acceptodds
Under review as a conference paper at ICLR 2027

Personalized Content Generation as Preference Distribution Completion

Abstract

Personalized content generation is becoming increasingly important in modern generative systems. Existing approaches typically personalize provided prompts, instructions, or references; directly generating content from user behavior requires inferring preferences from limited observations. This makes it difficult to determine both what to generate and how generation probabilities should be distributed across interests. In this paper, we formulate the personalized content generation task as preference distribution completion. To this end, we develop a collaborative framework that draws on cross-user behavioral patterns to supplement sparse individual preferences. The framework combines source-distribution modeling with history-conditioned Riemannian flow matching, which together determine how probability mass is allocated in a semantic latent space. A decoder maps samples from the resulting distributions directly to content, without explicit user prompts. Experiments on short-video user interaction benchmarks, with textual descriptions as the generation targets, demonstrate the effectiveness of our approach.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Personalized Content Generation, Preference Distribution Completion, Riemannian Flow Matching, Collaborative Filtering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vop1OpenReview ↗
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
  1. The paper makes a useful distinction between generating individually relevant outputs and matching the proportions of a user’s multiple interests.

  2. The proposed framework is technically coherent, and the authors evaluate it using both set-level semantic metrics and multi-model LLM judgments.

Critical weaknesses
  1. 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.

  2. 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.

  3. 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.

  4. 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
  1. 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?

  2. What evidence justifies treating positive interactions as samples from a user’s preference distribution rather than from a recommender-mediated exposure distribution?

  3. 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
AuthorsOpenReview ↗

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}$$
AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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 MMD2\mathrm{MMD}^{2}-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 MMD2\mathrm{MMD}^{2}-U across both datasets. These comparisons support the effectiveness of our preference-distribution modeling beyond the simpler distributional alternatives.

MethodTarget-best ↑\uparrowCoverage ↓\downarrowFMC ↑\uparrowEnergy ↓\downarrowMMD2\mathrm{MMD}^{2}-U ↓\downarrow
RecIF
PPlug0.7220.7220.7550.7550.6010.6010.4910.4910.0030.003
Conditional spherical mixture0.7130.7130.7690.7690.5210.5210.4990.4990.0020.002
Dual-view Euclidean FM0.7240.7240.7430.7430.6000.6000.4790.4790.0020.002
Ours0.7550.7550.7060.7060.6910.6910.4780.4780.0010.001
MicroLens-100K
PPlug0.7040.7040.7770.7770.5450.5450.5170.5170.0290.029
Conditional spherical mixture0.6980.6980.7870.7870.4740.4740.5100.5100.0180.018
Dual-view Euclidean FM0.7130.7130.7650.7650.5570.5570.5000.5000.0160.016
Ours0.7290.7290.7460.7460.6320.6320.5000.5000.0110.011

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 MMD2\mathrm{MMD}^{2}-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 p<0.01p<0.01.

MethodTarget-best ↑\uparrowCoverage ↓\downarrowFMC ↑\uparrowEnergy ↓\downarrowMMD2\mathrm{MMD}^{2}-U ↓\downarrow
RecIF
Dual-view Euclidean FM0.7240.7240.7430.7430.6000.6000.4790.4790.0020.002
Dual-view RFM0.7450.7450.7220.7220.6700.6700.4790.4790.0010.001
MicroLens-100K
Dual-view Euclidean FM0.7130.7130.7650.7650.5570.5570.5000.5000.0160.016
Dual-view RFM0.7250.7250.7480.7480.6200.6200.5050.5050.0120.012
Reviewer_F9tUOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper innovatively formulates personalized content generation as preference distribution completion problem. In practical scenarios, users' preference history is usually incomplete. Cross-user interactions can support the completion. Therefore, a novel framework via conditioned flow matching is proposed. During inference, semantic samples are drawn from personalized distribution and mapped by a frozen caption decoder. Experiments show the effectiveness of the proposed framework.

