Gaoyang Guo, Han Zhu, Jiakai Tang, Jian Wu, Mao Zhang, Wen Chen, Yifan Lu, et al. 2026. RecGPT-V3 Technical Report. arXiv preprint arXiv:2607.15591(2026)

    Pith/arXiv arXiv 2026

  58. [60]

    Yuhang Zhou, Kai Zheng, Haoling Li, Dengyun Peng, Can Xu, and Jingjing Chen. 2026. TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training.arXiv preprint arXiv:2607.05804(2026)

    Pith/arXiv arXiv 2026

  59. [62]

    Zhuang Zhuang, Tianxin Wei, Lingbo Liu, Heng Qi, Yanming Shen, and Baocai Yin. 2024. TAU: trajectory data augmentation with uncertainty for next POI recommendation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 22565–22573

    2024

  60. [64]

    SinkRec: Mitigating Semantic State Sink in Long Sequence Recommendation with Memory-Conditioned Gated Delta Networks.arXiv preprint arXiv:2606.09888 (2026)

    Pith/arXiv arXiv 2026

  61. [2015]

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

    Pith/arXiv arXiv 2015

  62. [2023]

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

    2023

  63. [2024]

    Mamba4rec: Towards efficient sequential recommendation with selective state space models.arXiv preprint arXiv:2403.03900(2024)

    Pith/arXiv arXiv 2024

  64. [2025]

    KaVa: Latent Reasoning via Compressed KV-Cache Distillation.arXiv preprint arXiv:2510.02312(2025)

    Pith/arXiv arXiv 2025

  65. [2026]

    Multi-Turn On-Policy Distillation with Prefix Replay.arXiv preprint arXiv:2607.04763(2026)

    Pith/arXiv arXiv 2026