Pith. sign in

REVIEW 4 major objections 5 minor 17 references

WhisperRec compresses explicit chain-of-thought reasoning into three latent tokens, exceeding verbose reasoning on accuracy while running at over ten times the throughput.

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 12:07 UTC pith:AQZMWNBV

load-bearing objection A promising latent-reasoning architecture for FRMs, but the headline 'latent beats explicit' claim is built on the weakest baseline and needs a stronger comparison. the 4 major comments →

arxiv 2607.26621 v2 pith:AQZMWNBV submitted 2026-07-29 cs.IR cs.AI

WhisperRec: Latent Reasoning for Efficient Foundation Recommendation Models

Hao Jiang , Peiru Du , Pengfei Yao , Mengting Li , Siyuan Lou , Kuo Cai , Sheng Yu , Qiang Luo
show 5 more authors
Jian Liang Ruiming Tang Fei Pan Peng Jiang Wenwu Ou
This is my paper
classification cs.IR cs.AI
keywords latent reasoningchain-of-thoughtfoundation recommendation modelsgenerative retrievalsemantic IDsknowledge distillationinference efficiencymulti-view CoT
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 the decision-relevant content of explicit chain-of-thought reasoning for recommendation can be distilled into a small set of learnable latent tokens, so the model reasons in latent space and then answers. This Latent-Reason-then-Answer paradigm avoids autoregressive rationale generation while preserving reasoning benefits. On an industrial-scale dataset and a public benchmark, WhisperRec improves SID@64 by 17.44% over an explicit-CoT Think baseline and 9.33% over a No-Think variant, while achieving more than ten times the online inference throughput. The authors argue that this shows reasoning helps recommendation by improving decision-relevant representations, not by generating long natural-language rationales.

Core claim

The central claim is that distilling teacher-generated chains of thought into a few learnable latent tokens transfers reasoning ability to a foundation recommendation model more effectively than using the explicit rationales at inference time. WhisperRec generates multi-view adaptive CoT supervision—exploration of intents, evaluation of a candidate item, and attribution of an observed conversion—then distills these traces into three latent tokens via a three-stage alignment process, and finally post-trains with a curriculum mixing standard and latent-reasoning inputs. During inference, only the latent tokens and user context are used to predict the next item's semantic ID. The paper reports

What carries the argument

The key machinery combines Multi-View Adaptive CoT (MV-ACoT) with latent-token alignment. MV-ACoT builds three complementary reasoning tasks—Exploration, Evaluation, and Attribution—and adapts reasoning length to instance difficulty through an unspecified complexity assessor d=g(U,H). A three-stage alignment procedure trains the model to generate each rationale conditioned on a shared set of latent tokens, forcing the tokens to encode common reasoning context, then aligns the tokens with next-item prediction. At inference, the latent tokens are fed as inputs and only the target semantic ID is decoded, so reasoning cost is bounded by the fixed number of tokens, not by rationale length.

Load-bearing premise

The complexity split d=g(U,H) in Equation 4 is never defined, so the claimed benefit of adaptive over fixed CoT rests on an unspecified function that might be trivial, arbitrary, or tuned on the validation set.

What would settle it

If a reproduction where g is fixed to always-'High' (or always-'Low') yields the same SID@64 as the reported adaptive MV-ACoT, then the adaptive complexity component is not the source of the gains; equivalently, a random g that preserves the gap would falsify the paper's adaptive-reasoning claim.

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

If this is right

  • Latency-sensitive recommender systems can adopt reasoning without paying the autoregressive generation cost, since the reasoning overhead is bounded by a fixed number of latent tokens.
  • The quality of teacher reasoning—measured by multi-view coverage and complexity adaptation—directly transfers to recommendation accuracy, so improving supervision quality is a viable path to better foundation recommendation models.
  • The reported ~0.8 semantic similarity between latent reconstructions and explicit rationales suggests latent tokens can serve as a compressed, interpretable reasoning artifact for auditing or explanation.
  • Mixing latent-reasoning and standard SFT in post-training lets one model serve both reasoning and non-reasoning modes, which is practical for heterogeneous traffic.

