acceptodds
Under review as a conference paper at ICLR 2027

Generalize or Remember? Rethinking Generative Recommendation from a Data-Centric Perspective

Abstract

abstract Modern generative recommendation (GR) is largely built around generalization: semantic IDs (SIDs) allow related items to share discrete tokens and transfer knowledge across items. Yet generalization alone is insufficient for recommendation: when interaction patterns recur across users, effectively reusing such recurring knowledge is equally important. This raises a key question: when should a generative recommender generalize, and when should it remember? We revisit GR from a data-centric perspective by distinguishing interaction instances according to whether their target transitions are supported by recurring cross-user behaviors, yielding memory and generalization subsets. Our analysis shows that memory specialization improves performance on the memory subset at the cost of generalization and overall performance, while gradients from the two subsets frequently conflict. To address this trade-off, we propose MemOPSD, a memory-driven on-policy self-distillation framework. MemOPSD trains a memory-specialized teacher on recurring interactions and selectively transfers its knowledge only to memory samples while retaining the original GR objective on all training instances, strengthening memory without indiscriminately imposing memory-oriented supervision on generalization data. Since conventional distillation relies on fixed prefixes that differ from student-generated states at inference time, MemOPSD further performs on-policy distillation along student-generated SID trajectories, transferring memory knowledge at the states actually visited by the student. Extensive experiments on three benchmark datasets and four GR backbones demonstrate consistent improvements in overall and memory recommendation performance. The code is available at https://anonymous.4open.science/r/MemOPSD-0627. abstract

Primary area:
generative models
Keywords:
Generative Recommendations

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

This paper examines generative recommendation using a partition based on recurring cross-user item transitions. An instance belongs to the Memory subset when the transition from its last history item to its target occurs in another user's training history; the remaining instances form the Generalization subset. The reference index uses training data only and excludes the current user. MemOPSD trains a same-architecture teacher on Memory data and then post-trains a student with the original backbone objective on all instances plus teacher distillation on Memory instances. Distillation is evaluated at prefixes sampled from the current student, with teacher and student conditioned on the same prefix. The deployed student requires neither the teacher nor target-dependent subset labels. Experiments on three Amazon Reviews 2023 datasets and four generative backbones report consistent Overall and Memory improvements, with mixed effects on the Generalization subset. Additional experiments examine teacher and distillation configurations, model size, language model backbones, runtime, and gradient conflicts.

Critical strengths

1. The supervision scope and distillation mechanism form a coherent, testable design. Sections 3.1–3.4 clearly separate the observed-instance label, which determines whether to distill, from the student-generated SID prefix, which determines where teacher guidance is evaluated. The training-only reference construction and exclusion of the current user are explicit, and inference uses a single student. Table 2 tests continued backbone training and ground-truth-prefix KD, while Table 9 separately examines teacher data, distillation scope, normalization, and additional Memory CE. In particular, Group F improves Overall and Memory performance over Group E across the three datasets and both reported metrics. These controls support a contribution beyond merely adding more supervised updates on Memory samples.

2. The empirical gains are consistent across several backbones, and the paper reports important boundaries. Table 1 shows improvements for TIGER, LETTER, CARE, and LatentR3 across all three datasets. Tables 8 and 10 extend the analysis to larger configurations and several language model families. Table 7 explicitly reports the cases where Generalization performance declines, and Section 3.1 states that the partition does not establish out-of-distribution or compositional generalization. Appendix F.5 also distinguishes reduced gradient conflict from a causal explanation of the gains. This combination of positive evidence and explicit qualifications makes the empirical contribution more credible.

Critical weaknesses

1. The experiments expose a remaining supervision trade-off that the overall improvements do not resolve. Table 9 allows a more precise reading than the aggregate results alone. On Video Games, moving from Group D to full MemOPSD, Group F, changes the denominator from the full batch to the selected Memory samples while retaining the same teacher and Memory-only scope. Memory NDCG@10 increases from 0.1988 to 0.2085, but Generalization decreases from 0.0226 to 0.0217. Using the 14.35% Memory proportion in Table 6, the rounded values imply an Overall contribution of about +0.00139 from Memory and -0.00077 from Generalization. Thus, the better Overall score includes a substantial offsetting loss on the complementary subset. Table 7 also reports Generalization declines for CARE and LatentR3 on Office Products. The authors acknowledge trade-offs, and I do not interpret this as a hidden contradiction. The remaining weakness is that the evidence does not establish an improved Memory–Generalization trade-off at comparable supervision strength. Appendix C itself shows that the local relative KD strength depends on the Memory fraction as well as lambda. A matched-strength comparison or trade-off curve is needed to separate the value of selective transfer from a stronger emphasis on Memory under the chosen normalization.

