acceptodds
Under review as a conference paper at ICLR 2027

GReCF: Generative Recommendation via Continuous-Space Collaborative Filtering

Abstract

Generative recommendation aims to create and recommend novel items beyond fixed catalogs, addressing users' evolving or unrepresented preferences that existing catalogs cannot adequately capture. It is important to generate content aligned with users' historical preferences while reliably extrapolating beyond observed history. We formulate this task as collaborative preference-field completion in continuous item space, with user-based and item-based collaborative filtering providing two completion mechanisms. We instantiate this formulation in GReCF, our proposed reference-free generative recommender in the image modality. GReCF represents users with adaptive multi-interest themes, anchors each generation to one theme, and injects user and theme conditions into a pretrained diffusion model through residual preference attention. Shared training transfers evidence across users, while masked-interest training learns relations among co-occurring themes to support extrapolation beyond observed history. To support generative recommendation research, we construct CIGR, a synthetic dataset with one million positive interactions. We also introduce an extrapolation-focused evaluation protocol specifically for the generative recommendation task. Across three datasets, GReCF achieves competitive ranking, strong personalization, favorable image quality, high coverage, and reliable CF extrapolation with few unsupported generations. We further investigate the cold-start generation capability of GReCF and show that it preserves initial interests while enabling CF extrapolation.

Primary area:
generative models
Keywords:
Generative Recommendation, 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_3uUFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper formulates catalog-free generative recommendation as collaborative preference-field completion in continuous item space. It proposes GReCF, which combines adaptive multi-interest representations, theme-anchored diffusion conditioning, and masked-interest training to generate personalized images and encourage extrapolation beyond observed preferences. The authors also introduce the synthetic CIGR dataset and an extrapolation-focused evaluation protocol, evaluating ranking performance, personalization, image quality, interest coverage, and collaborative support across three datasets.

Critical strengths
  1. Interesting problem setting. Directly generating new content from users’ preference histories is an interesting direction for recommendation. The image-generation setting explored here raises meaningful questions about both personalization and discovering interests beyond observed interactions.

  2. Natural and well-motivated approach. Collaborative filtering offers an intuitive starting point for this task: preferences shared across users can provide useful guidance for generating personalized content. Connecting this classical recommendation principle with content generation gives the approach a clear motivation.

  3. Comprehensive contributions. The paper presents a coherent study spanning problem formulation, methodology, dataset construction, and evaluation. The introduction of CIGR and an extrapolation-focused evaluation protocol, alongside experiments on three datasets, makes the contribution broader than a model proposal alone.

Critical weaknesses
  1. The CF formulation appears to serve largely as conceptual packaging for the implemented method. GReCF primarily trains a shared, history-conditioned diffusion model, interpreting shared training as user-based CF and masked-interest prediction as item-based CF. While these mechanisms could learn collaborative patterns, the paper does not sufficiently establish their correspondence to the similarity-weighted user aggregation and centered co-preference kernel introduced in the formulation. Shared training alone does not demonstrate collaborative preference transfer, and masked prediction may also exploit content similarity or global popularity. Moreover, matching generated images to neighboring users’ histories does not establish that collaborative learning caused those outputs. Consequently, the connection between the paper’s central CF framing and its technical contribution remains insufficiently substantiated.
Questions for authors

What concrete application scenarios or industries do you envision for GReCF? Has the approach been deployed in practice, or are there plans for deployment? If applicable, insights from real-world use and user feedback would help clarify its practical value.

Reviewer_HVkiOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies recommendation by generating new images rather than selecting existing items. It formulates the task as preference-field completion in continuous item space and proposes GReCF. The method represents each user with adaptive interest themes, conditions a diffusion model on the full theme set and one anchor theme, and masks the target theme during some training steps. The experiments use a new synthetic dataset and two datasets derived from real interactions.

Critical strengths
  1. The paper studies recommendation by generating new content beyond a fixed catalog. Formulating this task as preference-field completion offers an interesting perspective on extending collaborative filtering to continuous item spaces.

  2. The adaptive multi-interest representation and theme anchor provide a clear way to condition generation on several user interests. The residual preference branches allow the method to use a pretrained diffusion model while keeping its original parameters frozen.

