Pith. sign in

REVIEW 3 major objections 9 minor 6 cited by

RecGPT Technical Report

T0 review · 3 major / 9 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that RecGPT, an LLM-based intent-centric recommender, has been fully deployed on Taobao's homepage and raises user, merchant, and platform metrics together—clicked-item diversity by 6.96%, dwell time by 4.82%, CTR by…

desk verdict A coherent, deployed industrial LLM recommender pipeline with genuinely useful engineering detail, but the headline A/B gains are not statistically supported as reported. read the letter →

arxiv 2507.22879 v2 pith:XT4HKU5X submitted 2025-07-30 cs.IR cs.CL

Chao Yi , Dian Chen , Gaoyang Guo , Jiakai Tang , Jian Wu , Jing Yu , Mao Zhang , Sunhao Dai
show 46 more authors
Wen Chen Wenjun Yang Yuning Jiang Zhujin Gao Bo Zheng Chi Li Dimin Wang Dixuan Wang Fan Li Fan Zhang Haibin Chen Haozhuang Liu Jialin Zhu Jiamang Wang Jiawei Wu Jin Cui Ju Huang Kai Zhang Kan Liu Lang Tian Liang Rao Longbin Li Lulu Zhao Na He Peiyang Wang Qiqi Huang Tao Luo Wenbo Su Xiaoxiao He Xin Tong Xu Chen Xunke Xi Yang Li Yaxuan Wu Yeqiu Yang Yi Hu Yinnan Song Yuchen Li Yujie Luo Yujin Yuan Yuliang Yan Zhengyang Wang Zhibo Xiao Zhixin Ma Zile Zhou Ziqi Zhang
This is my paper · ORCID
classification cs.IRcs.CL
keywords largelanguagemodelsrecommendersystemsuserinterestminingitemtagpredictionintent-centricrecommendationonlinedeploymentTaobaoretrieval
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

RecGPT is a production recommender system that places explicit user intent, rather than historical click patterns, at the center of the pipeline. The paper claims that by adding three reasoning LLMs—one that mines a user's interests from compressed lifelong behavior, one that turns those interests into fine-grained item tags, and one that writes personalized explanations—Taobao's homepage recommendation scenario gains across all stakeholders: clicked-item category diversity (+6.96%), dwell time (+4.82%), click-through rate (+6.33%), item page views (+9.47%), and daily click active users (+3.72%). The design's core wager is that an LLM's world knowledge can surface latent interests that co-occurrence-based collaborative filtering misses, enlarging the candidate set without sacrificing relevance. If the online results are what they appear to be, intent-centric LLM components can be layered onto an existing industrial retrieval-ranking cascade without replacing it.

What carries the argument

The load-bearing mechanism is the closed loop 'User Interest Mining → Item Tag Prediction → Item Retrieval → Explanation Generation', carried by three LLMs (LLM_UI, LLM_IT, LLM_RE). To fit average histories of over 37k records into a 128k-token window, the paper's hierarchical behavior compression aggregates behaviors first by time-behavior type and then reverse-aggregates by item, yielding a 'Time (Behaviors) | Items' format that covers 98% of users and speeds inference by 29%. The retrieval stage is a tri-tower matcher: a user tower and item tower supply the collaborative score ŷ_col, a tag tower supplies the semantic score ŷ_sem, and online inference fuses them as ŷ_final = β ŷ_col + (1−β) ŷ_sem, giving operators a single knob (β) to trade exploitation of behavior patterns against exploration of LLM-inferred interests. Explanations are produced offline as an interest–item–explanation lookup table so that online serving is a hash lookup rather than a generation call.

What would settle it

For the roughly 88% of users whose raw sequences already fit in the 128k context, run LLM_UI on both the full sequence and the compressed version, and compare human-evaluated interest pass rates on the same users. If compressed input produces meaningfully lower pass rates, the compression step destroys the very signals the system depends on, and the reported gains would be attributable to something else.