Critical strengths
  1. The task formulation is clear and well motivated. The preference distribution completion directly addresses sparse user histories and makes the best use of cross-user interests.
  2. The evaluation and ablation support the core components and the overall results are consistent with the stated contributions.
Critical weaknesses
  1. The direction of 'coverage' metrics may mislead readers. While the higher target-best and lower coverage are better, FMC is also a coverage-style metric but higher the better.
  2. The set-level LLM evaluation relies heavily on closed API: gpt 5.6 sol and gemini 3.8 flash. This may raise reproducibility concerns. More open-weight judges will make the LLM-based assessment more robust.
Questions for authors

Please see weaknesses above.

3 comments
AuthorsOpenReview ↗

Replies to weakness 1 and weakness 2

W1. The direction of 'coverage' metrics may mislead readers. While the higher target-best and lower coverage are better, FMC is also a coverage-style metric but higher the better.

Thank you for your careful reading and pointing out this ambiguity. We agree that the shared term “coverage” may obscure the different meanings and directions of these metrics. Coverage (Cov.) measures the mean angular distance from each held-out target to its nearest generated sample, so lower is better. Future-mode coverage (FMC) measures the fraction of held-out target modes supported by generated samples, so higher is better. We have added a footnote in the Evaluation subsection explicitly clarifying this distinction and the direction of each metric.


W2. The set-level LLM evaluation relies heavily on closed API: gpt 5.6 sol and gemini 3.8 flash. This may raise reproducibility concerns. More open-weight judges will make the LLM-based assessment more robust.

Thank you for this suggestion. We have added evaluations with two open-weight judges, Gemma 4 31B and Qwen 3.8 27B, in Appendix K. Across both datasets and all six evaluated baselines, Q1 preference scores (win rates) exceed 50%, and improvements in Q2 and Q3 are consistently positive. These results support the robustness of our findings across judge families and reduce reliance on closed API-based evaluation.

Each metric reports Gemma (left) and Qwen (right). Differences are computed as Ours minus the baseline; pp denotes percentage points.

\hline \textbf{Ours}\\ \textit{vs.}\\ \text{Baseline} & \begin{array}{c} \textbf{Q1: Overall}\\\\ \textbf{Preference}\\\\ \hline \begin{array}{c} \text{WR}\uparrow(\\%) \end{array} \end{array} & \begin{array}{c} \textbf{Q2: Target-level}\\\\ \textbf{Support}\\\\ \hline \begin{array}{cc} \quad\Delta\mathrm{BM}\uparrow\quad & \Delta\mathrm{VC}\uparrow(\mathrm{pp}) \end{array} \end{array} & \begin{array}{c} \textbf{Q3: Incremental}\\\\ \textbf{Completion}\\\\ \hline \begin{array}{c} \Delta\mathrm{IVC}\uparrow(\mathrm{pp}) \end{array} \end{array} \\\\ \hline \textit{Dataset: RecIF} & & & \\\\ \begin{array}{l} \text{Task-only}\\\\ \text{Full-history}\\\\ \text{LaMP-IPA}\\\\ \text{PPlug}\\\\ \text{CFRAG}\\\\ \text{ClusterRAG} \end{array} & \begin{array}{r|r} 100.0 & 99.6\\\\ 60.9 & 63.2\\\\ 75.8 & 70.7\\\\ 66.8 & 59.3\\\\ 67.4 & 68.4\\\\ 82.8 & 78.2 \end{array} & \begin{array}{cc} \begin{array}{r|r} +1.43 & +1.29\\\\ +0.52 & +0.35\\\\ +0.63 & +0.48\\\\ +0.23 & +0.18\\\\ +0.42 & +0.25\\\\ +0.74 & +0.59 \end{array} & \begin{array}{r|r} +55.1 & +60.9\\\\ +26.0 & +20.1\\\\ +30.5 & +21.6\\\\ +15.9 & +11.3\\\\ +21.6 & +10.4\\\\ +36.3 & +24.6 \end{array} \end{array} & \begin{array}{r|r} +19.6 & +22.5\\\\ +16.1 & +9.4\\\\ +23.2 & +14.5\\\\ +12.5 & +8.4\\\\ +8.3 & +12.5\\\\ +19.6 & +10.9 \end{array} \\\\ \hline \textit{Dataset: MicroLens-100K} & & & \\\\ \begin{array}{l} \text{Task-only}\\\\ \text{Full-history}\\\\ \text{LaMP-IPA}\\\\ \text{PPlug}\\\\ \text{CFRAG}\\\\ \text{ClusterRAG} \end{array} & \begin{array}{r|r} \phantom{0}95.3 & 94.5\\\\ 82.7 & 80.9\\\\ 77.0 & 82.4\\\\ 62.8 & 64.5\\\\ 82.0 & 86.7\\\\ 79.6 & 79.3 \end{array} & \begin{array}{cc} \begin{array}{r|r} +1.26 & +1.13\\\\ +1.00 & +0.75\\\\ +0.92 & +0.65\\\\ +0.36 & +0.19\\\\ +1.00 & +0.82\\\\ +1.05 & +0.84 \end{array} & \begin{array}{r|r} +43.2 & +49.6\\\\ +41.0 & +40.6\\\\ +36.8 & +34.5\\\\ +13.9 & +13.5\\\\ +39.8 & +43.8\\\\ +39.3 & +41.0 \end{array} \end{array} & \begin{array}{r|r} +40.7 & +44.0\\\\ +39.4 & +36.3\\\\ +36.4 & +28.1\\\\ +14.0 & +11.5\\\\ +38.6 & +39.0\\\\ +36.9 & +35.8 \end{array} \\\\ \hline \end{array}$$