Where Pith is reading between the lines

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

  • The complexity assessor g is left unspecified; a natural next step is to learn it jointly from the recommendation loss or replace it with a confidence-based classifier, which would make the adaptive claim testable and could improve robustness.
  • The latent-token bottleneck may generalize beyond recommendation: any task where rationales are long but mostly redundant could benefit from distilling them into a few learned tokens, provided the alignment procedure transfers.
  • The paper's similarity evaluation uses a fixed embedding model; a stricter test would be decoding the latent tokens into human-judged explanations to see whether the compressed representation is genuinely faithful, not just embedding-close.

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 / 5 minor

Summary. The paper proposes WhisperRec, a latent reasoning framework for foundation recommendation models (FRMs). It introduces Multi-View Adaptive CoT (MV-ACoT) to generate diverse teacher rationales from exploration, evaluation, and attribution views, with adaptive reasoning complexity per instance. These rationales are distilled into a small set of learnable latent tokens via a three-stage Latent Token Alignment procedure, followed by curriculum-based post-training. During inference, the model prepends the latent tokens and decodes only the target Semantic ID sequence, avoiding autoregressive CoT generation. Experiments on an industrial Kuaishou dataset and the public Kuaishou LLM-Rec benchmark report that WhisperRec improves SID@64 by 17.44% over the explicit Think variant and 9.33% over the No-Think variant of OneReason, while achieving roughly 17.5x higher online inference throughput than explicit CoT.

Significance. If the central claims hold, WhisperRec offers a practical way to retain some of the benefits of reasoning for recommendation while avoiding the latency of explicit CoT generation—an important direction for real-world FRM deployment. The paper has notable strengths: it evaluates on both an industrial dataset and a public benchmark, initializes from a publicly released OneReason checkpoint, reports detailed ablations (latent-token count, CoT construction variants, efficiency), and includes appendices with prompt and rubric details. However, the magnitude of the reported advantage over explicit CoT is highly sensitive to baseline selection, and the key adaptive-complexity mechanism is underspecified. These issues materially affect the strength and reproducibility of the central claims.

major comments (4)
  1. [Table 1, Overall Results (RQ1)] The headline gains of 17.44% and 9.33% are computed against the OneReason-CoT 'OR' rows (industrial SID@64: Think 0.1697, NoThink 0.1823), not the strongest explicit-CoT variants. Against the strongest No-Think baseline (Eval, industrial SID@64 = 0.1962), WhisperRec-Merge (0.1993) improves by only 1.58%; on the public set, the gain vs NoThink-Eval is 12.2%. Several single-view comparisons are negative (e.g., WhisperRec-Attr industrial 0.1936 vs NoThink-Attr 0.1943; WhisperRec-OR public 0.0092 vs NoThink-OR 0.0104). No error bars or significance tests are reported. The claim that latent reasoning outperforms explicit CoT therefore needs to be re-benchmarked against the best explicit variant, with paired comparisons and statistical testing.
  2. [Eq. (4), Adaptive CoT Generation] The adaptive complexity split d = g(U, H) with d ∈ {Low, High} is load-bearing for the MV-ACoT contribution, yet the function g is never specified—no input features, threshold, training procedure, or validation criterion. Table 4 attributes clear gains to Adaptive CoT over OneReason-CoT, but without g the result is unreproducible and it is impossible to assess whether the advantage comes from principled complexity adaptation or from a heuristic tuned on the validation set. Appendix C gives only qualitative principles. A concrete operational definition of g is required.
  3. [Latent Reasoning Analysis (RQ3), Table 5 and Figure 3] The comparison between latent reasoning (WhisperRec) and explicit CoT is confounded by differences in CoT supervision. Table 5 compares OneReason-UnCoT and OneReason-CoT against WhisperRec, but WhisperRec is trained with MV-ACoT teacher traces while the explicit-CoT baseline appears to use generic OneReason-CoT. Figure 3 holds the CoT view fixed but compares only against the Think variant, not the No-Think variant that achieves the strongest baseline numbers. To isolate the effect of latent vs explicit reasoning, the CoT supervision, backbone, and training data should be identical, with only the reasoning format changed.
  4. [Interpretability, Table 7] The cosine-similarity result (0.788–0.804) measures similarity between latent-token reconstructions and the very CoT traces used to train the latent tokens. Since the reconstruction objective is trained to reproduce those traces, high similarity to the training data is expected and does not demonstrate that the latent tokens encode decision-relevant semantics beyond memorization. A more discriminative test would compare against a control (e.g., untrained tokens, tokens trained on permuted traces) or measure downstream steering effects on recommendation decisions.
minor comments (5)
  1. [Throughout] Typos: 'sucess' in Related Work; 'argure' in Section Multi-View CoT Task Design; 'no-attribute' in Appendix C should likely be 'no-attribution'.
  2. [References] Liu et al. 2025a and 2025b cite the same title 'Onerec-think: In-text reasoning for generative recommendation' with identical arXiv identifiers; please verify and correct the duplicate or distinct references.
  3. [Eq. (15) and Eq. (18)] The curriculum order (high→medium→low activity) and the 1:1 mixing ratio in the post-training objective are design choices that are not ablated. Given their potential impact, report sensitivity or justify the choices empirically.
  4. [Table 3 and Appendix D] The LLM-as-judge protocol uses explicit score caps (e.g., factuality ≤ 2 for certain violations). These caps may compress or amplify differences among CoT variants. Reporting inter-judge agreement or using multiple judge models would strengthen the validity of the CoT quality comparison.
  5. [Table 1, 'vs. SOTA' row] The 'vs. SOTA' percentages appear to be computed against the OneReason-CoT Think 'OR' row, which is not the strongest explicit baseline on the public benchmark (e.g., Think-Eval SID@64 = 0.0096 vs Think-OR = 0.0089). Using the strongest explicit baseline would yield more meaningful and less inflated improvement figures.

