Pith. sign in

REVIEW 4 major objections 4 minor 43 references

Recommender behavior, turned into a training reward, can teach a multimodal language model to write one shared item description that improves downstream recall while serving stays user-free.

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

2026-08-04 03:21 UTC pith:FO26D62Y

load-bearing objection Useful training-time reward for content-only description generation, but the headline offline gains are over-fit to the test set and need a validation split and more seeds. the 4 major comments →

arxiv 2607.25901 v2 pith:FO26D62Y submitted 2026-07-28 cs.IR

RecoReward: Recommender-Guided Multimodal Description Generation for Recommendation

Guohong Mu , Yueyang Liu , Jiangxia Cao , Changxin Lao , Zijie Zhuang , Yuhui Zhang , Jiaqi Feng , Ruochen Yang
show 3 more authors
Shuang Yang Zhaojie Liu Qibin Hou
This is my paper
classification cs.IR
keywords multimodal recommendationreinforcement learningrecommender rewardcontent-only inferencetwo-tower matchinglive-stream recommendationitem description generationrecall metrics
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 sets out to settle a trade-off in multimodal recommendation: item descriptions written from content alone ignore the downstream user signals that decide which items are relevant, while user-conditioned generation needs user histories at every request. RecoReward is the claim that this trade-off can be broken by moving user signal into the training reward instead. The authors show that a frozen two-tower recommender, trained on behavior, can score any candidate description against a historical target-user population relative to a non-target population, yielding a user-selective scalar reward. Policies trained with this reward generate a single shared description at serving time, and the 9-billion-parameter variant improves all seven measured recall metrics over its base model by 31.7–40.4%, with a positive one-week online A/B test. If true, the finding matters because it suggests recommendation utility can be baked into a reusable item representation without per-user inference costs.

Core claim

RecoReward claims that a behavior-trained two-tower recommender can serve as a training-time interface between user behavior and generated text, letting a multimodal language model learn which semantics matter for downstream matching without ever seeing a user at inference. The central construction is the Recommender Affinity Score (RAS), which averages compatibility of a candidate description with historically engaged users and subtracts a scaled average compatibility with observational non-target users—active users with no observed positive interaction with the item's author. This contrast removes affinity shared across the broad user population, leaving a target-selective reward. The poli

What carries the argument

The central object is the Recommender Affinity Score (RAS), a scalar reward defined as the inner product between a generated description's embedding and the difference of two empirical user-center embeddings: the mean embedding of historically engaged 'target' users minus λ times the mean embedding of observational 'non-target' users. The subtraction is meant to cancel affinity shared across broadly active users, leaving a target-selective signal. This score is computed by a frozen behavior-trained two-tower matching model that maps users and descriptions into a shared space, so the user signal never enters the language model's input. Group-relative policy optimization uses RAS as a sequence

Load-bearing premise

The reward gains rest on the assumption that observational non-target users—active users with no observed positive interaction—honestly estimate affinity shared across all users, so subtracting their mean representation removes the shared component without removing target-specific signal; if exposure, item popularity, or prior recommender policy drives non-target affinity instead, the subtraction removes the wrong component.

What would settle it

Run the same training with a non-target set of users who were exposed to the item but chose not to engage (true negatives). If the recall gain over the target-only baseline disappears or reverses, the claimed shared-affinity subtraction is actually removing exposure or popularity signal. If the gain persists, RAS is doing the work it claims.

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

If this is right

  • With recommender-derived rewards, a multimodal LLM can produce a single, shareable item description that internalizes downstream user preference; serving stays content-only, and per-user prompts are never needed.
  • The 9B policy trained this way beats all evaluated baselines, including larger and commercial ones, across every measured recall metric, so the training signal matters more than raw model scale.
  • The analysis suggests the target–non-target contrast, not just the target signal, is what drives the gain: the corrected direction improves future-user ROC-AUC and increases the target–non-target margin.
  • The design applies any time a recommender can score a user against a generated description, so the recipe extends to other item types and ranking surfaces, though the paper only tests live streams.