Thanks for the clarification. The results on open-source VLMs are great. I'll keep my positive rating.

AuthorsOpenReview ↗

Thank you for your feedback and suggestions

Thank you for your time and for your constructive suggestions. We appreciate your feedback!

Reviewer_nzLgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates behavior-driven personalized content generation as preference distribution completion. The proposed framework combines observed user histories with collaborative evidence through dual-view Riemannian flow matching, followed by source-distribution calibration. Samples from the learned semantic distribution are converted into captions using a frozen decoder. Experiments on two short-video interaction datasets evaluate semantic alignment, interest coverage, and distributional agreement, complemented by LLM-based assessments and component ablations.

Critical strengths
  1. The paper considers both which interests generated content should cover and how frequently those interests should appear. This offers a meaningful perspective on personalized generation.
  2. Separating preference modeling from content decoding provides a clear way to incorporate individual histories and collaborative evidence into generation.
  3. The method improves over the compared baselines across both datasets, with complementary evidence from semantic metrics and LLM-based assessments.
  4. The decoder conditioning checks and order-swapped LLM comparisons help address alternative explanations for the reported results.
Critical weaknesses
  1. The paper emphasizes completing preferences from limited observations, but does not report performance across different history lengths. It therefore remains unclear how much collaborative evidence helps when user histories are particularly sparse.
  2. Appendix H describes bootstrap confidence intervals and significance tests, but Tables 1–3 do not show the corresponding results. Some gains from source calibration are small, making it difficult to assess whether they are reliable.
  3. The paper specifies the user-level data split, but does not clearly explain how each test user’s interactions are divided into history and targets. The baseline adaptations and decoding settings are also insufficiently described, making the comparisons harder to assess and reproduce.
Questions for authors

See Weaknesses

4 comments
AuthorsOpenReview ↗

Replies to weakness 1

W1. The paper emphasizes completing preferences from limited observations, but does not report performance across different history lengths. It therefore remains unclear how much collaborative evidence helps when user histories are particularly sparse.

Thank you for suggesting an evaluation across different user history lengths. We conducted additional experiments on a subset of RecIF with user history lengths of 8, 16, 32, 64, and 128 interactions. As shown below, our method consistently outperforms PPlug, the strongest baseline, across all evaluated history lengths in target matching, target-mode coverage, and distributional agreement. Notably, these advantages persist even when only 8 historical interactions are available, demonstrating that our framework remains effective under particularly sparse user histories.

