Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

VRAgent-R1: Boosting Video Recommendation with MLLM-based Agents via Reinforcement Learning

T0 review · 4 major / 5 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 Two cooperating MLLM agents improve both video recommendation and next-video simulation on MicroLens-100k.

desk verdict A plausible user-simulation result whose headline gains hinge on a train/eval user split the paper never states; worth refereeing if that gets fixed. read the letter →

arxiv 2507.02626 v1 pith:Q7BELYXT submitted 2025-07-03 cs.MM

Siran Chen , Boyu Chen , Chenyun Yu , Yuxiao Luo , Ouyang Yi , Lei Cheng , Chengxiang Zhuo , Zang Li
show 1 more author
Yali Wang
This is my paper
classification cs.MM
keywords videorecommendationmultimodallargelanguagemodelsusersimulationreinforcementfine-tuningGRPOchain-of-thoughtreasoningitemperceptionagentMicroLens-100k
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

The paper sets out to show that multimodal LLM agents can strengthen video recommendation by doing two things a recommender cannot easily do itself: describe what a video is actually about, and predict what a real user would choose next. It proposes two collaborating agents: an Item Perception Agent that turns a video's key frames and title into a concise roughly 35-word recommendation caption, and a User Simulation Agent that reasons in chain-of-thought about the user's status and is fine-tuned with GRPO reinforcement learning. On the MicroLens-100k benchmark, the IP-Agent-enhanced recommender gains 4.3% in HR@10 and 6.0% in NDCG@10 over the best prior multimodal baseline, and the trained simulator reaches 64.1% accuracy at next-video selection, ahead of frozen commercial models and SFT baselines. If these results hold, they indicate that RL-trained user simulators can both audit and improve recommender systems in a multimodal domain.

What carries the argument

The machinery is the two-agent loop. The Item Perception Agent implements three steps: key-frame retrieval (uniformly sample 10 frames, use CLIP frame-title similarity to keep the top 3), collaborative multimodal perception (an MLLM reads the frames with the title and uses world knowledge to explain what the title means in context), and recommendation-relevant analysis (rewrite the understanding into a caption of about 35 words with topic tags). The User Simulation Agent turns the same captions, the user's watched-video history, and comments into a chain-of-thought user status, then answers a preference-judgment or next-video-selection question over m+1 candidates. Its policy is trained with GRPO (Group Relative Policy Optimization), a policy update that scores each sampled response relative to the group mean without training a critic model: for each question the model samples a group of G responses with different reasoning chains, assigns each a reward (format reward plus judgment/selection reward), and updates the policy using the group-relative advantage $(r_i - \mathrm{mean}(r))/\mathrm{std}(r)$; training starts on the easier judgment task and proceeds to selection. The captions are what let a text-based LLM agent process long multimodal interaction histories.

What would settle it

Look up the actual user IDs: check whether any of the 1,000 evaluation users, or the videos in their evaluation candidates, appear in the 2,000-user RFT training set; if they do, the 64.1% next-video accuracy and the roughly 45% gain over SFT would not be trustworthy. A clean rerun with an explicitly disjoint split and the same 2,000-user training budget would settle whether the reported user-simulation gains persist.

Watch

Extended reading notes

Core claim

The central claim is that a two-agent pipeline can jointly improve item modeling and user modeling for video recommendation. The IP Agent selects the top-3 frames by CLIP similarity to the title, has an MLLM interpret those frames together with the title to resolve ambiguous semantics, and rewrites the result into a short recommendation-oriented caption; feeding these captions into SASRec lifts the best prior multimodal baseline from 0.0953 to 0.0994 in HR@10 and from 0.0517 to 0.0548 in NDCG@10 on MicroLens-100k. The US Agent consumes the same captions along with the user's watch and comment history, produces a chain-of-thought user status, and is optimized by GRPO against rewards that check the final judgment or next-video selection against ground truth; this yields 71.5% preference-judgment accuracy and 64.1% next-video selection accuracy on 1,000 cold-start users, roughly 45% above the MLLM-MSR SFT baseline. The paper also reports that simulated 'like' feedback can rerank the recommender's candidate list to slightly better accuracy, while simulated 'dislike' feedback hurts.

Load-bearing premise

The central empirical claims rest on the assumption that the 1,000 cold-start users used to evaluate the User Simulation Agent are disjoint from the 2,000 users whose histories were used for reinforcement fine-tuning; the paper reports both numbers without ever stating the split, so any overlap would inflate the reported simulation gains.