Circularity Check

2 steps flagged

Central SID@64 results are externally benchmarked, but the interpretability evidence (Table 7) is a fitted-input-called-prediction check and the CoT-quality judge partially measures the method's own design goals.

specific steps
  1. fitted input called prediction [Latent Reasoning Analysis (RQ3), 'Interpretability' paragraph and Table 7; Stage I/II objectives Eqs. (9)-(10)]
    "For each MV-ACoT view, we reconstruct reasoning semantics from latent tokens and measure their cosine similarity to annotated Plain CoT counterparts using Qwen3-Emb-8B. As shown in Table 7, the reconstructed latent semantics achieve consistently high similarity across Exploration, Evaluation, and Attribution (0.788–0.804). These results provide evidence that latent tokens retain CoT semantics while reasoning in the latent space."

    Stage I (Eqs. 8-9) and Stage II (Eq. 10) train the latent tokens L with the objective of predicting the rationale token sequences e_R given the latent slots: L_single = -E_u Σ_t log p_φ(e_r | X_u, L, <v>, e_r,<t). The 'Plain CoT counterparts' in Table 7 are exactly these same training targets. Reconstructing from L and measuring similarity to the training rationales therefore verifies that the training objective was achieved; a high score is expected by construction (and cannot even distinguish CoT information stored in L from a plausible rationale generated from X_u alone). Reporting this as evidence that latent tokens 'retain CoT semantics' presents the fit as a prediction.

  2. self definitional [CoT Analysis (RQ2), 'LLM-as-a-Judge Evaluation' paragraph and Table 3; rubric in Appendix D; prompt principles in Table 10/Appendix C]
    "For each CoT construction view, we sample 100 instances and score them with a unified rubric covering nine dimensions: factuality, evidence selection, signal strength, intent translation, causal logic, ad-user fit, spatiotemporal reasoning, confidence, and task fulfillment. We use GPT-5.5 as the judge; the complete prompt and rubric are provided in Appendix D."

    The judge's nine quality dimensions are the operational definition of 'CoT quality,' and they largely mirror the constraints MV-ACoT's prompts were engineered to enforce (Table 10: interest-to-demand translation, adaptive reasoning path selection, fulfillment and industry knowledge, evidence grounding, confidence calibration). Scoring MV-ACoT against a rubric built from MV-ACoT's own design goals is partly self-verifying: OneReason-CoT scores lower on several dimensions mainly because it was never instructed to produce these components. The claim 'MV-ACoT provides higher-quality CoT' is therefore partially circular, although the independent downstream gains in Table 4 mitigate this.