  3. The paper introduces CIGR and evaluates GReCF on three datasets and three diffusion backbones. The experiments cover full-pool ranking, personalization, image quality, and preference extrapolation, with additional comparisons across LLM judges and human checks of their labels.

Critical weaknesses
  1. The paper does not directly evaluate whether the generated images help recommendation models learn user preferences. Retrieval-based metrics show alignment with held-out items, but do not isolate the benefit of using generated content as additional preference evidence.

  2. Generated images may contain irrelevant content or visual artifacts. Although the paper reports aesthetic and personalization scores, it does not explain how unreliable generations are detected or filtered. Such outputs could introduce noisy preference signals, particularly when reused as positive feedback.

  3. The Coverage implementation assigns every generated image to its nearest test-interest theme without a similarity threshold. An unrelated image can therefore count as covering a theme, so a high Coverage score need not indicate meaningful interest coverage.

  4. The cold-start experiment treats all generated images as positive feedback and evaluates later rounds against a history containing those outputs. Increasing similarity to this history may reflect self-reinforcement rather than a better understanding of the user's preferences.

Questions for authors
  1. How often do the generated images contain irrelevant content, and how does filtering these outputs affect personalization and subsequent preference learning?

  2. Do the cold-start gains persist when evaluated against a fixed held-out preference set or feedback obtained independently of the generator?

2 comments

Thank you for the detailed response. I still recognize the novelty and interesting idea of this paper. I have some more questions to ask: 1) For W2, would using only the similarity metric be a bit limited, and how is the Aesthetic range set? By the way, thank you for the careful response. I will keep my positive score.

We sincerely thank you for your valuable comments and for carefully considering our response. We will address your new questions point by point.

Regarding the limitations of similarity metrics. If you are concerned about the response to Weakness2, we are responding to your concern about the possibility of the model generating irrelevant content or visual artifacts. We separately analyzed the distribution of personalized scores and image aesthetic scores, proving that most of the images generated by the model are relevant and normal images. Traditional recommender systems generally use metrics such as HitRate and Recall to quantitatively measure recommendation accuracy. However, our generative recommendation method creates new content, making the evaluation of newly generated images harder. Although we use metrics from generative tasks and an LLM Judge, we still follow variants of metrics used in traditional recommendation tasks. In this setting, similarity measures provide the available approximation. Developing evaluation methods for generative recommendation will remain an important focus of our future work.

Regarding the Aesthetic Score, we adopt this metric to detect cases in which strong personalization may push the generation toward visually implausible artifacts, and to verify that the generated images remain aesthetically acceptable. The score is produced by a predictor trained from human aesthetic ratings and approximately follows a 0--10 scale, describing how much people like an image on a 1--10 rating scale. The official LAION dataset documentation provides useful empirical references: scores above 5 are generally associated with good visual aesthetics, scores above 7 are used for the LAION-Aesthetics high-aesthetic subset, and scores above 8 are used for the more artistic LAION-Art subset (LAION dataset documentation). In our CIGR SD1.5 evaluation, 98.36% of generated images score at least 5, indicating that nearly all outputs fall within the broadly acceptable aesthetic range.

Reviewer_3c8cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Existing recommendation systems select items from fixed catalogs. GReCF instead generates new images directly from user preferences, aiming to extrapolate beyond observed interests using collaborative evidence rather than generating arbitrary content.

Critical strengths
  • Formulate generative recommendation as continuous-space preference-field completion problem, coupled with a masked-interest training objective that randomly remove the target image's corresponding interest theme and train the diffusion model to reconstruct that image using the remaining themes. This is an interesting direction and the authors attempted it by applying diffusion modeling with a modified objective.
Critical weaknesses
  • Insufficient evidence that masked-interest training learns collaborative preference extrapolation. Since the diffusion model receives a noisy version of the target image and retains a pre-trained generative prior, it can minimize denoising loss without meaningfully using the remaining interest themes. Moreover, assigning an unrelated theme as the anchor introduces potentially conflicting supervision. The ablation results (Table 2) reinforce this concern: removing masking improves recommendation accuracy and personalization while reducing LLM-attributed CF extrapolation. Further analysis is needed to verify if masking learns merely weakens historical preference alignment.

