- Rating
- 2
- Confidence
- 2(−1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper formulates behavior-driven personalized content generation as preference distribution completion. Given a user's positive-interaction history and cross-user behavioral data, the proposed framework learns a personalized distribution in a frozen CLIP space using collaborative retrieval, dual-view Riemannian flow matching, and source-distribution calibration. Samples from this distribution are decoded into short-video captions by a frozen language model. Experiments on RecIF and MicroLens-100K use embedding-based distribution metrics and LLM-based judgments.
Critical strengths
-
The paper makes a useful distinction between generating individually relevant outputs and matching the proportions of a user’s multiple interests.
-
The proposed framework is technically coherent, and the authors evaluate it using both set-level semantic metrics and multi-model LLM judgments.
Critical weaknesses
-
The central research target is not well grounded in real-world preference behavior. Positive short-video interactions are shaped by recommendation exposure, ranking, item availability, and changing user interests; they are not IID samples from an intrinsic preference distribution. Since the paper explicitly omits exposure and has no explicit preference labels, controlled exposure data, or user study, predicting held-out interactions cannot establish that the method has recovered a user's true or previously unobserved preferences. It may only reproduce the platform’s logged interaction distribution.
-
The claimed “preference completion” is not identifiable from the evaluation protocol. Interests are considered unobserved because some existing interactions are withheld from the visible history. If the split is random, this creates an artificial missing-interest problem rather than testing discovery of genuinely new preferences. If it is temporal, the IID formulation is invalid because preferences and exposure may change over time. The theoretical result relies on smooth shared preferences and accurately identified neighbors, but the experiments do not verify these assumptions or measure whether collaborative evidence introduces correct versus spurious interests.
-
The experimental scope is substantially narrower than the paper’s claims. The method generates short-video captions rather than videos or other consumable content. It does not evaluate whether the generated concepts are visually feasible, novel, useful, or preferred by users. Consequently, the experiments support personalized caption-distribution modeling at most, not personalized content generation more broadly.
-
The evaluation and baselines do not convincingly isolate the proposed contribution. The model is trained to match caption embeddings in a frozen CLIP space, and most quantitative metrics are computed in that same space using objectives closely related to the calibration loss. Improvements may therefore reflect optimization of the selected representation rather than genuine preference completion. In addition, the paper lacks matched distributional baselines—such as collaborative empirical sampling, semantic mixture models, or conditional generative models—using the same retrieval evidence, decoder, sample count, and computational budget. The current ablation only adds components within the proposed pipeline and does not establish that Riemannian flow matching and source calibration are necessary.
The paper proposes a technically elaborate method, but its central object: the user’s latent preference distribution, is not identified by the available data. Logged positive interactions are treated as direct preference samples despite being mediated by exposure, ranking, and temporal dynamics. Holding out some interactions and predicting their caption embeddings therefore does not demonstrate completion of genuine unobserved preferences. The evidence is further limited to short-video caption generation and evaluation largely within the same semantic space optimized during training. Without exposure-aware or user-validated targets, independent evaluation, and matched distributional baselines, the experiments do not substantiate the paper’s main claims. Addressing these issues would require a revised evaluation design and substantial new experiments rather than clarification.
Questions for authors
-
How are visible histories and held-out targets divided? If the split is random, how does the task represent genuinely unobserved interests? If it is chronological, how are temporal preference and exposure shifts handled?
-
What evidence justifies treating positive interactions as samples from a user’s preference distribution rather than from a recommender-mediated exposure distribution?
-
Do the gains remain under an independent embedding model or human preference evaluation, and against simpler distributional baselines with identical collaborative evidence, decoder, generation budget, and supervision?
8 comments
Replies to question 2 and weakness 1
Q2 & W1. What evidence justifies treating positive interactions as samples from a user’s preference distribution rather than from a recommender-mediated exposure distribution?
The central research target is not well grounded in real-world preference behavior. Positive short-video interactions are shaped by recommendation exposure, ranking, item availability, and changing user interests; they are not IID samples from an intrinsic preference distribution. Since the paper explicitly omits exposure and has no explicit preference labels, controlled exposure data, or user study, predicting held-out interactions cannot establish that the method has recovered a user's true or previously unobserved preferences. It may only reproduce the platform’s logged interaction distribution.