full rationale

The paper's central quantitative claims—SID@64 gains over explicit-CoT OneReason variants and >10x throughput—are externally anchored: experiments use the public Kuaishou LLM-Rec benchmark and the public OneReason-0.8B-pretrain-competition checkpoint, and Table 1 reports concrete numbers rather than quantities derived from fitted inputs. No fitted parameter is renamed as a prediction in the main recommendation results; the unspecified difficulty function d=g(U,H) (Eq. 4) is a reproducibility gap, not circularity. Genuine circularity is confined to two pieces of supporting evidence. First, the interpretability check (Table 7) measures cosine similarity between latent-token reconstructions and the very teacher rationales used as training targets in Eqs. (9)-(10); because the latent tokens are optimized to predict those rationale tokens, high similarity is the training objective itself, so this 'evidence that latent tokens retain CoT semantics' reduces by construction to checking the fit. Second, the LLM-as-judge CoT-quality comparison (Table 3) scores MV-ACoT on a rubric whose dimensions mirror MV-ACoT's own prompt principles (Table 10/Appendix C), making the measured quality advantage partially self-verifying; this is mitigated by independent downstream gains in Table 4. The paper cites its own prior line (OneReason, OneRec, OpenOneRec; overlapping Kuaishou authors) heavily, but since the baseline checkpoint is public and the comparison is re-run in this paper, these self-citations are not load-bearing. Headline-adjacent concerns—the 9.33%/17.44% gains use the weakest OR rows, versus the best No-Think 'Evaluation' row the industrial gain is about 1.6%, several single-view comparisons are negative, and no significance tests are reported—are correctness and robustness risks under the given rules, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 7 axioms · 3 invented entities

The paper's contribution is an engineering pipeline. The only genuinely new mechanism is the latent-token distillation; its effectiveness is supported by internal experiments but not by independent formal or external evidence. The adaptive-difficulty estimator and the reconstruction decoder are the largest unaccounted components.

free parameters (6)
  • Number of latent tokens K = 3
    Selected by validation ablation (Figure 4): Hit rises from 1 to 3 tokens and saturates; this is a model-selection choice, not a prediction.
  • Adaptive difficulty dichotomy d in {Low, High} = unspecified
    Eq. (4) defines d=g(U,H) but the estimator is never described; the Low/High split is hand-set and load-bearing for the adaptive-Co T claim.
  • Post-training mixture ratio lambda=1:1 = 0.5/0.5
    Eq. (18) mixes latent-reasoning and standard SFT streams at 1:1; no sensitivity analysis is given.
  • Curriculum order high to low activity = high, medium, low
    Eq. (15) curriculum ordering is chosen by hand; no ablation is presented.
  • History length L = 50
    Appendix A keeps at most 50 interactions; a preprocessing choice that shapes inputs.
  • LLM-judge score caps = factuality <= 2 / <= 3
    Appendix D hard caps (e.g., fabricated evidence caps factuality at 2) are protocol choices that materially change CoT quality scores.
axioms (7)
  • domain assumption Teacher Qwen3-235B-Instruct-2507 generates decision-relevant CoT that is a valid supervision target
    MV-ACoT quality is only shown via the authors' GPT-5.5 judge rubric (Table 3); no external validation.
  • domain assumption Res-Kmeans SIDs with 3 levels and codebook 1024 faithfully encode item semantics
    Inherited from prior work (Luo et al. 2025; OneRec); no error analysis for SID collisions affecting ID@K metrics.
  • domain assumption OneReason-0.8B-pretrain-competition checkpoint is a capable FRM base
    The public checkpoint (Team 2026) is from the same lab; its quality bounds WhisperRec's gains.
  • domain assumption Language-model cross-entropy on rationale tokens transfers reasoning into latent vectors
    Stages I-II (Eqs. 8-10) assume the LM loss on CoT creates reusable latent semantics; the paper only checks similarity to the same CoT.
  • domain assumption SID@K and ID@K are adequate recommendation metrics
    No diversity/coverage metrics, no online A/B; leaderboard-style metrics may not reflect business value.
  • domain assumption GPT-5.5 LLM-as-judge scores are a valid proxy for CoT quality
    Appendix D rubric is authored by the authors; no judge-human agreement is reported.
  • domain assumption Qwen3-Emb-8B cosine similarity reflects semantic alignment of reasoning
    Used for the interpretability claim (Table 7); no threshold or calibration is given.
invented entities (3)
  • Latent reasoning tokens Z (K=3) no independent evidence
    purpose: Compressed storage of CoT-derived reasoning; fed as prefix at inference to predict target SID
    Internal to the model; only evidence is in-paper ablations and reconstruction similarity to the same CoT used for training.
  • View control tokens for exploration/evaluation/attribution no independent evidence
    purpose: Condition latent-reasoning training on one of the three CoT views
    Learnable vocabulary entries; no external handle.
  • Latent reconstruction decoder no independent evidence
    purpose: Decode latent tokens into text embeddings for the interpretability analysis (Table 7/Appendix B)
    Its architecture and training are not described in the methodology; appears only in the evaluation.