Watch

Extended reading notes

Core claim

The central claim is that the long-standing log-fitting paradigm of recommender systems—learning clicks from clicks—can be replaced by an intent-centric loop: compress a user's lifelong multi-behavior history into a compact, information-preserving sequence; have an LLM reason over it to produce an explicit natural-language interest profile; have a second LLM translate that profile into a set of specific item tags in 'modifier + core-word' form; retrieve items by fusing tag–item semantic similarity with the user–item collaborative score; and attach a pre-generated, cached natural-language explanation to each final item. The paper reports that this loop has been fully deployed in the 'Guess What You Like' scenario on Taobao's homepage and that the online A/B test shows consistent multi-stakeholder gains, including a 6.96% improvement in clicked-item category diversity and a 6.33% improvement in click-through rate, with the tag-based semantic channel also producing more uniform CTR across item popularity groups, which the authors read as mitigation of the Matthew effect.

Load-bearing premise

The whole pipeline assumes that compressing a user's lifelong behavior into a 'time-behavior | items' summary preserves the information the LLM needs to infer that user's true interests.

Editorial extensions

If this is right

  • An LLM-based intent layer can sit on top of an existing user–item dual-tower matcher and downstream ranking cascade, meaning the change is an additive upgrade rather than a green-field rebuild.
  • The fused score's balance between collaborative and semantic channels is a single tunable parameter β, so the explore/exploit trade-off can be adjusted per scenario without retraining.
  • Because tags are in a constrained 'modifier + core-word' grammar, the same pipeline generalizes to any domain with a tagged item corpus; only the taxonomy needs to be replaced.
  • Pre-generating explanations into a lookup table means personalization and transparency do not have to cost online inference latency.
  • The Human-LLM judge system implies data curation and model evaluation can scale to industrial iteration cycles while still being anchored to human standards.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The weakest link is likely the compression step: the paper validates that 98% of sequences fit in the context window, not that the compressed representation yields the same interest profiles as the full history; a user with a dense, idiosyncratic past may lose exactly the signals that distinguish them.
  • The reported redundancy reduction nearly doubles when ad cards are excluded (0.88% to 1.57%), suggesting ad placement confounds perceived diversity; an A/B test that isolates organic recommendations would give a cleaner estimate of the recommender's own effect.
  • The Matthew-effect claim rests on normalized CTR/PVR distributions across popularity groups; a direct measurement of tail-item exposure lift conditional on relevance would separate genuine exploration from simple re-ranking of popularity.
  • If the interest profiles are refreshed only every two weeks, the system is deliberately slower than the click stream itself; a faster refresh cadence, or continuous tag updates via the incremental-learning path, may be where further gains lie.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 9 minor

Summary. This technical report describes RecGPT, a production recommender system deployed on Taobao's homepage "Guess What You Like" scenario. The system places user intent at the center by integrating three LLM components: a User-Interest LLM that mines interests from compressed lifelong behavior sequences, an Item-Tag LLM that predicts product tags from those interests, and a Recommendation-Explanation LLM that generates personalized explanations. The paper details a multi-stage training paradigm (curriculum fine-tuning, reasoning-enhanced pre-alignment, self-training evolution) and a Human-LLM cooperative judge system used for data curation and evaluation. The authors report online A/B gains across user, platform, and merchant metrics (CICD +6.96%, DT +4.82%, CTR +6.33%, IPV +9.47%, DCAU +3.72%), human-evaluated pass rates for the three LLM components, and a user-experience study on perceived redundancy. The central claim is that RecGPT has been fully deployed and achieves consistent multi-stakeholder improvements, validating LLM-driven intent-centric recommendation at industrial scale.

Significance. If the reported online gains are statistically reliable, this paper would be a significant demonstration that LLM-based intent mining and tag retrieval can improve diversity, engagement, and merchant exposure in a large-scale real-world recommender system, with concrete evidence on filter-bubble and Matthew-effect mitigation. The manuscript is also strong in its level of system detail: the three-stage compression pipeline, the tri-tower retrieval architecture, the incremental learning process, and the judge-system training protocol are described concretely enough to be reproduced in other industrial settings. The paper's explicit limitation section is a welcome feature. However, the central deployment claim currently rests on an under-specified A/B test with inconsistent reported duration and no inferential statistics, and several internal numerical inconsistencies weaken the evidential value of the user-experience results. These issues are fixable, but they are load-bearing for the paper's headline claim.