Editorial extensions

If this is right

  • Adding IP-Agent captions to item representations lifts HR@10 from 0.0953 to 0.0994 (+4.3%) and NDCG@10 from 0.0517 to 0.0548 (+6.0%) over the best prior multimodal baseline on MicroLens-100k, with larger relative gains on cold-start users.
  • GRPO-trained US Agent reaches 71.5% preference-judgment accuracy and 64.1% next-video selection accuracy on 1,000 cold-start users, about 45% higher than the MLLM-MSR SFT baseline while training on fewer users.
  • Simulated 'like' feedback from the US Agent reranking the top-10 list improves HR@10 from 0.0994 to 0.1003 overall and from 0.0663 to 0.0678 for cold-start users, while simulated 'dislike' feedback lowers performance.
  • Every component of the pipeline contributes: removing recommendation-relevant analysis drops HR@10 to 0.0816 (below the no-caption baseline), removing CoT reasoning from the US Agent cuts accuracy from 0.715 to 0.580, and dropping comments or IP captions also degrades simulation.
  • The same US Agent, used alone on text-only MovieLens-1M, outperforms prior user-simulation agents, suggesting the RL training transfers beyond video.

Reading between the lines

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

  • A controlled comparison that applies the same SFT and RFT treatment to the same backbone would isolate how much of the gain is due to reinforcement learning versus simply updating the model weights; the current comparison only pits RL-trained Qwen2.5-7b against frozen GPT-4o and DeepSeek-R1.
  • The RRA ablation dropping below the no-caption baseline marks the 35-word caption as a potential bottleneck; testing longer captions, full-video inputs, or audio-plus-frame perception would show whether the gains come from semantic compression or from discarding information.
  • If the simulator generalizes, its 'like'/'dislike' labels could serve as pseudo-labels for cold-start users at larger scale, though the paper only tests this on 8,000 users and shows that dislike feedback must be filtered out.
  • The reported next-video-selection reward only checks the final pick, so the CoT reasoning is free to be gamed; rewarding intermediate reasoning quality or diversity of simulated actions could make the simulator more robust and more interpretable.
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

4 major / 5 minor

Summary. The paper proposes VRAgent-R1, a two-agent framework for video recommendation. The Item Perception (IP) Agent uses an MLLM to generate recommendation-oriented captions from key frames and titles via key-frame retrieval, collaborative multimodal perception, and recommendation-relevant analysis; these captions are used to augment a SASRec sequence recommender. The User Simulation (US) Agent is trained with GRPO on two tasks, preference judgment and next-video selection, using real user behavior as reward, and its simulated feedback is used to augment cold-start user modeling. Experiments on MicroLens-100K report that the IP Agent improves SASRec by 4.3% in HR@10 and 6.0% in NDCG@10, and that the US Agent reaches 64.1% next-video selection accuracy, about 45% above MLLM-MSR. Additional results on MovieLens-1M show gains in preference simulation.

Significance. If the empirical claims hold, the paper makes a useful practical contribution: it shows that MLLM-generated captions can improve a standard sequential recommender on a public video dataset, and it is among the first to apply reinforcement fine-tuning to LLM-based user simulation. The use of a public benchmark, ground-truth rewards from logged behavior, configurable negative sampling, and ablations of both agents are strengths. The claimed gains are large, especially in user simulation, but the central evaluation is compromised by an unspecified train/test user split, and the main recommendation table lacks error bars. The paper does not attempt analytic derivations, and the RL reward is computed from real behavior rather than from the model itself, so circularity is not a concern; the issue is empirical validity.