2. The principal comparative claim lacks the most directly relevant external control. Section 2 describes adaptive MemGen-GR, which addresses the same memorization–generalization problem, and the experiments use its implementation framework. Yet Table 1 reports the constituent ID-based and generative models without the adaptive combination. The paper also cites collaborative memory augmentation, but does not compare with OMEGA. These are specific alternatives to the proposed way of exploiting cross-user knowledge. Table 9's internal controls, including the useful extra-Memory-CE control, do not determine whether MemOPSD improves the accuracy/cost trade-off over those approaches. Please prioritize a matched-data comparison and account for their different inference costs. The four named backbones provide useful within-backbone evidence, but Appendix E.3 places them in a shared lightweight T5 setup; those gains alone do not establish superiority over the broader set of published systems. This limits the demonstrated comparative contribution, even though the plug-in gains are consistently positive.

Questions for authors
  1. Separate selection from effective weighting: For Table 9, were lambda and the extra-CE coefficient tuned with comparable validation budgets for each group? Please compare selected- versus batch-normalized distillation at comparable effective strengths and report the Memory–Generalization trade-off with paired uncertainty. How much of the Overall gain remains when Generalization performance is constrained to match the relevant baseline?

  2. Closest-method comparison: Can adaptive MemGen-GR or OMEGA be evaluated on the same splits with an explicit accounting of serving and training costs? This directly tests the proposed memory-transfer contribution. Newer general competitors such as ACERec or LLaDA-Rec are additional context, with decoder compatibility assessed separately from benchmark comparability.

  3. Statistical precision: Section 4.2 reports five-run averages, while Appendix E.3 lists seed 2026 and Table 1 provides significance marks without a test specification or variability. Please identify all seeds, the test and pairing unit, and confidence intervals for Overall and both subsets, especially the reported Generalization losses. Three Amazon categories with similar short histories also leave distributional scope limited, but that fact alone does not establish that the gains are accidental.

  4. Check the cross-user exclusion in the released evaluator: In the code archive retrieved on October 18, 2026, genrec/trainer.py builds FineGrainedEvaluator from pooled training sequences and calls get_case_labels(item_seq) without the current user. mem_gen_categorizer.py then tests transition membership without subtracting that user's contribution. In contrast, the training-label export explicitly performs leave-one-sequence-out exclusion. A transition recurring only in the evaluated user's own training history is therefore labeled Memory by this evaluation path, contrary to Equation (2). I verified that conditional behavior with a minimal source-level example, but have not measured its incidence in the paper's datasets. Does preprocessing guarantee that such repeated transitions cannot occur? Please identify the code path used for Tables 6–7 and provide either that guarantee/count or labels with consistent cross-user exclusion. This is an implementation question with a stated condition, not evidence that the reported Overall gains are invalid.

  5. Complete compute accounting: Appendix D openly excludes teacher training from Table 3, and the measured runs use R=1 although Table 5 selects R=2 or R=4 in several settings. Please report teacher-plus-post-training time for those selected configurations and identify the exact experiment commands. The anonymous repository is accessible through its source/download API and contains the core training implementation; this request concerns matching the reported experiments and costs, not a claim that code is absent.

2 comments

Dear Reviewer,

Thank you again for your valuable comments and feedback. As the rebuttal deadline is approaching, we would like to kindly follow up on our responses, in which we have addressed your questions and concerns.

We would greatly appreciate it if you could revisit our responses and consider whether they warrant an upward revision of your score. If you have any remaining questions or concerns, please do not hesitate to let us know. We would be more than happy to provide further clarification.

Thank you for your time and consideration.

Best regards,

The Authors

We apologize for following up again and fully understand that this is a busy period. As the discussion period closes tomorrow, we would be sincerely grateful for any brief feedback on whether our new comparison have addressed your concerns. Even a short comment would help us greatly.

Reviewer_ZPm8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the memory–generalization trade-off in generative recommendation and proposes MemOPSD. The method trains a memory-specialized teacher and selectively distills its knowledge into the student using student-generated SID prefixes. Experiments on three datasets and four GR backbones show consistent improvements in overall and memory recommendation performance.

Critical strengths
  1. This paper addresses the memory–generalization trade-off in generative recommendation through selective knowledge transfer, aiming to enhance memory while preserving the original model's generalization ability.

  2. It achieves consistent overall performance gains across three datasets and four GR backbones. The authors also provide comprehensive ablation studies, scaling experiments, and analyses of different distillation configurations to support the effectiveness of the proposed method.

Critical weaknesses
  1. Insufficient validation of the Memory Teacher. The specialization advantage of the Memory Teacher is primarily demonstrated on TIGER, without systematic validation across other backbones.

  2. Limited evidence for an improved memory–generalization trade-off. Although MemOPSD consistently improves overall and memory performance, generalization performance decreases in some settings. The current experiments do not sufficiently establish whether selective distillation achieves a better trade-off between memory enhancement and generalization preservation.

Questions for authors
  1. It would be helpful to report the teacher's performance across all four backbones and analyze its prediction reliability on student-generated SID prefixes.

  2. Could the authors compare the memory–generalization trade off curves of Selective KD, Full KD, and Memory CE Reweighting under different distillation weights?

