Pith. sign in

REVIEW 3 major objections 5 minor 43 references

RecoReward shows that a recommender's score can be used as a training-time reward to make an MLLM generate live-stream descriptions that improve downstream recall, with no user input at serving time.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 01:09 UTC pith:FO26D62Y

load-bearing objection A genuinely useful industrial recipe—training-time recommender reward, content-only serving—but the non-target subtraction likely removes exposure artifacts rather than shared preference, and the headline numbers are tuned on the test set. the 3 major comments →

arxiv 2607.25901 v1 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 learningdescription generationlive-stream recommendationtwo-tower modeluser-selective rewardcontent-only inferencerecall evaluation
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.

RecoReward tries to settle a design tension in multimodal recommendation: content-only description generation ignores user behavior, while user-conditioned generation needs user histories at serving time. The paper's proposal is to inject user information only through a reward during training. A frozen two-tower recommender scores candidate descriptions by how well they match historically engaged users minus a background estimate from non-target users, producing a user-selective signal. Reinforcement learning then teaches a multimodal LLM to write descriptions that score well, and at serving time the model sees only frames and speech and emits one shared description. If the results hold, item-side semantic features can be optimized for recommendation without personalizing generation per user.

Core claim

RecoReward claims a behavior-trained two-tower recommender can act as a training-time critic, telling an MLLM which content-valid description is most useful for recommendation. The Recommender Affinity Score contrasts the mean embedding of historically engaged users with a scaled mean of observational non-target users, yielding a user-selective scalar reward. After reinforcement learning, the policy maps frames and ASR alone to a single shared description. The paper reports 31.7–40.4% relative gains over its base model on all seven offline recall metrics and positive one-week online A/B changes.

What carries the argument

The load-bearing object is the frozen two-tower scorer plus the Recommender Affinity Score (RAS). The scorer maps user histories and generated text into one normalized space; RAS, RASλ(y;a) = (mₐ⁺ − λ mₐ⁻)ᵀ v_y, subtracts the non-target mean embedding from the target mean embedding so the reward rewards compatibility with engaged users while penalizing affinity shared with the broader active population. This scalar is mapped into [0,1], combined with a JSON-format reward, and fed into group-relative policy optimization. The work it does is to convert user behavior into a per-description training signal that never enters the generator input.

Load-bearing premise

That non-target users—active users with no observed positive behavior—are a fair stand-in for broadly shared affinity; if they are mostly unexposed rather than disinterested, the subtracted background carries exposure bias and the user-selective direction is confounded.

What would settle it

Compare the corrected direction (m⁺ − λm⁻) with one built from exposed negatives (users shown the stream but not engaged). If candidate rankings under RAS change materially, or if recall gains vanish when the DSSM is trained with exposure-aware negatives, the background-subtraction interpretation is wrong.

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

If this is right

  • Content-only serving is preserved: one generated description per item is cached and reused for all users, because user information only shaped the reward during training.
  • Offline recall improves on all seven metrics, with relative gains of 31.7–40.4% over the base 9B model when descriptions are used as item features in freshly trained two-tower recall models.
  • Non-target subtraction is load-bearing for the reward: ablations show λ=0 (no subtraction) is worse, and λ=2 gives the strongest retained policy.
  • Reward-sharpness trade-offs appear: a smaller target-user cap (M=25) outperforms larger caps, and rollout count G=12 beats G=4/G=16 in the reported settings.
  • A one-week online A/B test reports +0.265% key-page effective-user penetration and +0.791%/+0.740% outflow exposure/users.