major comments (4)
  1. [Section 4, Implementation Details and Section 4.2] The manuscript states that reinforcement fine-tuning uses information from 2000 users and that user simulation evaluation is conducted on 1000 randomly selected cold-start users, but it never states that the evaluation users are disjoint from the RFT training users. Because the RFT tasks are constructed in exactly the same way as the evaluation tasks (the user's real last watched item as the positive target, SASRec top-10 recommendations, and m randomly chosen negatives), overlap between the training and evaluation users would let the model memorize the target items rather than learn to simulate user decisions. This would directly inflate the Acc_m=3 and Acc_m=4 numbers in Table 2 and the paper's headline claim of about 45% improvement over MLLM-MSR. Please specify how the 2000 RFT users and the 1000 evaluation users were sampled, confirm that the evaluation set is disjoint from the training set, and report results on a held-out split.
  2. [Table 1 and Appendix A.3] The main recommendation claim (4.3% HR@10 and 6.0% NDCG@10 improvement) is reported without error bars, while the appendix merely says that three runs produced 'almost the same results' without giving the numbers. Because several entries in Table 1 are close (e.g., SASRecF at 0.0953 vs. SASRec+Ours at 0.0994), the reader cannot judge whether the improvement is statistically meaningful. Please report the per-run values and mean ± standard deviation for the proposed method and the strongest baselines, and provide a significance test for the main comparison.
  3. [Section 4.2, Table 2] The user-simulation evaluation protocol is underspecified in ways that affect the headline results. The paper does not define 'cold-start' for the 1000 evaluation users, does not describe how the m negative candidates are sampled per evaluation task, and does not state whether all baselines are evaluated on the same users and the same candidate sets. These details are necessary to interpret Acc_m=3 and Acc_m=4, since the difficulty of the next-video selection task depends heavily on how negatives are drawn. Please provide the full protocol, including the exact definition of the candidate pools, and state whether the evaluation data are shared across methods.
  4. [Section 4.3, Table 4] The paper claims that simulated user feedback can further boost recommendation accuracy, but the procedure for updating the recommender with the US Agent's liked/disliked feedback is not described. Table 4 reports gains for '+US like' and a drop for '+US dislike', yet the reader cannot see how the simulated interactions were incorporated into the sequence model or how many episodes were generated. Please add the experimental setup for this optimization step, or explicitly label the result as a preliminary demonstration rather than a supported claim.
minor comments (5)
  1. [Figure 1] Figure 1 contains garbled text, including 'use's historic…' and an unaligned 'SFT Ours' label, which makes the qualitative example difficult to read; please clean the figure and its annotations.
  2. [Table 2] Table 2 refers to 'ALLREC [16]' while the cited work is TALLREC; please correct the method name to match the reference.
  3. [Table 5 and Table 6] The entries in Tables 5 and 6 run together (e.g., 'w/o KFR0.0980'), making the tables hard to parse; please format the rows and columns clearly and define the abbreviations in the captions.
  4. [Section 3.1, RRA] The paper says the summary is limited to approximately 35 words, but it does not describe how this length constraint is enforced or verified; please state whether the prompt alone, decoding constraints, or post-processing are used.
  5. [NeurIPS Checklist, Item 5] The checklist indicates that code will be released only after commercial approval, which limits reproducibility of the empirical claims; please consider releasing the prompts, generated captions, evaluation scripts, and the exact user split in the meantime, as these do not require releasing the trained model.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found; results are empirical comparisons against ground-truth behavior, with no fitted constant or self-citation chain forcing the conclusion.

full rationale

The paper's derivation chain is empirical rather than analytic: the IP Agent produces MLLM captions from video frames and titles, and these captions are fed as features into SASRec-style recommenders, whose HR/NDCG are computed against held-out interactions (Table 1). The US Agent is trained with GRPO against reward functions (Eqs. 1-2, 4-6) that are computed directly from the user's actual next item and randomly chosen negatives, and it is evaluated on a separately described set of 1000 cold-start users (Section 4, Table 2). No parameter is fitted to the evaluation metric and then reported as a prediction; no uniqueness theorem or load-bearing result is imported from the authors' prior work (references [28] and [29] are only related-work citations); and no known empirical pattern is merely renamed. The only concern raised by the manuscript text is a possible data-leakage issue: Section 4 states RFT uses 'information from 2000 users' and evaluation is on '1000 randomly selected cold-start users' without explicitly stating that the evaluation users are disjoint from the training users, and the appendix significance paragraph does not address the split. That is a validity/reproducibility concern, not a circularity concern, because even if overlap existed it would inflate results through memorization rather than by construction rendering the claimed prediction equivalent to an input. Under the requirement to exhibit a specific reduction, no circular step can be quoted, so the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

All entries are domain assumptions or hand-chosen hyperparameters. The paper contains no analytic derivation, so the free-parameter list consists of design choices the empirical claims depend on. None of these are fitted physical constants, and no new physical or conceptual entities are introduced.

free parameters (6)
  • key frame count = 3
    IP Agent selects top-3 CLIP frames; chosen by hand and directly affects caption quality and compute cost.
  • negative samples m = 3 or 4
    Next-video selection uses m random negatives; accuracy depends on m and is easier with smaller m.
  • reward scores = Rjud +1/-1, Rsel +2/-1.5/-2, Rformat 1/0.5/0/-1
    Hand-designed reward shaping in Equations 1-6; relative reward weights determine what policy learns.
  • RFT training set size = 2000 users
    RL fine-tuning uses only 2000 users; the 'little training data' claim depends on this number.
  • caption length target = 35 words
    RRA prompt asks for roughly 35-word captions; the ablation w/o RRA shows output format strongly affects downstream performance.
  • GRPO hyperparameters = G=16, KL coefficient 0.001, batch 16, 4 epochs
    Chosen training configuration; no sensitivity analysis is reported.
assumptions (4)
  • domain assumption MicroLens-100K interaction logs and user comments reflect genuine user preferences.
    All rewards and evaluation use the next watched video as ground truth; if these signals are noisy or popularity-driven, reported accuracy is not a clean measure of user simulation.
  • domain assumption Qwen2.5-7B MLLM can reliably describe video content from three CLIP-selected frames and a title.
    IP Agent captions are the only added video signal; MLLM errors or hallucinations transfer directly into item features and user simulation inputs.
  • domain assumption GRPO rewards based on next-item ground truth produce a policy that generalizes beyond training users.
    No formal generalization guarantee is given; the claim depends on train/eval user disjointness and on candidate-set distribution.
  • domain assumption SASRec on MicroLens-100K is a valid host model for evaluating caption augmentation.
    Improvements are measured through one backbone and one dataset; a different backbone could change the observed gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VRAgent-R1: Boosting Video Recommendation with MLLM-based Agents via Reinforcement Learning." pith.science (2026). https://pith.science/paper/Q7BELYXT