Where Pith is reading between the lines

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

  • If the non-target subtraction is truly removing shared affinity, the same reward design could work in other domains—e-commerce, video, news—but the paper does not test this, and non-target definitions would need to match each platform's exposure policy.
  • Because the reward is a proxy for downstream utility, stronger optimization may push the generator toward reward-hacking descriptions that separate user groups without being factually grounded; the paper notes this and calls for independent factuality evaluation.
  • A testable refinement: replace observational non-targets with exposed non-engagers or control users. If RAS gains persist under counterfactual-accurate non-targets, the shared-affinity mechanism is verified; if they vanish, the current gain may come from popularity cancellation.

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 / 4 minor

Summary. The paper proposes RecoReward, a training-time framework that uses a frozen two-tower recommender scorer to assign a sequence-level reward, the Recommender Affinity Score (RAS), to candidate MLLM-generated item descriptions. RAS contrasts compatibility with historically engaged target users against affinity shared with observational non-target users. The policy is optimized with group-relative RL (GRPO/DAPO-style) and, at serving time, generates a single shared description from content alone. Offline experiments report that RecoReward-9B improves over its Qwen3.5-9B baseline by 31.7–40.4% across seven recall metrics, and a one-week online A/B test reports positive changes in key-page and outflow metrics. The paper also includes a behavioral analysis of target/non-target representation geometry, hyperparameter ablations, and an explicit limitations section.

Significance. The central idea—using downstream recommender behavior as a training-time reward while preserving content-only inference—is practically valuable and well-motivated. The behavioral analysis in Section 3 is a useful empirical contribution, and the paper is commendably explicit about the observational nature of its rewards and the single-actor-run limitation. If the offline gains survive a proper validation protocol and generalize beyond the same two-tower objective family used for both reward and evaluation, this would be a solid contribution. However, the current evaluation protocol makes the headline numbers difficult to interpret, and the same-objective reward/evaluation design leaves open the question of whether the improvements are specific to one recommender architecture.

major comments (4)
  1. [§5.3 / Tables 4–6 / §5.1.1 / Appendix D] The central offline claim (Table 3) is weakened by post-hoc hyperparameter selection. Section 5.1.1 defines only a chronological train/test split; Section 5.3 selects λ=2, G=12, and M=25 by comparing configurations on the same final-day test set used for Table 3. Each configuration is trained once (Appendix D: 'each ablation setting uses one actor run'), and the reported seeds vary only evaluator initialization. The headline 31.7–40.4% improvement is therefore a maximum over a grid evaluated on the test set. Please add a held-out validation split for all hyperparameter selection, report multiple actor runs or bootstrap CIs, and show that the selected configuration's gains hold on a truly held-out test set.
  2. [§4.2 vs §5.1.3] The reward and the offline evaluation use the same contrastive two-tower family. RAS is computed by a frozen two-tower DSSM (Section 4.2), and the evaluation retrains DSSMs of the same architecture and objective on each generator's output (Section 5.1.3). The policy is thus optimized and measured in the same objective family, which conflates 'better for the reward proxy' with 'better for downstream recommendation.' At minimum, evaluate RecoReward descriptions with an alternative retrieval model (e.g., a GRU4Rec/NextItNet-style encoder or a different two-tower variant) or argue that the DSSM is the production serving model. This is needed to support the claim that gains generalize beyond the reward's own inductive bias.
  3. [§4.3.1 / Appendix D] The RAS contrast assumes that non-target users form an adequate baseline for affinity shared broadly across users. As stated in Section 4.3.1 and Appendix D, non-target users are active users with no observed positive interaction with the author—not exposed negatives, dislike labels, or counterfactual users. Exposure patterns, item popularity, and prior recommender policies can affect both target and non-target groups, so the subtraction may remove the wrong component. The behavioral analysis (Figure 3b) shows the non-target center is closer to the global center (cosine 0.9255 vs 0.7752), consistent with shared affinity, but it does not establish that the residual is target-specific. Please provide an exposure-aware robustness check (e.g., comparing random active users vs. observed non-interacted users, or controlling for exposure) or temper the 'user-selective' interpretation.
  4. [§5.4 / Table 7] The one-week online A/B test reports only point estimates (+0.265%, +0.791%, +0.740%) with no confidence intervals, significance tests, or day-level variance. Given the small absolute effect and short duration, these numbers do not by themselves establish a statistically reliable gain. Please report uncertainty (CIs, p-values, or day-level breakdowns) or explicitly label the results as directional.
minor comments (4)
  1. [§5.1.3] With three evaluator seeds, reporting 'population standard deviation' is a small-sample estimate; consider reporting standard errors or bootstrap intervals.
  2. [§5.2 / Table 3] No significance tests are reported between RecoReward and baselines; for some metrics the standard deviations overlap substantially (e.g., RecoReward-4B), making it hard to assess whether differences are meaningful.
  3. [§5.3] The 'fixed selection protocol' for retaining checkpoints is never defined. Please state explicitly what criterion is used (e.g., best NDCG@128 on a validation split) so the ablation results are reproducible.
  4. [Figure 3 / Table 2] Several quantitative claims in Section 3.2 (e.g., cosine values 0.7752, 0.9255, 0.0960) appear only in prose. Include per-seed values or a table to show stability across the three evaluator seeds.