pith-pipeline@v1.3.0-daily-deepseek · 17946 in / 26178 out tokens · 280953 ms · 2026-08-01T12:07:35.814099+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) have demonstrated strong reasoning capabilities, motivating their adoption as backbones for foundation recommendation models (FRMs). Existing approaches typically enhance recommendation with explicit Chain-of-Thought (CoT) under the Think-then-Answer paradigm. However, generating lengthy rationales introduces substantial inference overhead, while fixed CoT templates struggle to model diverse, dynamic, and context-dependent user interests. We propose WhisperRec, an efficient latent reasoning framework for FRMs. WhisperRec compresses teacher-generated CoT into learnable latent reasoning tokens, enabling a Latent-Reason-then-Answer paradigm that performs reasoning in latent space without producing verbose rationales. This design retains decision-relevant reasoning information while avoiding the latency bottleneck of autoregressive rationale generation. Specifically, it first introduces Multi-View Adaptive CoT (MV-ACoT) to construct diverse, high-quality supervision from complementary perspectives on user interests. MV-ACoT also adapts reasoning complexity to each instance, applying lightweight analysis to clear cases and targeted multi-factor reasoning to challenging ones. Building on a pre-trained FRM, WhisperRec then employs a three-stage Latent Reasoning Alignment procedure to progressively internalize teacher CoT into latent representations. Finally, curriculum-based post-training activates latent-token reasoning for downstream recommendation while preserving standard recommendation capability. Experiments on an industrial-scale Kuaishou dataset and the public Kuaishou LLM-Rec benchmark show that WhisperRec consistently outperforms explicit-CoT methods and conventional baselines. Compared with explicit CoT Think and No-Think variants, WhisperRec improves SID@64 by 17.44% and 9.33%, respectively, and achieves over 10x higher online inference throughput.

Figures

Figures reproduced from arXiv: 2607.26621 by Fei Pan, Hao Jiang, Jian Liang, Kuo Cai, Mengting Li, Peiru Du, Pengfei Yao, Peng Jiang, Qiang Luo, Ruiming Tang, Sheng Yu, Siyuan Lou, Wenwu Ou.

Figure 1
Figure 1. Figure 1: Limitations of explicit CoT for recommendation: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the WhisperRec, containing Multi-view Adaptive CoT, Latent Token Alignment and Post-training stage. Methodology Problem Formulation and Framework For a user u, let Pu denote the user profile, Hu the chrono￾logically ordered interaction history, Cu the optional spa￾tiotemporal context, and i + u the next interacted item. Each item i is represented by a Semantic ID (SID) sequence Si = [s (1) i ; … view at source ↗
Figure 3
Figure 3. Figure 3: SID@64 performance comparison between OneReason Think and WhisperRec under different CoT. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on the number of latent tokens. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: General and recommendation capability. Conclusion In this paper, we propose WhisperRec, a latent reasoning framework for foundation recommendation models. Efficient reasoning is essential for deploying reasoning-enabled FRMs in industrial recommendation systems. Built upon OneRea￾son, a strong FRM baseline, WhisperRec addresses the lim￾itations of explicit CoT through MV-ACoT for diverse CoT construction, … view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between latent reconstruction and explicit CoT. The two traces share core commercial semantics, including [PITH_FULL_IMAGE:figures/full_fig_p012_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

17 extracted references · 11 linked inside Pith

  1. [1]

    OneSearch- V2:TheLatentReasoningEnhancedSelf-distillationGener- ative Search Framework.arXiv preprint arXiv:2603.24422

    Chen, B.; Wang, S.; Ma, Y.; Liang, Z.; Zhang, X.; Lv, Y.; Yang,Y.;Dai,H.;Mao,L.;Zhao,T.;etal.2026. OneSearch- V2:TheLatentReasoningEnhancedSelf-distillationGener- ative Search Framework.arXiv preprint arXiv:2603.24422. Dai,S.;Tang,J.;Wu,J.;Wang,K.;Zhu,Y.;Chen,B.;Hong, B.; Zhao, Y.; Fu, C.; Wu, K.; et al

