acceptodds
Under review as a conference paper at ICLR 2027

Can Generative Recommenders serve as User Data Simulators?

Abstract

User behavior traces are the foundation of recommender systems, but high-quality traces are difficult to obtain. Online A/B testing provides direct feedback but is costly and risky, while existing logs are fixed records of past exposure policies. User simulation offers an offline alternative, yet its usefulness depends on whether generated traces provide useful signals for downstream tasks. This paper explores whether generative recommenders can serve as user data simulators. A generative recommender models the conditional distribution of future user behavior given user history, profile, and context; recommendation and user behavior simulation are two interfaces to the same learned behavior distribution. Building on this observation, we present GenSim, a unified generative user data simulation framework that turns the generative recommender backbone into a controlled simulation engine. To make the framework practical, we enhance it with two capabilities: the first on long-horizon generation, which combines test-time scaling with reinforcement learning on verifiable rewards to control repetition, context drift, and suffix degradation across generated traces; the second on new user synthesis, which synthesizes new user profiles together with compatible behavior traces to construct additional user contexts. We instantiate the framework with OpenOneRec and evaluate it along two axes: behavioral characteristics, measuring response-prediction performance and distributional alignment between generated and observed data, and downstream utility, measuring whether generated traces improve recommendation training. Our results show that generative recommenders can serve as practical engines for synthetic behavior trace generation, providing useful supervision for downstream training, with benefits varying across models and tasks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Generative Recommendation, User Simulator

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_LjzAOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GenSim, a framework that repurposes Generative Recommendation model as a generator of synthetic user profiles and interaction traces. It combines candidate selection and reinforcement learning to improve long-horizon generation, and constructs additional users through profile fusion and behavior generation.

Critical strengths

The evaluation covers several complementary aspects: behavioral distribution, response prediction, future-item prediction, and RL policy training. Testing multiple downstream architectures provides a useful view of where synthetic data helps and where its benefits are limited.

Critical weaknesses
  • The intended deployment scenario is insufficiently defined. GenSim updates the OpenOneRec backbone through RL, but the downstream evaluation primarily uses its generated data to train other recommenders. The paper does not establish when this indirect route is preferable to directly deploying the updated OpenOneRec as the recommender. Constraints such as model size, serving cost, or compatibility with an existing recommender could justify this choice, but they are not clearly specified or evaluated.

  • The necessity and mechanisms of the added components are insufficiently isolated. In Sec. 3.2, candidate selection and RL optimization share several behavioral scoring terms, but Tables 3–4 only remove the overall long-horizon module or profile-consistency component. These ablations do not distinguish the contributions of candidate selection, RL, and the suffix-specific reward. Their necessity is also not consistent across models: removing LH improves TIGER’s target-prediction results in Table 4, while removing PC slightly improves InterFormer’s AUC in Table 3.

Questions for authors
  1. Are total optimization steps matched between the no-augmentation and augmentation settings in Table 3/4? Would repeating real-data samples under the same training budget yield similar improvements?
  2. How does GenSim compare with existing recommendation data-generation methods, such as DR4SR, under a matched data budget? What are the additional training and generation costs?
  3. For Table 1, does “freq = 0” also imply absence from the OpenOneRec-8B-Pro pretraining corpus? Are the training data and available item information comparable with KuaiSim?
  4. It seems that the reference of DGDPO in Table 5 is missing.
Reviewer_s6woOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies whether a pretrained generative recommender can be reused as a user-data simulator. GenSim uses OpenOneRec as an autoregressive behavior generator and adds: (i) long-horizon generation using chunk-wise best-of-N/test-time scaling plus GSPO-style RL with verifier rewards, and (ii) synthesis of new user profile–trace pairs. The evaluation covers cold-item response prediction, distributional diagnostics, downstream data augmentation, and RL-based recommendation. The central claim is that generative recommenders can act as practical offline simulators and synthetic-data engines for recommendation.

Critical strengths
  1. Important practical question: reusing a strong generative recommender as an offline simulator could reduce dependence on costly online experimentation.

  2. Broad validation intent: the paper uses cold-item AUC, distributional diagnostics, downstream augmentation, ablations, and RL transfer rather than a single synthetic-data metric.

  3. Engineering care in parts of the supplement: provenance hashes, deterministic split logic, replay checks, matched seeds, and unit tests are present.

Critical weaknesses
  1. Novelty is limited relative to prior user simulation and data-centric SR. Existing literature already covers generative/adversarial user models for RL recommendation, LLM/generative-agent simulators, and dataset regeneration/synthetic sequence augmentation. GenSim appears primarily to combine a pretrained generative recommender with existing search/post-training/profile-mixing components. The paper needs a sharper novelty boundary and stronger controls against these prior paradigms.