Circularity Check

1 steps flagged

Offline gain is the best hyperparameter configuration selected on the same test set that is then reported as the headline result.

specific steps
  1. fitted input called prediction [Sections 5.1.1, 5.2, 5.3, Appendix D]
    "We use the first six days for training and hold out the final day as the test set for offline recall evaluation (5.1.1). λ=2 yields the strongest retained policy across all measured recall metrics (5.3.1). RecoReward-9B leads all seven metrics ... with gains of 31.7–40.4% (5.2). each ablation setting uses one actor run (Appendix D)."

    The same final-day test set is used both to select the reward hyperparameters (λ=2, G=12, M=25) in Tables 4–6 and to report the headline comparison in Table 3. The reported RecoReward-9B improvement is therefore the maximum over a small grid evaluated on the test set, not the out-of-sample performance of a pre-specified configuration. With one actor run per setting and standard deviations that only reflect evaluator seed, the 31.7–40.4% gain is statistically inflated by selection (winner's curse), making the central quantitative claim a fitted value rather than an independent prediction.

full rationale

The derivation of RAS from the behavioral analysis is not itself circular: the analysis uses temporally separated future users and fixed tower parameters, and the paper repeatedly states that RAS is a proxy, not an equivalent definition of Urec. The shared two-tower family between the reward and the evaluation DSSMs is a limitation (the policy may optimize specifically for DSSM-like compatibility), but it is not an equation-level reduction because the evaluation DSSMs are retrained from scratch on the generated description distribution; the paper candidly acknowledges this in Appendix D. The primary circularity concern is the evaluation protocol: the test set used for hyperparameter selection is the same test set used for the headline recall numbers. This makes the reported 31.7–40.4% improvement a selection-maximized estimate rather than an unbiased prediction, although the online A/B test provides some independent, if system-level, evidence that the method has real-world value.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central method rests on a compact set of empirical proxies and tuning choices. The main free parameters are the RAS contrast weight, rollout count, user cap, reward mixing weight, and the unreported behavior weights in tower training. The key axioms are the historical-target proxy, the non-target background estimate, the validity of the two-tower inner product, and the transferability of proxy-reward optimization.