@misc{pith2026250702626,
  author       = {Pith},
  title        = {Pith review of: VRAgent-R1: Boosting Video Recommendation with MLLM-based Agents via Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q7BELYXT}},
  note         = {Machine review of arXiv:2507.02626}
}
read the original abstract

Owing to powerful natural language processing and generative capabilities, large language model (LLM) agents have emerged as a promising solution for enhancing recommendation systems via user simulation. However, in the realm of video recommendation, existing studies predominantly resort to prompt-based simulation using frozen LLMs and encounter the intricate challenge of multimodal content understanding. This frequently results in suboptimal item modeling and user preference learning, thereby ultimately constraining recommendation performance. To address these challenges, we introduce VRAgent-R1, a novel agent-based paradigm that incorporates human-like intelligence in user simulation. Specifically, VRAgent-R1 comprises two distinct agents: the Item Perception (IP) Agent and the User Simulation (US) Agent, designed for interactive user-item modeling. Firstly, the IP Agent emulates human-like progressive thinking based on MLLMs, effectively capturing hidden recommendation semantics in videos. With a more comprehensive multimodal content understanding provided by the IP Agent, the video recommendation system is equipped to provide higher-quality candidate items. Subsequently, the US Agent refines the recommended video sets based on in-depth chain-of-thought (CoT) reasoning and achieves better alignment with real user preferences through reinforcement learning. Experimental results on a large-scale video recommendation benchmark have demonstrated the effectiveness of our proposed VRAgent-R1 method, e.g., the IP Agent achieves a 6.0\% improvement in NDCG@10 on the MicroLens-100k dataset, while the US Agent shows approximately 45.0\% higher accuracy in user decision simulation compared to state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2507.02626 by the authors.

Figure 1
Figure 1. Qualitative examples of VRAgent-R1 for Video Recommendation. Our VRAgent-R1 stands out from previous supervised fine-tuning of MLLM, which fails to give the correct prediction due to the lack of understanding of video items and deep thinking on user status. Benefiting from the significant advancements in Multimodal Large Language Models (MLLMs), recent studies tend to utilize their strengths and potential to boost m… view at source ↗
Figure 2
Figure 2. Overview of our VRAgent-R1 framework. We propose a framework with two novel agents for better video recommendation. The IP Agent conducts collaborative multimodal understand￾ing to obtain enhanced video features for the recommendation system and the US Agent. Meanwhile, the US Agent simulates user behavior via deep CoT reasoning based on user status. By reinforcement learning with actual behavior rewards, VRAgent-R1… view at source ↗
Figure 3
Figure 3. Video understanding by the IP Agent. We simulate the human video comprehension process through a progressive approach involving retrieval, collaborative perception, and analysis, so as to obtain a summary of the key video information that is applicable for recommendation. Key Frame Retrieval (KFR). Given that videos contain a wealth of visual information, directly utilizing all video frames would introduce substanti… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: User Group Performance. Our method has a great advantage in modeling cold-start users due to the good multimodal understanding, outperforming the baseline by more than 10%. and 719,405 interactions, with a sparsity level of 99.96%. Notably, MicroLens is the first micro…
Figure 4
Figure 4. Figure 4: It is evident that the ID-based method suffers the most significant performance degradation for [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison with other user simulation for recommendation. A.3 Additional Experiment Discussion Experiments on MovieLens. To verify the generality of our method in other domains, we also conduct user preference simulation tests on the widely used MovieLens-1M [48] datas…
Figure 6
Figure 6. Figure 6: Additional Visualization Results for Video Understanding. NeurIPS Paper Checklist 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The main claims made in t…

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. RecoReward: Recommender-Guided Multimodal Description Generation for Recommendation

    cs.IR 2026-07 conditional novelty 6.0 of 10

    Training an MLLM with a recommender-affinity reward that contrasts historical engaged users with non-target users improves downstream recall for content-only item descriptions.

  2. Twins: Learn to Predict Unified Representations with Focal Loss

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Channel-wise concatenation of SigLIP2 and Flux VAE features into one token, trained with a focal-style flow-matching loss, yields a unified representation with 1.59 gFID on ImageNet 256 and VAE-level reconstruction.