Where Pith is reading between the lines

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

  • If non-target users are mostly unexposed rather than genuinely disinterested, the subtraction may be removing exposure bias rather than shared preference; re-running the reward with exposed negatives or propensity-weighted controls would separate these.
  • The same training-reward scheme could transfer to other item domains (e-commerce listings, video captions) where a frozen retrieval tower exists, and even to multiple reusable descriptions per item assigned to different user segments.
  • Because RAS does not check factuality, stronger optimization may favor details that separate user groups but are unsupported by content; adding a grounding constraint is a natural testable extension.
  • The paper's matched-tower evaluation re-trains a new DSSM per generator; a stronger test would re-train the full industrial ranking stack or test cold-start generalization.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces RecoReward, a reinforcement-learning framework for training an MLLM to generate a single structured description of a live stream from frames and ASR, using a frozen two-tower recommender as a scoring function. A new reward, the Recommender Affinity Score (RAS), contrasts mean compatibility with a target set of historically engaged users against mean compatibility with an observational non-target set, with a subtraction coefficient λ. The policy is trained with GRPO/DAPO-style group-relative updates; at serving time only content is used. Offline, the authors report that RecoReward-9B improves all seven recall metrics by 31.7–40.4% relative to Qwen3.5-9B in a matched-tower evaluation where DSSMs are retrained on generated descriptions. A one-week online A/B test reports increases in key-page effective-user penetration and outflow. The paper also includes a temporally separated behavioral analysis arguing that the historical target center predicts future target users and that non-target subtraction improves target specificity.

Significance. If the results hold, the main contribution is a concrete demonstration that downstream recommender behavior can be incorporated at training time while keeping inference content-only, with large offline recall gains and positive online signals. The paper is unusually candid about limitations: RAS is explicitly a proxy, non-target users are not exposed negatives, and each ablation uses one actor run (Sections 4.3.1, 5.3.2, Appendix D). These candid statements are a strength but also define the burden of proof. The central claim—that the gain is attributable to user-selective reward design—is not yet established because the non-target contrast may be confounded by exposure, and the headline configuration is selected on the same benchmark used for reporting final numbers.

major comments (3)
  1. [§4.3.1, Eq. (7)–(10), Appendix D] Non-target users are defined as 'active users with no observed positive interaction' and Appendix D concedes they are 'not labeled as exposed negatives.' If these users are predominantly unexposed to the author's stream, then S−(y;a) measures exposure/popularity rather than shared affinity. Subtracting m− then removes a global exposure direction, and the λ=2 advantage in Table 4 may reflect exploitation of scorer exposure patterns rather than user-selective semantics. The Section 3.2 future-user ROC-AUC/AP validation uses the same non-target sampling, so it cannot distinguish the two explanations. The authors should test with exposed negatives (users who saw the stream but did not engage) or with propensity/exposure controls, and report whether the corrected direction remains target-specific under that definition.
  2. [§5.3, Tables 4–6; Table 3] The reported RecoReward-9B configuration (λ=2, M=25, G=12) is the best-performing setting on the same seven-metric next-day benchmark that Table 3 uses for the headline comparison. The paper does not describe a validation split or selection protocol that would protect against selection bias; 'the checkpoint retained under the fixed selection protocol' (Section 5.3.1) appears to mean selection on the test set. This inflates the 31.7–40.4% gains relative to untuned baselines and makes RQ2's trends hard to interpret. Use a chronological validation split for hyperparameter selection, or report both a default configuration and the selected configuration on a truly held-out test.
  3. [§5.3.2, §5.1.3, Appendix D] Each ablation configuration uses one actor run, and the three-seed standard deviations in Table 3 only cover evaluator initialization. Hence the differences between adjacent λ values (Table 4) or G=12 vs 16 (Table 5) may be within training noise, yet they are used to support causal-sounding interpretations ('aggressive suppression produces a more user-selective reward ordering,' Section 5.3.1). Report multiple actor seeds at a few key settings or provide appropriate significance tests before claiming a user-selectivity pattern.
minor comments (5)
  1. [§5.3.1 and §5.3.3] The prose values for NDCG@128, HR@128, and MRR are off by a factor of 100 relative to Tables 4 and 6 (e.g., '0.014118' vs 1.4118%, '0.004417' vs 0.4417%). Please make the percentage units consistent.
  2. [§3.2] Define the 'global user center' explicitly and state whether the target/non-target centers are normalized before the cosine computations. Without this, the reported cosines (0.7752, 0.9255, 0.0960) are not reproducible.
  3. [§5.1.1] Specify the size of the 'shared candidate universe' and how negative items are sampled. This is needed to interpret the absolute HR/NDCG values.
  4. [§5.4, Table 7] Report confidence intervals and significance for the online A/B metrics. The effect sizes (0.265%, 0.791%, 0.740%) are small and could be within daily platform noise.
  5. [Appendix B] The green/red bold annotations described in the text are not visible in the tables; please render them or replace with explicit annotations in the table.