free parameters (5)
  • lambda (non-target subtraction coefficient) = 2
    Chosen from Table 4 ablation as the setting with best offline recall metrics; controls the strength of shared-affinity subtraction in RAS (Eq. 10).
  • G (rollout count) = 12
    Chosen from Table 5 ablation; number of sampled descriptions per input used for group-relative advantage estimation.
  • M (reward user cap) = 25
    Chosen from Table 6 ablation; number of target and non-target users averaged to form m+ and m- in Eq. 7.
  • alpha (semantic reward weight) = 0.9
    Set in Section 5.1.2; semantic vs format reward mixture in Eq. 12.
  • behavior weights w_b
    Weights for room entry, long viewing, following, commenting, liking, and gifting in the tower objective (Eq. 5) are not reported; they shape the frozen scorer the entire method relies on.
axioms (5)
  • domain assumption Historical target users (engaged with the author during a construction window) are a valid proxy for future target users.
    Section 3.2 reports ROC-AUC 0.7603 for the proxy on future users, so it is empirically motivated, but the paper states it is not a complete or causal definition of user preference (Section 4.3.1).
  • domain assumption Observational non-target users (active users without observed positive interaction) provide an estimate of shared affinity that can be subtracted.
    Section 4.3.1 and Appendix D: non-target users are not exposed negatives or dislike labels; exposure and popularity confounds may affect both groups, so the subtraction's validity is assumed.
  • domain assumption Inner product in the frozen behavior-trained two-tower space measures recommendation utility of descriptions.
    Eqs. 3-5 define matching probability from positive interactions; both RAS and the offline evaluation rely on this space's validity for unseen generated descriptions.
  • domain assumption Optimizing a proxy reward with GRPO/DAPO transfers to genuine downstream recommendation utility without catastrophic overoptimization.
    The paper cites reward overoptimization work (Section 2.3) and in Appendix D says stronger optimization may amplify proxy-specific patterns; the evaluation with retrained DSSMs is the only check.
  • domain assumption The item description tower excludes author IDs, so score differences among descriptions of the same stream reflect semantics rather than identity.
    Section 4.2 states author identifiers are excluded from v_y; if identity leaked through text, the reward could game author popularity.
invented entities (1)
  • Recommender Affinity Score (RAS) no independent evidence
    purpose: Sequence-level reward for RL training of the content-only description policy (Eq. 10).
    RAS is a linear contrast of learned user-center embeddings; its validity is demonstrated only through the paper's offline/online metrics, and the paper itself calls it a behavior-derived proxy, not an independently measurable quantity.

pith-pipeline@v1.3.0-alltime-deepseek · 18316 in / 16060 out tokens · 206081 ms · 2026-08-04T03:21:46.838489+00:00 · methodology

0 comments
read the original abstract

Multimodal large language models (MLLMs) can convert multimodal item content into structured descriptions used as semantic features for recommendation. Conventional content-only generation, however, cannot use downstream user signals to determine which semantics should be emphasized. Recent user-conditioned methods incorporate these signals through user histories or profiles, but they require user information at inference and make generation user-dependent. In this paper, we introduce RecoReward, which instead uses behavior-derived rewards during training and preserves content-only inference. To instantiate this idea in live-stream recommendation, we treat historically engaged users as a proxy for future target users and use observational non-target users to estimate affinity shared broadly across users. The Recommender Affinity Score (RAS) contrasts these signals to provide user-selective feedback for reinforcement learning, allowing the learned policy to generate a single shared description without user inputs. In our offline benchmark, RecoReward-9B outperforms its Qwen3.5-9B baseline and all other evaluated models across seven recall metrics. Online A/B testing also shows performance gains. These results show that RecoReward trains the MLLM to produce item features that benefit downstream recommendation while retaining content-only serving.

Review history (2 revisions) →

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

43 extracted references · 8 linked inside Pith

  1. [1]

    Qwen3-VL technical report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, et al. Qwen3-VL technical report. arXiv preprint arXiv:2511.21631, 2025. 8

    Pith/arXiv arXiv 2025

  2. [2]

    TALLRec: An effective and efficient tuning framework to align large language model with recommendation

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. TALLRec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems, pages 1007–1014, New York, NY , USA, 2023. Association for Computing Machinery. 3

    2023

  3. [3]

    VRAgent-R1: Boosting video recom- mendation with MLLM-based agents via reinforcement learning

    Siran Chen, Boyu Chen, Chenyun Yu, Yuxiao Luo, Ouyang Yi, Lei Cheng, Chengxiang Zhuo, Zang Li, and Yali Wang. VRAgent-R1: Boosting video recom- mendation with MLLM-based agents via reinforcement learning. arXiv preprint arXiv:2507.02626, 2025. 1, 3

    Pith/arXiv arXiv 2025

  4. [4]

    Describe what you see with multimodal large language models to enhance video recommendations

    Marco De Nadai, Andreas Damianou, and Mounia Lal- mas. Describe what you see with multimodal large language models to enhance video recommendations. InProceedings of the Nineteenth ACM Conference on Recommender Systems, pages 1159–1163, New York, 9 NY , USA, 2025. Association for Computing Machinery. 1, 3

    2025

  5. [5]