  3. Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems

    cs.IR 2026-07 accept novelty 5.0 of 10

    Agentic recommender systems are organized by agent role (assisted, as-recommender, as-simulator) crossed with autonomy levels L2–L5, yielding a roadmap of architectures, evaluation limits, and open challenges.

Reference graph

Works this paper leans on

73 extracted references · 42 canonical work pages · cited by 3 Pith papers

  1. [1]

    Tencentrec: Real-time stream recommendation in practice,

    Y . Huang, B. Cui, W. Zhang, J. Jiang, and Y . Xu, “Tencentrec: Real-time stream recommendation in practice,” in Proceedings of the 2015 ACM SIGMOD international conference on management of data, 2015, pp. 227–238

    work page 2015

  2. [2]

    Spectral collaborative filtering,

    L. Zheng, C.-T. Lu, F. Jiang, J. Zhang, and P. S. Yu, “Spectral collaborative filtering,” inProceedings of the 12th ACM conference on recommender systems, 2018, pp. 311–319

    work page 2018

  3. [3]

    Graph convolutional neural networks for web-scale recommender systems,

    R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec, “Graph convolutional neural networks for web-scale recommender systems,” in Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 974–983

    2018

  4. [4]

    Mixed negative sampling for learning two-tower neural networks in recommendations,

    J. Yang, X. Yi, D. Zhiyuan Cheng, L. Hong, Y . Li, S. Xiaoming Wang, T. Xu, and E. H. Chi, “Mixed negative sampling for learning two-tower neural networks in recommendations,” inCompanion proceedings of the web conference 2020, 2020, pp. 441–447

    work page 2020

  5. [5]

    Parameter-efficient transfer from sequential behaviors for user modeling and recommendation,

    F. Yuan, X. He, A. Karatzoglou, and L. Zhang, “Parameter-efficient transfer from sequential behaviors for user modeling and recommendation,” in Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, 2020, pp. 1469–1478

    work page 2020

  6. [6]

    Notellm-2: Multimodal large representation models for recommendation,

    C. Zhang, H. Zhang, S. Wu, D. Wu, T. Xu, X. Zhao, Y . Gao, Y . Hu, and E. Chen, “Notellm-2: Multimodal large representation models for recommendation,” arXiv preprint arXiv:2405.16789, 2024

    arXiv 2024

  7. [7]

    Hllm: Enhancing sequential recommendations via hierarchical large language models for item and user modeling,

    J. Chen, L. Chi, B. Peng, and Z. Yuan, “Hllm: Enhancing sequential recommendations via hierarchical large language models for item and user modeling,” arXiv preprint arXiv:2409.12740, 2024

    arXiv 2024

  8. [8]

    Qarm: Quantitative alignment multi-modal recommendation at kuaishou,

    X. Luo, J. Cao, T. Sun, J. Yu, R. Huang, W. Yuan, H. Lin, Y . Zheng, S. Wang, Q. Hu et al., “Qarm: Quantitative alignment multi-modal recommendation at kuaishou,” arXiv preprint arXiv:2411.11739 , 2024

    arXiv 2024

Show all 73 references
  1. [9]

    Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer,

    F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang, “Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer,” in Proceedings of the 28th ACM international conference on information and knowledge management, 2019, pp. 1441–1450

    2019

  2. [10]

    Representation learning with large language models for recommendation,

    X. Ren, W. Wei, L. Xia, L. Su, S. Cheng, J. Wang, D. Yin, and C. Huang, “Representation learning with large language models for recommendation,” in Proceedings of the ACM Web Conference 2024, 2024, pp. 3464–3475

    2024

  3. [11]

    Star: A simple training-free approach for recommendations using large language models,

    D.-H. Lee, A. Kraft, L. Jin, N. Mehta, T. Xu, L. Hong, E. H. Chi, and X. Yi, “Star: A simple training-free approach for recommendations using large language models,” arXiv preprint arXiv:2410.16458, 2024

    2024 arXiv

  4. [12]

    Learn: Knowledge adaptation from large language model to recommendation for practical industrial application,

    J. Jia, Y . Wang, Y . Li, H. Chen, X. Bai, Z. Liu, J. Liang, Q. Chen, H. Li, P. Jiang et al. , “Learn: Knowledge adaptation from large language model to recommendation for practical industrial application,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol....