Furthermore, our original results on MicroLens-100K, a dataset characterized by sparse user interactions, provide complementary evidence. As reported in Table 1, our method also outperforms all baselines on this dataset across these evaluation dimensions. Together, these results support the effectiveness of our approach when user preference observations are limited.

\begin{array}{ccc} \begin{array}{c} \textbf{Target-best}\\;\uparrow \\\\[5pt] \text{User history length} \\\\[2pt] \begin{array}{lccccc} \hline \text{Method} & 8 & 16 & 32 & 64 & 128 \\\\ \hline \textbf{Ours} & 0.739 & 0.743 & 0.748 & 0.752 & 0.754 \\\\ \text{PPlug} & 0.719 & 0.722 & 0.724 & 0.726 & 0.727 \\\\ \hline \Delta & +0.020 & +0.021 & +0.024 & +0.026 & +0.027 \\\\ \hline \end{array} \end{array} & \begin{array}{c} \textbf{Coverage}\\;\downarrow \\\\[5pt] \text{User history length} \\\\[2pt] \begin{array}{lccccc} \hline \text{Method} & 8 & 16 & 32 & 64 & 128 \\\\ \hline \textbf{Ours} & 0.730 & 0.723 & 0.719 & 0.711 & 0.708 \\\\ \text{PPlug} & 0.760 & 0.754 & 0.751 & 0.748 & 0.747 \\\\ \hline \Delta & +0.030 & +0.032 & +0.032 & +0.037 & +0.039 \\\\ \hline \end{array} \end{array} & \begin{array}{c} \textbf{Future-mode coverage}\\;\uparrow \\\\[5pt] \text{User history length} \\\\[2pt] \begin{array}{lccccc} \hline \text{Method} & 8 & 16 & 32 & 64 & 128 \\\\ \hline \textbf{Ours} & 0.658 & 0.653 & 0.679 & 0.692 & 0.700 \\\\ \text{PPlug} & 0.588 & 0.612 & 0.609 & 0.620 & 0.613 \\\\ \hline \Delta & +0.071 & +0.041 & +0.070 & +0.073 & +0.087 \\\\ \hline \end{array} \end{array} \\\\[16pt] \begin{array}{c} \textbf{Energy score}\\;\downarrow \\\\[5pt] \text{User history length} \\\\[2pt] \begin{array}{lccccc} \hline \text{Method} & 8 & 16 & 32 & 64 & 128 \\\\ \hline \textbf{Ours} & 0.484 & 0.481 & 0.480 & 0.479 & 0.478 \\\\ \text{PPlug} & 0.498 & 0.493 & 0.489 & 0.488 & 0.487 \\\\ \hline \Delta & +0.014 & +0.012 & +0.008 & +0.008 & +0.009 \\\\ \hline \end{array} \end{array} & \begin{array}{c} \mathbf{MMD}^{2}\textbf{-U}\\;\downarrow \\\\[5pt] \text{User history length} \\\\[2pt] \begin{array}{lccccc} \hline \text{Method} & 8 & 16 & 32 & 64 & 128 \\\\ \hline \textbf{Ours} & 0.0027 & 0.0024 & 0.0021 & 0.0020 & 0.0020 \\\\ \text{PPlug} & 0.0048 & 0.0041 & 0.0037 & 0.0037 & 0.0037 \\\\ \hline \Delta & +0.0021 & +0.0017 & +0.0016 & +0.0016 & +0.0018 \\\\ \hline \end{array} \end{array} & \begin{array}{c} \textbf{History-nearest} \\\\[5pt] \text{User history length} \\\\[2pt] \begin{array}{lccccc} \hline \text{Method} & 8 & 16 & 32 & 64 & 128 \\\\ \hline \textbf{Ours} & 0.663 & 0.693 & 0.719 & 0.743 & 0.760 \\\\ \text{PPlug} & 0.680 & 0.703 & 0.724 & 0.744 & 0.762 \\\\ \hline \Delta & -0.017 & -0.010 & -0.005 & -0.001 & -0.001 \\\\ \hline \end{array} \end{array} \end{array}$$ **Note.** For metrics with a preferred direction, $\Delta$ measures our advantage over PPlug: $\mathrm{Ours}-\mathrm{PPlug}$ for $\uparrow$ metrics and $\mathrm{PPlug}-\mathrm{Ours}$ for $\downarrow$ metrics. Larger $\Delta$ indicates a greater advantage. For History-nearest, $\Delta=\mathrm{Ours}-\mathrm{PPlug}$ is a diagnostic difference with no preferred direction.
AuthorsOpenReview ↗

Replies to weakness 2 and weakness 3

W2. Appendix H describes bootstrap confidence intervals and significance tests, but Tables 1–3 do not show the corresponding results. Some gains from source calibration are small, making it difficult to assess whether they are reliable.