    Scaling laws for reward model overoptimization

    Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. InProceedings of the 40th International Conference on Machine Learn- ing, volume 202 ofProceedings of Machine Learning Research, pages 10835–10866, Honolulu, Hawaii, USA,

  6. [6]

    Recommendation as language pro- cessing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5)

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. Recommendation as language pro- cessing (RLP): A unified pretrain, personalized prompt & predict paradigm (P5). InProceedings of the 16th ACM Conference on Recommender Systems, pages 299– 315, New York, NY , USA, 2022. Association for Com- puting Machinery. 3

    2022

  7. [7]

    Gemini 3.1 pro model card

    Google DeepMind. Gemini 3.1 pro model card. Google DeepMind model card, Feb. 2026. Published February 19, 2026. 8

    2026

  8. [8]

    LGMRec: Local and global graph learning for multimodal recommendation.Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, 38(8):8454–8462, 2024

    Zhiqiang Guo, Jianjun Li, Guohui Li, Chaoyang Wang, Si Shi, and Bin Ruan. LGMRec: Local and global graph learning for multimodal recommendation.Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, 38(8):8454–8462, 2024. 3

    2024

  9. [9]

    VBPR: Visual bayesian personalized ranking from implicit feedback

    Ruining He and Julian McAuley. VBPR: Visual bayesian personalized ranking from implicit feedback. Proceedings of the AAAI Conference on Artificial Intel- ligence, 30(1):144–150, 2016. 1, 3

    2016

  10. [10]

    LightGCN: Simplifying and powering graph convolution network for recommen- dation

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong- dong Zhang, and Meng Wang. LightGCN: Simplifying and powering graph convolution network for recommen- dation. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in In- formation Retrieval, pages 639–648, New York, NY , USA, 2020. Association for Computing Machinery. 8

    2020

  11. [11]

    Session-based recommen- dations with recurrent neural networks

    Bal´azs Hidasi, Alexandros Karatzoglou, Linas Bal- trunas, and Domonkos Tikk. Session-based recommen- dations with recurrent neural networks. In4th Inter- national Conference on Learning Representations, San Juan, Puerto Rico, 2016. OpenReview.net. 8

    2016

  12. [12]

    Learning deep structured semantic models for web search using clickthrough data

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. Learning deep structured semantic models for web search using clickthrough data. InProceedings of the 22nd ACM International Con- ference on Information and Knowledge Management, pages 2333–2338, New York, NY , USA, 2013. Associa- tion for Computing Machinery. 8

    2013

  13. [13]

    Large language models meet collaborative filtering: An efficient all-round LLM-based recommender system

    Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. Large language models meet collaborative filtering: An efficient all-round LLM-based recommender system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1395– 1406, New York, NY , USA, 2024. Association for Com- puting Ma...

    2024

  14. [14]

    Pick-a-pic: An open dataset of user preferences for text-to-image gen- eration

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image gen- eration. InAdvances in Neural Information Processing Systems, volume 36, pages 36652–36663, Red Hook, NY , USA, 2023. Curran Associates, Inc. 3

    2023

  15. [15]

    LLaRA: Large language-recommendation assistant

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. LLaRA: Large language-recommendation assistant. In Proceedings of the 47th International ACM SIGIR Con- ference on Research and Development in Information Retrieval, pages 1785–1795, New York, NY , USA, 2024. Association for Computing Machinery. 3

    2024

  16. [16]

    ReLLa: Retrieval-enhanced large language models for lifelong sequential behavior com- prehension in recommendation

    Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. ReLLa: Retrieval-enhanced large language models for lifelong sequential behavior com- prehension in recommendation. InProceedings of the ACM Web Conference 2024, pages 3497–3508, New York, NY , USA, 2024. Association for Computing Ma- chinery. 1, 3

    2024

  17. [17]

    Rec-R1: Bridg- ing generative large language models and user-centric recommendation systems via reinforcement learning

    Jiacheng Lin, Tian Wang, and Kun Qian. Rec-R1: Bridg- ing generative large language models and user-centric recommendation systems via reinforcement learning. arXiv preprint arXiv:2503.24289, 2025. 1, 3

    arXiv 2025

  18. [18]

    AlignRec: Aligning and training in multimodal recommendations

    Yifan Liu, Kangning Zhang, Xiangyuan Ren, Yanhua Huang, Jiarui Jin, Yingjie Qin, Ruilong Su, Ruiwen Xu, Yong Yu, and Weinan Zhang. AlignRec: Aligning and training in multimodal recommendations. InProceed- ings of the 33rd ACM International Conference on Infor- mation and Knowledge Management, pages 1503–1512, New York, NY , USA, 2024. Association for Comp...