    2025

  5. [13]

    Precise: Pre-training sequential recommenders with collaborative and semantic information,

    C. Song, C. Shen, H. Gu, Y . Wu, L. Yi, J. Wen, and C. Chen, “Precise: Pre-training sequential recommenders with collaborative and semantic information,” arXiv preprint arXiv:2412.06308, 2024

    2024 arXiv

  6. [14]

    Llm-powered user simulator for recommender system,

    Z. Zhang, S. Liu, Z. Liu, R. Zhong, Q. Cai, X. Zhao, C. Zhang, Q. Liu, and P. Jiang, “Llm-powered user simulator for recommender system,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 12, 2025, pp. 13 339–13 347

    2025

  7. [15]

    Harnessing multimodal large language models for multimodal sequential recommendation,

    Y . Ye, Z. Zheng, Y . Shen, T. Wang, H. Zhang, P. Zhu, R. Yu, K. Zhang, and H. Xiong, “Harnessing multimodal large language models for multimodal sequential recommendation,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 12, 2025, pp. 13 069–13 077

    2025

  8. [16]

    Tallrec: An effective and efficient tuning frame- work to align large language model with recommendation,

    K. Bao, J. Zhang, Y . Zhang, W. Wang, F. Feng, and X. He, “Tallrec: An effective and efficient tuning frame- work to align large language model with recommendation,” in Proceedings of the 17th ACM Conference on Recommender Systems, 2023, pp. 1007–1014

    2023

  9. [17]

    Usimagent: Large language models for simulating search users,

    E. Zhang, X. Wang, P. Gong, Y . Lin, and J. Mao, “Usimagent: Large language models for simulating search users,” in Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2024, pp. 2687–2692. 10

    2024

  10. [18]

    On generative agents in recommendation,

    A. Zhang, Y . Chen, L. Sheng, X. Wang, and T.-S. Chua, “On generative agents in recommendation,” in Proceedings of the 47th international ACM SIGIR conference on research and development in Information Retrieval, 2024, pp. 1807–1817

    2024

  11. [19]

    Simuser: Generating usability feedback by simulating various users interacting with mobile applications,

    W. Xiang, H. Zhu, S. Lou, X. Chen, Z. Pan, Y . Jin, S. Chen, and L. Sun, “Simuser: Generating usability feedback by simulating various users interacting with mobile applications,” in Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, 2024, pp. 1–17

    2024

  12. [20]

    Vista: A visually, socially, and temporally-aware model for artistic recommendation,

    R. He, C. Fang, Z. Wang, and J. McAuley, “Vista: A visually, socially, and temporally-aware model for artistic recommendation,” in Proceedings of the 10th ACM conference on recommender systems, 2016, pp. 309–316

    2016

  13. [21]

    A content-driven micro-video recommendation dataset at scale,

    Y . Ni, Y . Cheng, X. Liu, J. Fu, Y . Li, X. He, Y . Zhang, and F. Yuan, “A content-driven micro-video recommendation dataset at scale,” arXiv preprint arXiv:2309.15379, 2023

    2023 arXiv

  14. [22]

    Enhancing adversarial robustness of multi-modal recommendation via modality balancing,

    Y . Shang, C. Gao, J. Chen, D. Jin, H. Ma, and Y . Li, “Enhancing adversarial robustness of multi-modal recommendation via modality balancing,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 6274–6282

    2023

  15. [23]

    Multimodal recommender systems: A survey,

    Q. Liu, J. Hu, Y . Xiao, X. Zhao, J. Gao, W. Wang, Q. Li, and J. Tang, “Multimodal recommender systems: A survey,”ACM Computing Surveys, vol. 57, no. 2, pp. 1–17, 2024

    2024

  16. [24]

    A comprehensive survey on multimodal recommender systems: Taxonomy, evaluation, and future directions,

    H. Zhou, X. Zhou, Z. Zeng, L. Zhang, and Z. Shen, “A comprehensive survey on multimodal recommender systems: Taxonomy, evaluation, and future directions,”arXiv preprint arXiv:2302.04473, 2023

    2023 arXiv

  17. [25]

    User behavior simulation with large language model-based agents,

    L. Wang, J. Zhang, H. Yang, Z.-Y . Chen, J. Tang, Z. Zhang, X. Chen, Y . Lin, H. Sun, R. Songet al., “User behavior simulation with large language model-based agents,” ACM Transactions on Information Systems, vol. 43, no. 2, pp. 1–37, 2025

    2025

  18. [26]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,” arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  19. [27]

    Gpt-4o system card,

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford et al., “Gpt-4o system card,” arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  20. [28]

    Super encoding network: Recursive association of multi-modal encoders for video understanding,

    B. Chen, S. Chen, K. Li, Q. Xu, Y . Qiao, and Y . Wang, “Super encoding network: Recursive association of multi-modal encoders for video understanding,” arXiv preprint arXiv:2506.07576, 2025

    2025

  21. [29]

    Percept, chat, and then adapt: Multimodal knowledge transfer of foundation models for open-world video recognition,

    ——, “Percept, chat, and then adapt: Multimodal knowledge transfer of foundation models for open-world video recognition,” arXiv preprint arXiv:2402.18951, 2024

    2024

  22. [30]