Thank you for your suggestion. Due to space limitations, we did not include confidence intervals in the main-text tables. We have now added the complete results for Tables 1–3 with 95%95\% confidence intervals to Appendix J (Tables 7–9). We also separately report the gains from source calibration, their 95%95\% confidence intervals, and two-sided significance tests in Table 10, reproduced below. All gain intervals lie above zero, with p<0.01p<0.01 for nine comparisons and p<0.05p<0.05 for Cov. on MicroLens-100K. These results support the statistical reliability of the calibration gains, including those with small absolute magnitudes.

\begin{array}{lccccc} \hline \text{Statistic} & \text{TB}\\,\uparrow & \text{Cov.}\\,\downarrow & \text{FMC}\\,\uparrow & \text{ES}\\,\downarrow & \mathrm{MMD}^{2}\text{-U}\\,\downarrow \\\\[3pt] \hline \textit{Dataset: RecIF} & & & & & \\\\[5pt] \text{Gain }\Delta & \begin{array}{c} +0.010 \\\\[2pt] {\scriptsize [0.008863,\\,0.011117]} \end{array} & \begin{array}{c} +0.016 \\\\[2pt] {\scriptsize [0.014288,\\,0.017693]} \end{array} & \begin{array}{c} +0.021 \\\\[2pt] {\scriptsize [0.010257,\\,0.031828]} \end{array} & \begin{array}{c} +0.001 \\\\[2pt] {\scriptsize [0.000705,\\,0.001300]} \end{array} & \begin{array}{c} +0.00024 \\\\[2pt] {\scriptsize [0.000159,\\,0.000325]} \end{array} \\\\[8pt] \text{Significance} & p<0.01 & p<0.01 & p<0.01 & p<0.01 & p<0.01 \\\\[5pt] \hline \textit{Dataset: MicroLens-100K} & & & & & \\\\[5pt] \text{Gain }\Delta & \begin{array}{c} +0.004 \\\\[2pt] {\scriptsize [0.003055,\\,0.004945]} \end{array} & \begin{array}{c} +0.002 \\\\[2pt] {\scriptsize [0.000328,\\,0.003643]} \end{array} & \begin{array}{c} +0.012 \\\\[2pt] {\scriptsize [0.007635,\\,0.016329]} \end{array} & \begin{array}{c} +0.005 \\\\[2pt] {\scriptsize [0.004319,\\,0.005687]} \end{array} & \begin{array}{c} +0.00094 \\\\[2pt] {\scriptsize [0.000322,\\,0.001569]} \end{array} \\\\[8pt] \text{Significance} & p<0.01 & p<0.05 & p<0.01 & p<0.01 & p<0.01 \\\\[5pt] \hline \end{array}$$ **Note**. Each gain is accompanied by its $95\%$ confidence interval below. Positive $\Delta$ indicates improvement: calibrated minus uncalibrated for TB and FMC, and uncalibrated minus calibrated for Cov., ES, and $\mathrm{MMD}^{2}$-U. Significance refers to the reported two-sided tests. --- **W3. The paper specifies the user-level data split, but does not clearly explain how each test user’s interactions are divided into history and targets. The baseline adaptations and decoding settings are also insufficiently described, making the comparisons harder to assess and reproduce.** Thank you for your suggestions on the experimental description. 1. 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. 2. To clarify the “task-adapted” baselines listed in Section 5.1, we have added “Baseline adaptations” to Appendix G. We use descriptions of the user’s observed videos as preference history. Each baseline applies its personalization mechanism to this history to condition the generation of new video captions. For methods requiring explicit task inputs, we define the task as generating new video descriptions based on the user's history and derive the required personalization conditions from that history. The resulting caption collections are evaluated under the same set-level evaluation protocol. 3. Section 5.1 specifies 64 captions per user, generated using 32-step spherical Heun integration and greedy decoding without reranking. Section 4.4 describes the deterministic decoding process. Appendix G explains how the frozen decoder converts sampled semantic representations into video captions.