major comments (3)
  1. [Sections 4.1–4.2, Table 8] The headline claim of consistent performance gains is not statistically supported. Section 4.1 describes a one-month online A/B experiment on the top one-third of active users with 1% traffic per group, while Table 8's caption states that the results were "conducted from June 17 to June 20, 2025," a four-day window. No confidence intervals, standard errors, p-values, or daily sample sizes are reported for any metric in Table 8. Given the small traffic fraction and the restricted user population, the reported gains (CTR +6.33%, IPV +9.47%, DCAU +3.72%) could fall within normal day-to-day fluctuation. The manuscript must reconcile the one-month versus four-day discrepancy and provide uncertainty quantification or explicit pre-registered significance tests before the deployment claim can be accepted.
  2. [Section 4.5] The redundancy-reduction arithmetic is internally inconsistent. The text reports that the repetition rate decreased from 37.1% to 36.2%, an absolute drop of 0.9 percentage points, yet it also says "the redundancy improvement nearly doubled from 0.88% (with ads included) to 1.57% (without ads)"; separately, the top-4 repetition decrease from 27.7% to 25.3% is an absolute drop of 2.4 percentage points. These numbers cannot all be percentage-point changes, and the manuscript does not state whether they are relative or absolute. Because the user-diversity benefit is one of the paper's key claims, this section needs to be rewritten with a single, clearly defined measure of redundancy reduction.
  3. [Section 3, Tables 2, 4, 7] The evaluation loop is partially circular. The LLM judge is trained on human-annotated data that uses the same willingness/reasonableness criteria (interest mining) and relevance/consistency/specificity/validity criteria (tag prediction) that were used to filter the self-training data, and the same judge is then used for model monitoring. The human pass rates in Tables 2, 4, and 7 may therefore partly reflect the model's ability to conform to the filtering criteria rather than independent recommendation quality. The online A/B metrics are external and substantially mitigate this concern, but the internal evaluation claims would be stronger with an explicit acknowledgment of this dependence or with a held-out human evaluation using criteria not used during training-data curation.
minor comments (9)
  1. [Section 2.2.2] The model name is first introduced as "Tag-to-Cate model φ(·)" in Step 3 and later referred to as the "Tag-to-Gate model φ(·)"; please unify the terminology.
  2. [Table 9] There is a typo in the table header: "Item Tag Prediciton" should be "Item Tag Prediction."
  3. [Appendix B] In the Recommendation Explanation Generation prompt, the output format field is written as "Explation"; this should be "Explanation."
  4. [Table 3] The typo "Sepcificity" appears in the label column; it should be "Specificity."
  5. [Table 10] The task list contains typographical errors: "Adavanced" should be "Advanced" and "Recommandation" should be "Recommendation."
  6. [Section 4.3] The fine-tuned judge is inconsistently named: Section 4.3 refers to "Qwen3-SFT-Judge" and "Qwen3-SFT-Judge" in one passage, while Table 9 and the evaluation setup use "Qwen3-Judge-SFT"; please standardize.
  7. [Section 1 and Section 4.1] The contribution bullet states that RecGPT is the first deployed "hundred-billion-scale recommendation foundation model," but the deployed model is described as TBStars-MoE-42B-A3.5B (42B total parameters, 3.5B active) and Qwen3-14B for explanations; the claimed scale is unclear and should be substantiated or reworded.
  8. [Section 2.1.1] The compression validation reports sequence coverage (98% within 128k tokens) and inference speed (29% faster), but it does not measure whether the compressed representation yields the same interest-mining quality as the full sequence; an ablation on a held-out set comparing compressed versus uncompressed inputs would strengthen this design choice.
  9. [Section 4.1] The experimental setup should specify the randomization unit (users vs. requests) and whether the 1% traffic per group refers to users, sessions, or requests, as this affects the interpretation of the reported metric gains.

Circularity Check

1 steps flagged · score 4.0 of 10

The judge-in-the-loop validation is partially circular (same judge filters self-training data and assesses performance), but the headline deployment claim is grounded in independent online A/B metrics.