    Large language models are zero-shot rankers for recommender systems,

    Y . Hou, J. Zhang, Z. Lin, H. Lu, R. Xie, J. McAuley, and W. X. Zhao, “Large language models are zero-shot rankers for recommender systems,” in European Conference on Information Retrieval. Springer, 2024, pp. 364–381

    2024

  23. [31]

    Rethinking the evaluation for conversational recommendation in the era of large language models,

    X. Wang, X. Tang, W. X. Zhao, J. Wang, and J.-R. Wen, “Rethinking the evaluation for conversational recommendation in the era of large language models,” arXiv preprint arXiv:2305.13112, 2023

    2023 arXiv

  24. [32]

    Suber: An rl environment with simulated human behavior for recommender systems,

    N. Corecco, G. Piatti, L. A. Lanzendörfer, F. X. Fan, and R. Wattenhofer, “Suber: An rl environment with simulated human behavior for recommender systems,” arXiv preprint arXiv:2406.01631, 2024

    2024 arXiv

  25. [33]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat et al., “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  26. [34]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,” arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  27. [35]

    Qwen2. 5-math tech- nical report: Toward mathematical expert model via self-improvement,

    A. Yang, B. Zhang, B. Hui, B. Gao, B. Yu, C. Li, D. Liu, J. Tu, J. Zhou, J. Linet al., “Qwen2. 5-math tech- nical report: Toward mathematical expert model via self-improvement,”arXiv preprint arXiv:2409.12122, 2024

    2024 arXiv

  28. [36]

    Internlm-math: Open math large language models toward verifiable reasoning,

    H. Ying, S. Zhang, L. Li, Z. Zhou, Y . Shao, Z. Fei, Y . Ma, J. Hong, K. Liu, Z. Wanget al., “Internlm-math: Open math large language models toward verifiable reasoning,”arXiv preprint arXiv:2402.06332, 2024

    2024 arXiv

  29. [37]

    o1-coder: an o1 replication for coding,

    Y . Zhang, S. Wu, Y . Yang, J. Shu, J. Xiao, C. Kong, and J. Sang, “o1-coder: an o1 replication for coding,” arXiv preprint arXiv:2412.00154, 2024. 11

    2024 arXiv

  30. [38]

    Codedpo: Aligning code models with self generated and verified source code,

    K. Zhang, G. Li, Y . Dong, J. Xu, J. Zhang, J. Su, Y . Liu, and Z. Jin, “Codedpo: Aligning code models with self generated and verified source code,” arXiv preprint arXiv:2410.05605, 2024

    2024 arXiv

  31. [39]

    Visual-rft: Visual reinforcement fine-tuning,

    Z. Liu, Z. Sun, Y . Zang, X. Dong, Y . Cao, H. Duan, D. Lin, and J. Wang, “Visual-rft: Visual reinforcement fine-tuning,” arXiv preprint arXiv:2503.01785, 2025

    2025 arXiv

  32. [40]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  33. [41]

    Qwen2. 5 technical report,

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Weiet al., “Qwen2. 5 technical report,” arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  34. [42]

    Learning deep structured semantic models for web search using clickthrough data,

    P.-S. Huang, X. He, J. Gao, L. Deng, A. Acero, and L. Heck, “Learning deep structured semantic models for web search using clickthrough data,” in Proceedings of the 22nd ACM international conference on Information & Knowledge Management, 2013, pp. 2333–2338

    2013

  35. [43]

    Lightgcn: Simplifying and powering graph convolution network for recommendation,

    X. He, K. Deng, X. Wang, Y . Li, Y . Zhang, and M. Wang, “Lightgcn: Simplifying and powering graph convolution network for recommendation,” inProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, 2020, pp. 639–648

    2020

  36. [44]

    Deepfm: a factorization-machine based neural network for ctr prediction,

    H. Guo, R. Tang, Y . Ye, Z. Li, and X. He, “Deepfm: a factorization-machine based neural network for ctr prediction,” arXiv preprint arXiv:1703.04247, 2017

    2017 arXiv

  37. [45]

    A simple convolutional generative network for next item recommendation,

    F. Yuan, A. Karatzoglou, I. Arapakis, J. M. Jose, and X. He, “A simple convolutional generative network for next item recommendation,” in Proceedings of the twelfth ACM international conference on web search and data mining, 2019, pp. 582–590

    2019

  38. [46]

    Session-based recommendations with recurrent neural networks,

    B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk, “Session-based recommendations with recurrent neural networks,” arXiv preprint arXiv:1511.06939, 2015

    2015 arXiv

  39. [47]

    Self-attentive sequential recommendation,

    W.-C. Kang and J. McAuley, “Self-attentive sequential recommendation,” in 2018 IEEE international conference on data mining (ICDM). IEEE, 2018, pp. 197–206

    2018

  40. [48]