    2024

  19. [19]

    Towards com- prehensible recommendation with large language model fine-tuning

    Yunze Luo, Yinjie Jiang, Gaode Chen, Xinghua Zhang, Jun Zhang, Jian Liang, and Kaigui Bian. Towards com- prehensible recommendation with large language model fine-tuning. arXiv preprint arXiv:2508.07595, 2025. 1, 3

    Pith/arXiv arXiv 2025

  20. [20]

    Openai GPT-5 system card

    OpenAI. Openai GPT-5 system card. arXiv preprint arXiv:2601.03267, 2025. 8

    Pith/arXiv arXiv 2025

  21. [21]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training lan- guage models to follow instructions with human feed...

    2022

  22. [22]

    Qwen3.5: Towards native multimodal agents

    Qwen Team. Qwen3.5: Towards native multimodal agents. Official technical blog, Feb. 2026. Official model release. 8

    2026

  23. [23]

    Manning, Stefano Ermon, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D. Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. InAdvances in Neural Infor- mation Processing Systems, volume 36, pages 53728– 10 53741, Red Hook, NY , USA, 2023. Curran Associates, Inc. 3

    2023

  24. [24]

    Representation learning with large language models for recommendation

    Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. Representation learning with large language models for recommendation. InProceedings of the ACM Web Con- ference 2024, pages 3464–3475, New York, NY , USA,

    2024

  25. [25]

    Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel

    Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. Self-critical sequence training for image captioning. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7008–7024, Honolulu, HI, USA,

  26. [26]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  27. [27]

    Minimum risk train- ing for neural machine translation

    Shiqi Shen, Yong Cheng, Zhongjun He, Wei He, Hua Wu, Maosong Sun, and Yang Liu. Minimum risk train- ing for neural machine translation. InProceedings of the 54th Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), pages 1683–1692, Berlin, Germany, 2016. Association for Computational Linguistics. 3

    2016

  28. [28]

    Christiano

    Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea V oss, Alec Radford, Dario Amodei, and Paul F. Christiano. Learning to summarize with human feedback. InAdvances in Neural Informa- tion Processing Systems, volume 33, pages 3008–3021, Red Hook, NY , USA, 2020. Curran Associates, Inc. 3

    2020

  29. [29]

    MISSRec: Pre- training and transferring multi-modal interest-aware se- quence representation for recommendation

    Jinpeng Wang, Ziyun Zeng, Yunxiao Wang, Yuting Wang, Xingyu Lu, Tianxiang Li, Jun Yuan, Rui Zhang, Hai-Tao Zheng, and Shu-Tao Xia. MISSRec: Pre- training and transferring multi-modal interest-aware se- quence representation for recommendation. InProceed- ings of the 31st ACM International Conference on Mul- timedia, pages 6548–6557, New York, NY , USA, 20...

    2023

  30. [30]

    In- ternVL3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency

    Weiyun Wang, Zhangwei Gao, Lixin Gu, et al. In- ternVL3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025. 8

    Pith/arXiv arXiv 2025

  31. [31]

    GRCN: Graph-refined convolu- tional network for multimedia recommendation with implicit feedback

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, and Tat-Seng Chua. GRCN: Graph-refined convolu- tional network for multimedia recommendation with implicit feedback. InProceedings of the 28th ACM Inter- national Conference on Multimedia, pages 3541–3549, New York, NY , USA, 2020. Association for Computing Machinery. 3

    2020

  32. [32]

    MMGCN: Multi- modal graph convolution network for personalized rec- ommendation of micro-video

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong, and Tat-Seng Chua. MMGCN: Multi- modal graph convolution network for personalized rec- ommendation of micro-video. InProceedings of the 27th ACM International Conference on Multimedia, pages 1437–1445, New York, NY , USA, 2019. Associa- tion for Computing Machinery. 1, 3