  3. Although the experiments are conducted within the MemGen-GR framework, the paper only adopts a binary Memory/Generalization partition without considering the existing fine-grained generalization categories. Could the authors provide performance analyses on different generalization types, such as Transitivity, Symmetry, and Substitutability, to examine whether memory enhancement affects different generalization capabilities differently?

Reviewer_Hoc4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MemOPSD, a framework for improving generative recommendation through selective on-policy distillation from a memory-specialized teacher. An instance is classified as a memory sample if its target transition appears in another user’s training history. Teacher supervision is applied only to these samples, conditioned on student-generated semantic ID prefixes, while the original recommendation objective is retained on the full training set. Experiments cover three Amazon datasets and four recommendation backbones, with additional evaluations across model sizes and language model families. The results show generally consistent gains in overall and memory-subset performance, although memory–generalization trade-offs remain in some settings.

Critical strengths

The performance differences between the standard model and the memory-specialized teacher provide a reasonable motivation for selective knowledge transfer. The method retains supervised training on the full dataset and uses only the student at inference, without requiring the teacher or target-dependent memory labels making it practical at inference time.

The paper reports generally consistent improvements across three datasets and four recommendation backbones. Table 2 and Appendix F.3 include comparisons with continued training, ground-truth-prefix KD, alternative teacher training data, different distillation scopes and normalization schemes, and memory-sample CE reweighting. These experiments support the effectiveness of the proposed combination. The authors also acknowledge that individual design choices do not improve every setting.

Critical weaknesses

The practical advantage over the closest prior approach has not been established through a direct comparison. The authors acknowledge the contributions of MemGen-GR and prior on-policy distillation methods and include relevant distillation baselines. However, they do not report a direct comparison with MemGen-GR’s adaptive fusion approach. Since the contribution centers on combining memory specialization, selective transfer, and existing distillation mechanisms, a comparison of recommendation accuracy and computational cost under matched protocols is important for assessing its value relative to the existing alternative. Single-student inference is a clear practical benefit, but the accuracy–cost trade-off relative to adaptive fusion has not been quantified.

The statistical support for the reported gains is not sufficiently documented. The paper reports averages over five runs and marks statistically significant improvements, but does not specify the test, the unit of analysis, or variability across runs.

Questions for authors

Are results available comparing MemOPSD with MemGen-GR’s adaptive fusion approach under the same data and evaluation protocols? How do recommendation accuracy, inference cost, and total training cost compare? Appendix D explicitly states that Table 3 covers only R=1 and excludes teacher training, so it cannot be taken as the full measured training cost of the configurations selected in Table 5.

What statistical test and unit of analysis were used for the significance markers? Could you report variability across runs or confidence intervals for the key comparisons, and clarify whether the Generalization-subset decreases in Table 7 are statistically significant?

Appendix C analyzes how normalization affects the effective distillation weight, and Appendix F.3 separately examines distillation scope and normalization. Was \lambda selected independently on the validation set for each configuration in Table 9? Please report the values and other relevant settings used for each group.

2 comments

Dear Reviewer Hoc4,

Thank you sincerely for your time and valuable comments, which have greatly improved our work. We have done our best to address your concerns in our responses and the revised manuscript, and we would be truly grateful to learn your current view. If any concerns remain, we would be honored to continue the discussion before the deadline.

With our deepest respect,

The Authors

We apologize for following up again and fully understand that this is a busy period. As the discussion period closes tomorrow, we would be sincerely grateful for any brief feedback on whether our new comparison have addressed your concerns. Even a short comment would help us greatly.

Comments on the paper

5 comments
Area_Chair_Wkt4OpenReview ↗

Discussion encouraged during this period!

Hi all reviewers,

We’re about halfway through the discussion period. This is a friendly reminder to check the authors' replies. If the authors have answered questions or responded to critiques in the reviews, please leave a comment for the authors!

Thank you so much!

Best, Your AC

Dear Area Chair,

Thank you for encouraging the discussion. We appreciate your help and would be happy to clarify any points in our rebuttal or answer any follow-up questions from the reviewers.

Best regards,

The Authors

AuthorsOpenReview ↗

Revised PDF Uploaded: Summary of Changes

Dear Area Chair and Reviewers,

Thank you for encouraging the discussion. We appreciate your help and would be happy to clarify any points in our rebuttal or answer any follow-up questions.

We have uploaded a revised PDF that incorporates the results and clarifications from our rebuttal, with all changes highlighted in blue. The main text contains only wording revisions, and the new results are in the appendix:

  • Appendix B: teacher diagnostics across all four backbones and on student-generated prefixes.
  • Appendix D: full training cost of the selected configurations, including teacher training.
  • Appendix E: seeds, the statistical test, and verification of cross-user exclusion.
  • Appendix F: significance of Generalization decreases, fine-grained generalization types, selected hyperparameters per configuration, and memory–generalization trade-offs under different weights.
  • Appendix G (new): comparison with MemGen-GR's adaptive fusion and OMEGA, including inference and training costs.

We would be grateful if the reviewers could take a look and let us know whether their concerns have been addressed.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.