Thanks for the rebuttal. I will keep my positive score.

AuthorsOpenReview ↗

Thank you for your feedback and suggestions

Thank you for your time and for your constructive suggestions. We appreciate your feedback!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Revisions and General Response

Dear Reviewers and Area Chairs,

Thank you for taking the time to review our paper.

Our work formulates personalized content generation as preference-distribution completion, addressing both what to generate from limited user histories and how to allocate generation probabilities across interests. We provide a conditional theoretical analysis of how cross-user interactions can support completion and develop a collaborative framework combining dual-view Riemannian flow matching with source-distribution calibration. Experiments on RecIF and MicroLens-100K demonstrate improved target alignment, interest coverage, and distributional agreement in personalized video-description generation.

In our rebuttal and revision, we clarified the task assumptions and evaluation scope, expanded the experimental comparisons, and strengthened the statistical and reproducibility analyses. The additional results provide further evidence regarding performance under sparse histories, the relevance of collaborative evidence, and robustness across evaluation models.

Below, we summarize each reviewer’s main concerns and our corresponding responses:

ReviewerMain concerns (very brief)What we addressed in the rebuttal and revision
nzLgPerformance with sparse user histories; statistical reliability of small calibration gains; insufficient details on history–target splits, baseline adaptations, and decoding.Added history-length experiments on a RecIF subset with 8,16,32,64,8, 16, 32, 64, and 128128 interactions. Our method consistently outperforms PPlug in target matching, target-mode coverage, and distributional agreement, including with only eight interactions. Added 9595\\% confidence intervals for Tables 1–3 in Appendix J, together with separate calibration-gain intervals and two-sided significance tests: all ten gain intervals lie above zero, with p<0.01p<0.01 for nine comparisons and p<0.05p<0.05 for the remaining comparison. Specified the seeded within-user interaction split and baseline adaptations in Appendix G, and clarified the shared evaluation protocol and generation/decoding settings.
vop1Preference completion and exposure assumptions; collaborative evidence quality; evaluation robustness and controlled baselines; rationale for video-description generation, intended application, and personalization benefits.Clarified preference completion and the exposure-omitting assumption, explaining how held-out targets and IVC assess interests underrepresented in visible histories. Further clarified the rationale for video-description generation. Used existing LLM evaluations and the Figure 3 case study to illustrate concept-level personalization through better interest allocation and coverage supported by held-out behavior. Added collaborative-versus-random evidence comparisons, showing better target alignment, including for weak-history targets. Verified that gains persist under BGE-M3 and Qwen3-Embedding-8B. Added Conditional Spherical Mixture and Dual-view Euclidean FM with matched evidence, supervision, decoder, and generation settings; isolated the benefit of spherical transport with calibration disabled; and supported source calibration through ablations and statistical tests.
F9tUConfusing directions of coverage-related metrics; reproducibility and robustness of evaluation relying on closed API judges.Clarified the meanings and preferred directions of Coverage and FMC in an evaluation footnote. Added evaluations with two open-weight judges, Gemma 4 31B and Qwen 3.8 27B, in Appendix K. Across both datasets and all six evaluated baselines, both judges assign our method overall preference win rates above 5050\\%, with consistently positive gains in target-level support, valid coverage, and incremental valid coverage. These results strengthen robustness across judge families and reduce reliance on closed API-based assessment.

We understand that this year’s ICLR review process is especially demanding, and we are sincerely grateful that you took the time to read our detailed rebuttal and additional experiments. We remain confident that our work offers a meaningful contribution to the study of preference-distribution completion for personalized content generation!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.