Thank you for highlighting the distinction between a user’s preference distribution and the interaction distribution shaped by the platform’s exposure mechanism. We agree that exposure, ranking, and content availability affect which interactions are observed. Positive interactions also record users’ responses to the content they encounter, providing evidence about their preferences. Our goal is to complete users’ preference distributions from this behavioral evidence. Section 3.1 states the simplifying assumption explicitly: we omit the unobserved exposure process and model each user’s positive interactions as i.i.d. samples from their preference distribution. Recent work also uses interaction histories to infer preferences. PreferDiff [1] models user preference distributions from historical behavior, while Pigeon [2] assumes that images from users’ interaction histories still reflect their visual preferences despite noise and uses these images to construct personalized generation targets.
For our task, detailed user-specific preference data that separate individual interests from exposure effects are not readily available because of privacy and access constraints [4], as well as the practical difficulty of collecting controlled observations from individual users. PMG [3] also relies on interaction-derived supervision: it uses the last item in a user’s interaction sequence as the target, noting that a personalized target image is not directly available as ground truth. Given these constraints, we evaluate our method on public interaction datasets in a simplified setting that uses positive interaction histories and their associated content without explicitly modeling exposure. The results show improvements in semantic alignment, coverage, and distributional agreement with held-out positive targets, supporting the effectiveness of the proposed distribution completion method in this setting.
References
[1] Shuo Liu, An Zhang, Guoqing Hu, Hong Qian, and Tat-Seng Chua. Preference Diffusion for Recommendation. ICLR, 2025.
[2] Yiyan Xu, Wenjie Wang, Yang Zhang, Biao Tang, Peng Yan, Fuli Feng, and Xiangnan He. Personalized Image Generation with Large Multimodal Models. WWW, 2025.
[3] Xiaoteng Shen, Rui Zhang, Xiaoyan Zhao, Jieming Zhu, and Xi Xiao. PMG: Personalized Multimodal Generation with Large Language Models. WWW, 2024.
[4] Zhongzhou Liu, Yuan Fang, and Min Wu. Estimating Propensity for Causality-based Recommendation without Exposure Data. NeurIPS, 2023.
Replies to question 3 and weakness 4
Q3 & W4. Do the gains remain under an independent embedding model or human preference evaluation, and against simpler distributional baselines with identical collaborative evidence, decoder, generation budget, and supervision?
The evaluation and baselines do not convincingly isolate the proposed contribution. The model is trained to match caption embeddings in a frozen CLIP space, and most quantitative metrics are computed in that same space using objectives closely related to the calibration loss. Improvements may therefore reflect optimization of the selected representation rather than genuine preference completion. In addition, the paper lacks matched distributional baselines—such as collaborative empirical sampling, semantic mixture models, or conditional generative models—using the same retrieval evidence, decoder, sample count, and computational budget. The current ablation only adds components within the proposed pipeline and does not establish that Riemannian flow matching and source calibration are necessary.
Thank you for this suggestion. We have strengthened the evaluation with independent embedding models and additional controlled comparisons. The gains persist under BGE-M3 and Qwen3-Embedding-8B, as reported in Appendix L, Tables 12–13. Our method achieves the best Target-best, Coverage, FMC, Energy, and -U scores among the compared methods on both datasets under both encoders. This supports the robustness of our results beyond the CLIP space used for training.
Our original baseline suite (Table 1) already includes strong conditional text-generation methods, particularly PPlug, which generates personalized text conditioned on learned user representations. We additionally include Conditional Spherical Mixture and Dual-view Euclidean FM, covering a simpler mixture model and an alternative conditional generative model. Both follow the original experimental configuration, using the same collaborative evidence, supervision data, decoder, and generation settings (Appendix G). Our method outperforms both added baselines on Target-best, Coverage, FMC, and -U across both datasets. These comparisons support the effectiveness of our preference-distribution modeling beyond the simpler distributional alternatives.
| Method | Target-best | Coverage | FMC | Energy | -U |
|---|---|---|---|---|---|
| RecIF | |||||
| PPlug | |||||
| Conditional spherical mixture | |||||
| Dual-view Euclidean FM | |||||
| Ours | |||||
| MicroLens-100K | |||||
| PPlug | |||||
| Conditional spherical mixture | |||||
| Dual-view Euclidean FM | |||||
| Ours |
To isolate the contribution of spherical transport, we compare Dual-view Euclidean FM with Dual-view RFM under the same experimental configuration. Both models use the same dual-view conditioning, with source calibration disabled in both. As shown below, RFM improves Target-best, Coverage, FMC, and -U on both datasets, supporting the empirical benefit of spherical transport under matched conditioning. The effectiveness of source calibration is separately demonstrated by our ablation in Section 5.3, Table 3, with statistical assessment in Appendix J, Table 10. Adding calibration improves Target-best, Coverage, and FMC on both datasets, with the FMC improvements supported by paired-bootstrap tests at .