  1. fitted input called prediction [Section 2.1.2 Stage 3; Section 3.1 Dataset Construction; Section 4.3 Evaluation Setup]
    "To efficiently filter these self-generated outputs and evaluate model performance at low cost, we adopt a Human-LLM collaborative paradigm with LLM-as-a-Judge capabilities for data quality control and assessment. ... The annotated samples and results are stored in a Judge Data Buffer, which is subsequently used to fine-tune corresponding LLM-Judges. ... Qwen3-Judge-SFT denotes the fine-tuned version trained on human judgment data."

    The same LLM-Judge is used both to select which self-generated samples enter training and to assess model performance. A model trained on judge-filtered data tends to score well on that same judge by construction, so judge-based performance monitoring in the self-training loop is not an independent measurement. In addition, the reported judge-human agreement (Table 9) is measured against human judgment data of the same kind and criteria used to fine-tune the judge, and no held-out split is reported. If the Judge Data Buffer is reused for evaluation, the agreement is a fitted-input prediction. The online A/B metrics (Section 4.2) are measured on real user behavior and are independent of this loop, so the central deployment claim remains externally grounded.

full rationale

RecGPT's central claim, full deployment with multi-stakeholder gains, rests on the online A/B test in Section 4.2, which compares the full system against a baseline on real user behavior (CTR, IPV, DCAU, CICD, DT). Those metrics are not derived from the LLM-judge loop or from the training criteria, so the headline is not circular. The partial circularity is confined to the Human-LLM cooperative judge system: the judge filters self-training data and is also used to assess model performance, and its claimed reliability is validated against human annotations of the same type used to train it, without a stated held-out split. This makes the internal low-cost assessment and Table 9 agreement partly self-referential, though the human pass-rate tables (Tables 2, 4, 7) and the HR@30 evaluation on real next-interaction categories provide some independent component-level evidence. Section 5 also acknowledges a nontrivial limitation, namely that about 2% of sequences still exceed the 128K token limit and long-sequence noise can bias user understanding; that is a weakness, not a circularity. The A/B evidence itself lacks significance testing and uses a short, potentially inconsistent window, but statistical-validity concerns are distinct from circular-derivation concerns.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claims rest mainly on domain assumptions about which behaviors are reliable, how well LLM-generated tags can be matched to real items, and whether human-defined quality criteria are consistent. The only numeric choices reported are a handful of hyperparameters (alpha, beta, tag counts, sampling caps). No new physical or conceptual entities are introduced. The assumptions are standard for industrial recommender papers but are not independently validated beyond the authors' internal evaluations.

free parameters (6)
  • alpha (loss balance) = 0.5
    Balances tag contrastive loss and category contrastive loss in LTAR (Section 2.3.2). Chosen by hand, not swept.
  • beta (fusion weight) = not reported
    Weights user tower vs tag tower in online inference (Section 2.3.3). Not specified, so online results are not reproducible.
  • minimum tag count = 50
    Prompt constraint enforcing diversity in item tag prediction (Section 2.2.1). Changes candidate set size and affects retrieval.
  • number of interests to reason over = 10
    Prompt requires reasoning over at least 10 interests (Appendix B). Affects interest pool coverage.
  • predicted tag count for HR@30 = 30
    Used in incremental learning evaluation (Section 2.2.2). The metric's value depends on this number.
  • maximum samples per category in data balancing = 2
    Secondary sampling cap in incremental learning (Section 2.2.2).
assumptions (6)
  • domain assumption Ordinary product clicks are excluded because they contain noise and are less reflective of interest than favorites, purchases, add-to-cart, detailed views, review reads, and searches.
    Section 2.1.1 Reliable Behavior Extraction. If ordinary clicks encode latent interest, the pipeline loses those signals before the LLM sees them.
  • domain assumption LLM-generated item tags in 'Modifier + Core-Word' format can be mapped to concrete in-domain items through the tri-tower retrieval model.
    Section 2.3. The tag tower and item tower are trained with contrastive losses, but the mapping quality depends on embedding alignment across tag and item text, which is not directly evaluated.
  • domain assumption The human-annotated quality criteria (willingness, reasonableness, relevance, consistency, specificity, validity, factuality, clarity, safety) are stable, complete, and consistently applied across annotators.
    Sections 2.1.2, 2.2.1, 2.4.1. These criteria define the pass-rate evaluations and judge training data; inter-annotator agreement is not reported.
  • domain assumption User interests inferred from up to two-week-old behavior remain valid for service until the next refresh.
    Section 2.1.2 Online Deployment states interests are refreshed every two weeks; rapidly shifting interests may be stale within that window.
  • domain assumption DeepSeek-R1 generated reasoning samples, after manual curation and distillation, transfer competently to the smaller Qwen3 and TBStars models.
    Section 2.1.2 Stage 2, Section 2.2.1, Section 2.4.1. The pass-rate results support this for the tested tasks, but the distillation may not preserve all teacher capabilities.
  • domain assumption The Tag-to-Cate model phi(.) maps item tags to product categories accurately enough for evaluation and interest-item pairing.
    Section 2.2.2 and Section 2.4.2. HR@30 and offline explanation lookup both depend on this mapping, and no accuracy of phi is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RecGPT Technical Report." pith.science (2026). https://pith.science/paper/XT4HKU5X