    Pith/arXiv arXiv 2026

  2. [3]

    Fang, Y.; Wang, W.; Zhang, Y.; Zhu, F.; Wang, Q.; Feng, F.; and He, X

    Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment.arXiv preprint arXiv:2502.18965. Fang, Y.; Wang, W.; Zhang, Y.; Zhu, F.; Wang, Q.; Feng, F.; and He, X

    Pith/arXiv arXiv

  3. [4]

    Gao, Y.; Sheng, T.; Xiang, Y.; Xiong, Y.; Wang, H.; and Zhang, J

    Reason4rec: Large language models for recommendation with deliberative user preference align- ment.arXiv preprint arXiv:2502.02061. Gao, Y.; Sheng, T.; Xiang, Y.; Xiong, Y.; Wang, H.; and Zhang, J

    Pith/arXiv arXiv

  4. [6]

    Hidasi,B.;Karatzoglou,A.;Baltrunas,L.;andTikk,D.2015

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948. Hidasi,B.;Karatzoglou,A.;Baltrunas,L.;andTikk,D.2015. Session-based recommendations with recurrent neural net- works.arXiv preprint arXiv:1511.06939. Hou, Y.; Zhang, J.; Lin, Z.; Lu, H.; Xie, R.; McAuley, J.; andZhao,W.X.2024. Largelanguagem...

    Pith/arXiv arXiv 2015

  5. [8]

    Luo, X.; Cao, J.; Sun, T.; Yu, J.; Huang, R.; Yuan, W.; Lin, H.; Zheng, Y.; Wang, S.; Hu, Q.; et al

    GOT4Rec: Graph of Thoughts for Sequential Recommendation.arXiv preprint arXiv:2411.14922. Luo, X.; Cao, J.; Sun, T.; Yu, J.; Huang, R.; Yuan, W.; Lin, H.; Zheng, Y.; Wang, S.; Hu, Q.; et al

    arXiv

  6. [10]

    Xing, H.; Deng, H.; Mao, Y.; Mu, L.; Hu, J.; Xu, Y.; Zhang, H.; Wang, J.; Wang, S.; Zhang, Y.; et al

    OneReason Technical Report.arXiv preprint arXiv:2606.06260. Xing, H.; Deng, H.; Mao, Y.; Mu, L.; Hu, J.; Xu, Y.; Zhang, H.; Wang, J.; Wang, S.; Zhang, Y.; et al

    Pith/arXiv arXiv

  7. [11]

    Yi, C.; Chen, D.; Guo, G.; Tang, J.; Wu, J.; Yu, J.; Zhang, M.; Chen, W.; Yang, W.; Luo, Y.; et al

    Reg4rec: Reasoning-enhanced generative model for large-scale rec- ommendation systems.arXiv preprint arXiv:2508.15308. Yi, C.; Chen, D.; Guo, G.; Tang, J.; Wu, J.; Yu, J.; Zhang, M.; Chen, W.; Yang, W.; Luo, Y.; et al. 2025a. RecGPT-V2 Technical Report.arXiv preprint arXiv:2512.14503. Yi,C.;Chen,D.;Guo,G.;Tang,J.;Wu,J.;Yu,J.;Zhang,M.; Dai, S.; Chen, W.; Y...

    arXiv

  8. [12]

    InProceedings of the ACM Web Conference 2026, 5698–5709

    ThinkRec: Thinking-based recommendation via LLM. InProceedings of the ACM Web Conference 2026, 5698–5709. Zhai, J.; Liao, L.; Liu, X.; Wang, Y.; Li, R.; Cao, X.; Gao, L.; Gong, Z.; Gu, F.; He, M.; et al

    2026

  9. [13]

    Zhang,B.;Luo,L.;Chen,Y.;Nie,J.;Liu,X.;Guo,D.;Zhao, Y.; Li, S.; Hao, Y.; Yao, Y.; et al

    Ac- tionsspeaklouderthanwords:Trillion-parametersequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152. Zhang,B.;Luo,L.;Chen,Y.;Nie,J.;Liu,X.;Guo,D.;Zhao, Y.; Li, S.; Hao, Y.; Yao, Y.; et al

    Pith/arXiv arXiv

  10. [14]