  • Reliability of the evaluation protocol. The evaluation relies heavily on CIGR, a synthetic dataset with explicitly constructed collaborative structures, and an LLM judge that identifies CF extrapolation through broad semantic matching. Although human validation is provided in the appendix, annotators only verify existing LLM labels rather than independently annotating them. Different LLM judges also show substantial disagreement between History and CF attribution. More fundamentally, semantic similarity to neighboring users' interests does not establish that the target user would actually prefer the generated content. Stronger evaluation on real user preference data is needed to support the central claims.

Questions for authors
  1. Could the authors compare denoising performance under authentic, shuffled, and absent theme conditions across diffusion noise levels to assess if shortcuts appears during masked-interest training?

  2. Can the authors isolate collaborative learning from semantic associations? For example, could they evaluate under an entire-theme holdout recovery setting?

  3. How reliable is the LLM-based CF attribution? Could the authors report independent human annotations and inter-annotator agreement, or introduce randomized-neighbor controls to assess whether the judge distinguishes collaborative evidence from generic semantic similarity?

  4. Do CF-attributed generations reflect actual user preferences? Could the authors provide direct human preference comparisons or evaluate held-out positive interests on less selectively filtered real-world datasets, especially considering the recommendation accuracy degradation introduced by masked training?

7 comments

Dear Reviewer 3c8c,

Thank you again for your valuable comments and constructive suggestions on our submission. With only one day remaining in the discussion period, we would like to kindly follow up and ask whether our rebuttal has sufficiently addressed your concerns. Should any questions or concerns remain, we would be grateful for the opportunity to provide further clarification.

Best regards, The Authors

Thank you for the rebuttal. I appreciate the extended explanations on the masked training design, however the rise in CF attribution could simply mean that the model fits user history less well (especially given the accuracy dropped), and the concern about conflicting anchors is still not addressed. For evaluation, the authors agree CF attribution does not prove real user preferences, and they do not address the concerns about judge disagreement or bias in human checks. While I appreciate the clarifications, these unresolved issues remain central to my assessment of the paper, so I am unable to increase my score at this stage.

We sincerely thank you for taking the time to respond during the busy review period. We are sorry that our previous response did not fully resolve your concerns and questions. We use this final rebuttal opportunity to address the remaining points more directly.

Our task is recommendation but not personalized generation. Recommendation should extrapolate beyond observed interests rather than only reinforce existing preferences, and our training objective is therefore not designed to maximize history reconstruction alone. Our evaluation likewise does not focus only on personalization; it also considers full-pool ranking, multi-interest coverage, image quality, and interest-source attribution. Regarding your concern that the increase in CF attribution may simply reflect weaker fitting to the user's observed history, we clarify that this can indeed occur relative to the w/o Mask group, but it represents the intended trade-off between reproducing history and extrapolating beyond it. We therefore conduct the following experiment to distinguish whether Full only moves away from the user's history or also receives additional support from the true CF pool after history alignment is controlled. This is a controlled analysis of CF support, not a direct test of whether users would like the generated images. Full and w/o Mask use the same users, seeds, generation settings, and CF construction. For each output, we compute History Sim, CF Sim, and Random-CF Sim, and match outputs from the same user and seed whose History Sim differs by at most 0.02. We define CF Margin as CF Sim minus History Sim. | Condition | Matched pairs | History Sim | CF Sim | Random-CF Sim | CF Margin | True-CF - Random-CF | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Full | 640 | 0.6788 | 0.7378 | 0.6944 | 0.0590 | 0.0434 | | w/o Mask | 640 | 0.6798 | 0.7155 | 0.6925 | 0.0357 | 0.0230 | | Full - w/o Mask | -- | -0.0010 | +0.0223 | +0.0019 | +0.0233 | +0.0204 |

After matching, the two conditions have nearly identical History Sim, but Full retains higher CF Sim and CF Margin. Its advantage over the true CF pool is also larger than its advantage over the random-neighbor pool. This indicates that the CF increase cannot be explained entirely by weaker history alignment. We do not interpret this result as a complete causal proof of collaborative learning, but as evidence that masked training contributes additional CF-supported signal beyond history mismatch.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.