@misc{pith2026250722879,
  author       = {Pith},
  title        = {Pith review of: RecGPT Technical Report},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XT4HKU5X}},
  note         = {Machine review of arXiv:2507.22879}
}
read the original abstract

Recommender systems are among the most impactful applications of artificial intelligence, serving as critical infrastructure connecting users, merchants, and platforms. However, most current industrial systems remain heavily reliant on historical co-occurrence patterns and log-fitting objectives, i.e., optimizing for past user interactions without explicitly modeling user intent. This log-fitting approach often leads to overfitting to narrow historical preferences, failing to capture users' evolving and latent interests. As a result, it reinforces filter bubbles and long-tail phenomena, ultimately harming user experience and threatening the sustainability of the whole recommendation ecosystem. To address these challenges, we rethink the overall design paradigm of recommender systems and propose RecGPT, a next-generation framework that places user intent at the center of the recommendation pipeline. By integrating large language models (LLMs) into key stages of user interest mining, item retrieval, and explanation generation, RecGPT transforms log-fitting recommendation into an intent-centric process. To effectively align general-purpose LLMs to the above domain-specific recommendation tasks at scale, RecGPT incorporates a multi-stage training paradigm, which integrates reasoning-enhanced pre-alignment and self-training evolution, guided by a Human-LLM cooperative judge system. Currently, RecGPT has been fully deployed on the Taobao App. Online experiments demonstrate that RecGPT achieves consistent performance gains across stakeholders: users benefit from increased content diversity and satisfaction, merchants and the platform gain greater exposure and conversions. These comprehensive improvement results across all stakeholders validates that LLM-driven, intent-centric design can foster a more sustainable and mutually beneficial recommendation ecosystem.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

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

  1. RECAP: Feedback-Driven Streaming Semantic User Profiles for Short-Video Recommendation

    cs.IR 2026-07 conditional novelty 6.0 of 10

    RECAP trains a streaming LLM profile updater with GRPO rewards from a dual-tower evaluator, gaining +0.0084 uAUC (cleaned eval) and +0.139% online usage time.

  2. ShopX: A Foundation Model for Intent-to-Item Fulfillment in Agentic Shopping

    cs.IR 2026-06 unverdicted novelty 6.0 of 10

    ShopX is a single foundation model combining intent understanding, planning, and SID-native item fulfillment for agentic shopping, with claimed improvements over tool-mediated systems on Taobao logs.

  3. UniVA: Unified Value Alignment for Generative Recommendation in Online Advertising at Tencent

    cs.IR 2026-05 unverdicted novelty 6.0 of 10

    UniVA unifies value alignment in generative recommendation via a Commercial SID tokenizer, eCPM-aware RL decoder, and personalized beam search, reporting 37% offline Hit Rate gains and 1.5% online GMV lift on Tencent ...

  4. Deep Interest Mining for Intent-Enriched Semantic IDs in Multimodal Generative Recommendation

    cs.IR 2026-03 unverdicted novelty 6.0 of 10

    A new framework integrating deep interest mining, cross-modal semantic alignment, and quality-aware reinforcement learning generates higher-quality Semantic IDs and outperforms prior methods on recommendation benchmarks.

  5. LLM-Based Generative Retrieval for Snapchat Content Recommendation

    cs.IR 2026-07 conditional novelty 5.0 of 10

    SnapLGR, a production LLM-based generative retrieval system for Snapchat short video, lifted View Time 0.37% and related engagement metrics in a 7-day A/B test, with offline ablation attributing most of the gain to de...