    2019

  33. [33]

    Im- ageReward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Im- ageReward: Learning and evaluating human preferences for text-to-image generation. InAdvances in Neural In- formation Processing Systems, volume 36, pages 15903– 15935, Red Hook, NY , USA, 2023. Curran Associates, Inc. 3

    2023

  34. [34]

    SARM: LLM-augmented semantic anchor for end-to-end live-streaming ranking

    Ruochen Yang, Yueyang Liu, Zijie Zhuang, Changxin Lao, Yuhui Zhang, Jiangxia Cao, Jia Xu, Xiang Chen, Haoke Xiao, Xiangyu Wu, Xiaoyou Zhou, Xiao Lv, Shuang Yang, Tingwen Liu, Zhaojie Liu, Han Li, and Kun Gai. SARM: LLM-augmented semantic anchor for end-to-end live-streaming ranking. arXiv preprint arXiv:2602.09401, 2026. 1, 3

    arXiv 2026

  35. [35]

    Multi-view graph convolutional network for mul- timedia recommendation

    Penghang Yu, Zhiyi Tan, Guanming Lu, and Bing-Kun Bao. Multi-view graph convolutional network for mul- timedia recommendation. InProceedings of the 31st ACM International Conference on Multimedia, pages 6576–6585, New York, NY , USA, 2023. Association for Computing Machinery. 1, 3

    2023

  36. [36]

    DAPO: An open-source LLM reinforcement learning system at scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gao- hong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Jin- hua Zhu, Jiaze Chen, Jiangjie Chen, Chengyi Wang, Hongli Yu, Yuxuan Song, Xiangpeng Wei, Hao Zhou, Jingjing Li...

    Pith/arXiv arXiv 2025

  37. [37]

    Jose, and Xiangnan He

    Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M. Jose, and Xiangnan He. A simple convolu- tional generative network for next item recommendation. InProceedings of the Twelfth ACM International Confer- ence on Web Search and Data Mining, pages 582–590, New York, NY , USA, 2019. Association for Computing Machinery. 8

    2019

  38. [38]

    NoteLLM-2: Multimodal large representation models for recommendation

    Chao Zhang, Haoxin Zhang, Shiwei Wu, Di Wu, Tong Xu, Xiangyu Zhao, Yan Gao, Yao Hu, and Enhong Chen. NoteLLM-2: Multimodal large representation models for recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V .1, pages 2815–2826, New York, NY , USA, 2025. Association for Computing Machinery. 3

    2025

  39. [39]

    Mining latent struc- tures for multimedia recommendation

    Jinghao Zhang, Yanqiao Zhu, Qiang Liu, Shu Wu, Shuhui Wang, and Liang Wang. Mining latent struc- tures for multimedia recommendation. InProceedings of the 29th ACM International Conference on Multi- media, pages 3872–3880, New York, NY , USA, 2021. Association for Computing Machinery. 3

    2021

  40. [40]

    A tale of two graphs: Freez- ing and denoising graph structures for multimodal rec- 11 ommendation

    Xin Zhou and Zhiqi Shen. A tale of two graphs: Freez- ing and denoising graph structures for multimodal rec- 11 ommendation. InProceedings of the 31st ACM Inter- national Conference on Multimedia, pages 935–943, New York, NY , USA, 2023. Association for Computing Machinery

    2023

  41. [41]

    Bootstrap latent representations for multi-modal recom- mendation

    Xin Zhou, Hongyu Zhou, Yong Liu, Zhiwei Zeng, Chun- yan Miao, Pengwei Wang, Yuan You, and Feijun Jiang. Bootstrap latent representations for multi-modal recom- mendation. InProceedings of the ACM Web Conference 2023, pages 845–854, New York, NY , USA, 2023. As- sociation for Computing Machinery. 3

    2023

  42. [42]

    Golden Mountains

    Jinguo Zhu, Weiyun Wang, Zhe Chen, et al. In- ternVL3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. 8 12 A Full Generation Prompt Table 8 presents an English translation of the complete Chi- nese prompt used for the reported base-model and policy evaluations. The bracketed plac...

    Pith/arXiv arXiv 2025

  43. [2024]

    Association for Computing Machinery. 1, 3