    Zhang, L.; Huang, Y.; Lv, H.; Zhi, X.; Yin, M.; Ye, Y.; Guo, W.; Wang, H.; and Chen, E

    Wukong: Towards a scaling law for large-scale recommendation.arXiv preprint arXiv:2403.02545. Zhang, L.; Huang, Y.; Lv, H.; Zhi, X.; Yin, M.; Ye, Y.; Guo, W.; Wang, H.; and Chen, E

    Pith/arXiv arXiv

  11. [15]

    Zheng, B.; Hou, Y.; Lu, H.; Chen, Y.; Zhao, W

    Why Thinking Hurts:DiagnosingandRectifyingLinguisticInertiainLarge Language Models for Recommendation.arXiv preprint arXiv:2602.16587. Zheng, B.; Hou, Y.; Lu, H.; Chen, Y.; Zhao, W. X.; Chen, M.; and Wen, J.-R

    Pith/arXiv arXiv

  12. [16]

    In2024 IEEE 40th International Conference on Data Engi- neering (ICDE), 1435–1448

    Adapting large language models by integrating collaborative semantics for recommendation. In2024 IEEE 40th International Conference on Data Engi- neering (ICDE), 1435–1448. IEEE. Zhou, C.; Wang, M.; Ma, Y.; Wu, C.; Chen, W.; Qian, Z.; Liu, X.; Zhang, Y.; Wang, J.; Xu, H.; et al. 2025a. From perception to cognition: A survey of vision-language inter- activ...

    arXiv

  13. [2017]

    Deep interest network for click-through rate prediction.arXiv preprint arXiv:1706.06978. Appendix A: Dataset Construction and Parameters Setting Implementation details In our experiment, we perform full-parameter supervised fine-tuning on two complementary training streams, con- sisting of standard recommendation instances and chain-of- thought(CoT) augme...

    Pith/arXiv arXiv

  14. [2023]

    Geng,S.;Liu,S.;Fu,Z.;Ge,Y.;andZhang,Y.2022

    Chat-rec: Towards interactive and explain- able llms-augmented recommender system.arXiv preprint arXiv:2303.14524. Geng,S.;Liu,S.;Fu,Z.;Ge,Y.;andZhang,Y.2022. Recom- mendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). InProceed- ings of the 16th ACM conference on recommender systems, 299–315. Guo, D.; Y...

    Pith/arXiv arXiv 2022

  15. [2024]

    Kang,W.-C.;andMcAuley,J.2018.Self-attentivesequential recommendation

    Openai o1 system card.arXiv preprint arXiv:2412.16720. Kang,W.-C.;andMcAuley,J.2018.Self-attentivesequential recommendation. In2018IEEEinternationalconferenceon data mining (ICDM), 197–206. IEEE. Kong,X.;Jiang,J.;Liu,B.;Xu,Z.;Zhu,H.;Xu,J.;Zheng,B.; Wu,J.;andWang,X.2026. ThinkbeforeRecommendation: AutonomousReasoning-enhancedRecommender.Advances in Neural ...

    Pith/arXiv arXiv 2018

  16. [2025]

    Deng, J.; Wang, S.; Cai, K.; Ren, L.; Hu, Q.; Ding, W.; Luo, Q.; and Zhou, G

    Onepiece: Bring- ing context engineering and reasoning to industrial cascade ranking system.arXiv preprint arXiv:2509.18091. Deng, J.; Wang, S.; Cai, K.; Ren, L.; Hu, Q.; Ding, W.; Luo, Q.; and Zhou, G

    arXiv

  17. [2026]

    Rajput, S.; Mehta, N.; Singh, A.; Hulikal Keshavan, R.; Vu, T.; Heldt, L.; Hong, L.; Tay, Y.; Tran, V.; Samost, J.; etal.2023

    Qwen3.5: Towards Native Multimodal Agents. Rajput, S.; Mehta, N.; Singh, A.; Hulikal Keshavan, R.; Vu, T.; Heldt, L.; Hong, L.; Tay, Y.; Tran, V.; Samost, J.; etal.2023. Recommendersystemswithgenerativeretrieval. Advances in Neural Information Processing Systems, 36: 10299–10315. Rajput,S.;Mehta,N.;Singh,A.;Keshavan,R.;Vu,T.;Heldt, L.; Hong, L.; Tay, Y.; ...

    2023