  6. When Language Models Meet NeuroGraphs: Exploring Enhanced Agentic LLM Framework Towards Brain Network Analysis

    cs.MA 2026-07 reject novelty 5.0 of 10

    BrainAgent, a training-free agentic LLM framework with graph understanding, knowledge retrieval, case retrieval, and reflection, claims improved but still moderate connectome classification and interpretability.

Reference graph

Works this paper leans on

15 extracted references · 2 canonical work pages · cited by 6 Pith papers

  1. [4]

    J. Gu, X. Jiang, Z. Shi, H. Tan, X. Zhai, C. Xu, W. Li, Y. Shen, S. Ma, H. Liu, et al. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594,

    arXiv

  2. [5]

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025a. J. Guo, Z. Chi, L. Dong, Q. Dong, X. Wu, S. Huang, and F. Wei. Reward reasoning model.arXiv preprint arXiv:2505.14674, 2025b. J. M. Johnson and...

    arXiv

  3. [6]

    R. Kirk, I. Mediratta, C. Nalmpantis, J. Luketina, E. Hambro, E. Grefenstette, and R. Raileanu. Under- standing the effects of rlhf on llm generalisation and diversity.arXiv preprint arXiv:2310.06452,

    arXiv

  4. [7]

    Y. C. Liu and M. Q. Huang. Examining the matthew effect on youtube recommendation system. In 2021 International Conference on Technologiesand Applications of Artificial Intelligence (TAAI), pages 146–148. IEEE,

    work page 2021

  5. [10]

    G. Son, H. Ko, H. Lee, Y. Kim, and S. Hong. Llm-as-a-judge & reward model: What they can and cannot do.arXiv preprint arXiv:2409.11239,

    arXiv

  6. [12]

    W. Wang, S. Xiong, G. Chen, W. Gao, S. Guo, Y. He, J. Huang, J. Liu, Z. Li, X. Li, et al. Reinforcement learning optimization for large-scale learning: An efficient and user-friendly scaling library.arXiv preprint arXiv:2506.06122,

    arXiv

  7. [13]

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    arXiv

  8. [2009]

    Casper, X

    S. Casper, X. Davies, C. Shi, T. K. Gilbert, J. Scheurer, J. Rando, R. Freedman, T. Korbak, D. Lindner, P. Freire, et al. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217,

    arXiv

Show all 15 references
  1. [2019]

    W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang, Z. Dong, et al. A survey of large language models.arXiv preprint arXiv:2303.18223, 1(2),

    arXiv

  2. [2020]

    Schroeder and Z

    K. Schroeder and Z. Wood-Doughty. Can you trust llm judgments? reliability of llm-as-a-judge.arXiv preprint arXiv:2412.12509,

    arXiv

  3. [2021]

    Z. Liu, P. Wang, R. Xu, S. Ma, C. Ruan, P. Li, Y. Liu, and Y. Wu. Inference-time scaling for generalist reward modeling.arXiv preprint arXiv:2504.02495,

  4. [2022]

    J. Tang, J. Zhang, Z. Tian, X. Feng, L. Wang, and X. Chen. Hf4rec: Human-like feedback-driven optimization framework for explainable recommendation.arXiv preprint arXiv:2504.14147,

    arXiv

  5. [2023]

    B. Chen, X. Gao, C. Hu, P. Yu, H. Zhang, and B.-K. Bao. Reasongrm: Enhancing generative reward models through large reasoning models.arXiv preprint arXiv:2506.16712,

    arXiv

  6. [2024]

    J. Deng, S. Wang, K. Cai, L. Ren, Q. Hu, W. Ding, Q. Luo, and G. Zhou. Onerec: Unifying re- trieve and rank with generative recommender and iterative preference alignment.arXiv preprint arXiv:2502.18965,

    arXiv

  7. [2025]

    J. Ye, Y. Wang, Y. Huang, D. Chen, Q. Zhang, N. Moniz, T. Gao, W. Geyer, C. Huang, P.-Y. Chen, et al. Justice or prejudice? quantifying biases in llm-as-a-judge.arXiv preprint arXiv:2410.02736,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.