| Method | Target-best | Coverage | FMC | Energy | -U |
|---|---|---|---|---|---|
| RecIF | |||||
| Dual-view Euclidean FM | |||||
| Dual-view RFM | |||||
| MicroLens-100K | |||||
| Dual-view Euclidean FM | |||||
| Dual-view RFM |
Replies to questions 1 and weakness 2
Q1 & W2. How are visible histories and held-out targets divided? If the split is random, how does the task represent genuinely unobserved interests? If it is chronological, how are temporal preference and exposure shifts handled?
Thank you for your comments.
Section 5.1 introduces observed history and held-out targets. We have added “Within-user interaction split” to Appendix G to specify how they are constructed. We order each evaluation user’s valid positive interactions using a hash with a fixed seed and determine the holdout size from a 20% base rate, constrained to 16–32 interactions for RecIF and 2–4 for MicroLens. The remaining interactions form the observed history, and all methods use the same partition.
Our task is to complete a user's preference distribution from a limited history. A user may already have an interest that does not appear in the history given to the model. Random withholding lets us test whether the model can recover such interests, using the same user's positive interactions outside the visible history as the reference. We do not treat every held-out item as an unobserved interest. In Appendix H, the evaluator first identifies the interests reflected in the held-out targets without seeing the generated captions, then checks whether those interests appear in the visible history. IVC measures coverage of targets whose interests are weakly represented or absent from that history. Only valid generated captions that match these targets count toward coverage. Table 2 reports IVC gains over PPlug of 6.0 percentage points on RecIF and 17.0 percentage points on MicroLens.
Our existing experiments examine the role of collaborative information and the completion objective developed in Section 3.2. First, Proposition 1 analyzes how observations from related users can support preference estimation, with the error bound accounting for cross-user transfer bias and the effective reference sample size. The bound includes the neighborhood width and the state-estimation errors of both the target user and the reference users, allowing for imperfect neighbor identification. Section 4.2 puts this idea into practice by finding users with related interaction patterns and using their additional interactions as conditioning information. Table 3 evaluates the contribution of this information: before source calibration, adding the collaborative view improves FMC by 2.7 percentage points on RecIF and 4.0 percentage points on MicroLens, together with improvements in target alignment. This provides empirical support for the usefulness of the collaborative information selected by our method. Second, Section 3.2 explains how finite interaction samples can supervise and evaluate the learned preference distribution under the stated sampling assumption. For the full model, IVC in Table 2 measures valid coverage of interests that are weakly represented or absent in the visible history but supported by held-out interactions. ES and MMD in Table 1 assess agreement between the generated set and the held-out content distribution in the fixed semantic space. The IVC gains show better coverage of under-observed interests, while the lower ES and MMD values show that this broader coverage is accompanied by better distributional agreement.
To further examine the relevance of the selected collaborative evidence, we conducted an additional comparison with random reference evidence and added the experiment to Appendix I. Both conditions use eight evidence items and the same held-out targets. For each target, we compute its highest cosine similarity to the eight evidence items and average these scores over the targets. We report results for all held-out targets and for weak-history targets. The latter are targets whose interests are weakly represented or absent in the user's visible history, as identified by the history-support annotations in Appendix H. On both datasets, the collaborative evidence is more closely aligned with the user's held-out interactions than the same number of random reference items. This also holds for targets whose interests are weakly represented or absent in the visible history.
\hline & \text{All targets}\uparrow & \text{Weak-history targets}\uparrow \\\\ \hline \begin{array}{l} \text{Dataset} \\\\ \hline \text{RecIF} \\\\ \text{MicroLens-100K} \end{array} & \begin{array}{cc} \text{Random evidence} & \text{Collaborative evidence} \\\\ \hline 0.627 & \mathbf{0.646} \\\\ 0.614 & \mathbf{0.637} \end{array} & \begin{array}{cc} \text{Random evidence} & \text{Collaborative evidence} \\\\ \hline 0.607 & \mathbf{0.625} \\\\ 0.614 & \mathbf{0.625} \end{array} \\\\ \hline \end{array}$$