    The movielens datasets: History and context,

    F. M. Harper and J. A. Konstan, “The movielens datasets: History and context,” Acm transactions on interactive intelligent systems (tiis), vol. 5, no. 4, pp. 1–19, 2015

    2015

  41. [49]

    Virtual-taobao: Virtualizing real-world online retail environment for reinforcement learning,

    J.-C. Shi, Y . Yu, Q. Da, S.-Y . Chen, and A.-X. Zeng, “Virtual-taobao: Virtualizing real-world online retail environment for reinforcement learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 01, 2019, pp. 4902–4909

    2019

  42. [50]

    Generative adversarial user model for reinforcement learning based recommendation system,

    X. Chen, S. Li, H. Li, S. Jiang, Y . Qi, and L. Song, “Generative adversarial user model for reinforcement learning based recommendation system,” in International conference on machine learning. PMLR, 2019, pp. 1052–1061

    2019

  43. [51]

    Kuaisim: A comprehensive simulator for recommender systems,

    K. Zhao, S. Liu, Q. Cai, X. Zhao, Z. Liu, D. Zheng, P. Jiang, and K. Gai, “Kuaisim: A comprehensive simulator for recommender systems,” Advances in Neural Information Processing Systems, vol. 36, pp. 44 880–44 897, 2023

    2023

  44. [52]

    Factual and personalized recommendations using language models and reinforcement learning,

    J. Jeong, Y . Chow, G. Tennenholtz, C.-W. Hsu, A. Tulepbergenov, M. Ghavamzadeh, and C. Boutilier, “Factual and personalized recommendations using language models and reinforcement learning,” arXiv preprint arXiv:2310.06176, 2023

    arXiv

  45. [53]

    Rlrf4rec: Reinforcement learning from recsys feedback for enhanced recommendation reranking,

    C. Sun, Y . Liang, Y . Yang, S. Xu, T. Yang, and Y . Tong, “Rlrf4rec: Reinforcement learning from recsys feedback for enhanced recommendation reranking,” arXiv preprint arXiv:2410.05939, 2024

    2024 arXiv

  46. [54]

    Rec-r1: Bridging generative large language models and user-centric recommendation systems via reinforcement learning,

    J. Lin, T. Wang, and K. Qian, “Rec-r1: Bridging generative large language models and user-centric recommendation systems via reinforcement learning,” arXiv preprint arXiv:2503.24289, 2025

    2025

  47. [55]

    R. S. Sutton, A. G. Barto et al., Reinforcement learning: An introduction. MIT press Cambridge, 1998, vol. 1, no. 1

    1998

  48. [56]

    Reinforcement learning: Theory and algorithms,

    A. Agarwal, N. Jiang, S. M. Kakade, and W. Sun, “Reinforcement learning: Theory and algorithms,” CS Dept., UW Seattle, Seattle, WA, USA, Tech. Rep, vol. 32, p. 96, 2019

    2019

  49. [57]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017. 12 A Technical Appendices and Supplementary Material Technical appendices with additional results, figures, graphs and proofs may b...

    2017 arXiv

  50. [58]

    Guidelines: • The answer NA means that the abstract and introduction do not include the claims made in the paper

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope of MLLM-b...

  51. [59]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We discuss the limitations at the end of the main paper. Guidelines: • The answer NA means that the paper has no limitation while the answer No means...

  52. [60]

    Guidelines: • The answer NA means that the paper does not include theoretical results

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [NA] Justification: We don’t consider theoretical proof or result in the paper. Guidelines: • The answer NA means...

  53. [61]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data...

  54. [62]

    Guidelines: • The answer NA means that paper does not include experiments requiring code

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: The data is open source to acc...

  55. [63]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: We specify the details in the e...

  56. [64]

    We report the bar of our main results in the appendix

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate informa- tion about the statistical significance of the experiments? Answer: [Yes] Justification: The main table do not provide error bars since s...

  57. [65]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We use 4 80G A100 GPUs fo...

  58. [66]

    Guidelines: • The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: We follow the NeurIPS code of Ethics. Guidelines: • The answer NA means that th...

  59. [67]

    But we establish the user profile which may have privacy consideration problems

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: Our work aims to improve video recommendation for users, enabling them to get more satisfying recommendat...

  60. [68]

    Guidelines: • The answer NA means that the paper poses no such risks

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: The paper poses...

  61. [69]

    Guidelines: • The answer NA means that the paper does not use existing assets

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: We cite all rele...

  62. [70]

    We prepare the documentation of our code for future reproduction and will release it afterward

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] 20 Justification: The new assets are well documented. We prepare the documentation of our code for future reproduction and will rel...

  63. [71]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  64. [72]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  65. [73]

    Answer: [Yes] Justification: We use MLLM to help understand video contents and and train the LLM with reinforce- ment fine-tuning to simulate user decision

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

    2025

Pith tools

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