Circularity Check

0 steps flagged

No significant circularity: RAS is a training-time proxy, not the evaluation metric; offline evaluation uses retrained DSSMs on temporally held-out data and an online A/B test provides an external anchor.

full rationale

The derivation chain is not circular. RecoReward optimizes a training-time proxy RAS (Eq. 10) computed with a frozen two-tower scorer, whereas the offline claim is evaluated by retraining new DSSMs from scratch on the generated description distribution and measuring next-day recall on items and users not used in policy training (Section 5.1.3). The reward and the evaluation are not the same function: the former is an inner product with a fixed scorer over historical target/non-target centers, while the latter is ranking performance of a re-fit model on temporally held-out interactions. The paper explicitly disclaims that RAS is "a behavior-derived proxy rather than an equivalent definition of U_rec" (Section 4.1), and Appendix D concedes that non-target users are "not labeled as exposed negatives." Those are validity/exposure concerns, not definitional circularity. The only overlapping self-citation (SARM [34]) appears in related work and is not load-bearing. The online A/B test provides an external behavioral anchor. The fact that offline evaluation shares the two-tower behavior-trained paradigm with the reward scorer is a legitimate generalization risk, but no equation or fitted parameter reduces the reported recall to the RAS objective by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

RAS depends on the geometry of a behavior-trained two-tower space and on the interpretation of historical target/non-target user sets; no new physical or model-level entities are postulated. The main free parameters (λ, M, G) are tuned on the same benchmark used for headline results.

free parameters (4)
  • λ (non-target subtraction coefficient) = 2 (best of {0, 0.5, 1, 2} on the same recall benchmark)
    Controls how much non-target affinity is subtracted in RAS; chosen because it maximizes the reported offline recall metrics.
  • M (reward user cap) = 25 (best of {25, 50, 100, 200})
    Number of target and non-target users averaged in RAS; selected on the evaluation benchmark.
  • G (rollout count) = 12 (best of {4, 8, 12, 16})
    Number of sampled descriptions per input in GRPO/DAPO; selected on the evaluation benchmark.
  • α (semantic reward weight) = 0.9
    Hand-set balance between semantic RAS reward and format reward; not ablated.
axioms (5)
  • domain assumption A behavior-trained two-tower DSSM inner product s(u_i, y) is a valid measure of user-description compatibility for recall.
    Introduced in Eqs. (3)–(5); the entire RAS reward is built on this embedding geometry.
  • domain assumption Historical engaged users are a valid proxy for future target users.
    Used in Section 4.3.1 to define U+; empirically supported in Section 3.2 but not causal.
  • domain assumption Active users with no observed positive interaction provide a background estimate of affinity shared broadly across users.
    Section 4.3.1 and Appendix D; unexposed users may confound this assumption.
  • domain assumption Retraining DSSMs on generated descriptions and measuring next-day recall is a valid measure of downstream recommendation utility.
    Section 5.1.3; shares the same two-tower objective family as the scorer used as reward, so offline gains are partly self-referential.
  • standard math GRPO/DAPO group-relative optimization with a scalar reward improves the policy toward higher-utility descriptions.
    Relies on standard RL literature [26,36]; the paper does not prove convergence or robustness for this specific reward.

pith-pipeline@v1.3.0-alltime-deepseek · 18324 in / 13599 out tokens · 123603 ms · 2026-08-01T01:09:01.031240+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.

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