  2. There is a direct numerical contradiction below Table 3. The text states that vanilla OneRec traces “even hurt TIGER and HSTU,” but the table shows improvements for both: TIGER 0.6729→0.6751 and HSTU 0.6703→0.6745. For InterFormer, removing PC gives 0.6549 versus 0.6545 for full GenSim, so the table does not support a positive PC contribution for that model.

Questions for authors
  1. The downstream TIGER/HSTU implementations appear to be adaptations rather than verified faithful reproductions of the original reference implementations. This is acceptable in principle, but the paper treats them as standard strong baselines without quantifying implementation fidelity, weakening cross-architecture conclusions.

  2. Major mismatch in new-user/profile synthesis. The paper describes demographics/intents/textual profile fusion, an 8,192-token context, and an “eighth-layer hidden representation.” The submitted Table-3 code defaults to a prefix-derived profile built from coarse SID-cluster summaries. Its profile embedding routine averages several final hidden layers rather than using only the eighth layer. The evaluated implementation therefore appears materially simpler than the manuscript description.

  3. The manuscript calls the label task “profile-and-sequence to label prediction,” but the supplied downstream response student for TIGER/HSTU consumes history, a candidate item, and optional feedback-rate features rather than the natural-language user profile described in the paper. Therefore the claimed profile-conditioned interpretation is not supported by the submitted implementation.

Reviewer_oQkpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies whether a generative recommender can be used as a user simulator. The idea is that such a model already learns what a user is likely to do next, so sampling from it can produce synthetic behavior traces.

The authors build GenSim on top of OpenOneRec and add two parts. For long traces, they generate several candidate chunks, keep the best-scoring one, and also fine-tune the generator with GSPO using shared behavioral reward components and an additional tail-quality reward. For new users, they pair similar training users, merge their profiles, and generate a new history starting from a mix of the two parents’ histories.

Experiments on RecIF-Bench cover cold-item AUC (Table 1), how closely generated traces match real ones (Table 2), whether synthetic traces help train downstream recommenders (Tables 3–4), and RL agents trained in different simulators (Table 5).

Critical strengths
  1. The framing is simple and useful. A generative recommender already predicts what a user will do next, and sampling traces from it is a natural way to build a simulator. This matters to my assessment because the direction is worth pursuing even though the current evidence is weak, and it is the main reason I see value in the paper.

  2. Table 1 gives concrete evidence for one advantage of this choice. On items never seen in training, the OpenOneRec backbone reaches 0.696 AUC, while KuaiSim is almost random (0.494). This is consequential because cold items are exactly where logged data and discriminative simulators fall short. I think the result can clearly support the paper’s motivation.

Critical weaknesses
  1. The two components added on top of OpenOneRec are not well justified.

(1) The paper does not show that long traces are needed. The paper does not report a controlled comparison between long traces and shorter continuations under matched settings, or results across different generation horizons. Although repetition and tail quality enter the reward design, the experiments do not report direct diagnostics of repetition rates or degradation across trace positions. In addition, it remains unclear whether the consistency-related reward terms distinguish undesirable generation drift from legitimate changes in user interests. The method incorporates recent as well as long-term interests, but the paper does not evaluate whether generated traces exhibit realistic interest changes.

(2) Each new user is built from two existing users who are chosen because they are similar, so the result is essentially a mix of two known users. This conflicts with the stated goal of covering rare or underrepresented users. Figures 1–2 show a rare-user discovery step, but Section 3.3 never explains it. The paper also does not report how similar the new users are to their parents, and it does not evaluate on cold-start users.

Why it matters: these two components are the paper’s main contribution beyond sampling from an existing model. However, the authors failed to establish the advantage and necessity of their core designs. It is unclear what GenSim adds over plain OneRec generation.

  1. The experiments do not support the main claims.

(1) For TIGER and HSTU in Table 3, the additional AUC gains over plain OneRec augmentation are modest: 0.0017 and 0.0013, respectively. Their gains over training without synthetic data are 0.0039 and 0.0055. InterFormer shows substantially larger gains, so the benefits vary across learners. There is no comparison with standard data-augmentation methods or with the LLM-based simulators discussed in Section 2. Because the generator is much larger than the downstream models, the gains may come from passing the generator’s knowledge to smaller models rather than from simulating users.

(2) The ablation results do not match the conclusions. For TIGER in Table 4, removing either component gives higher Recall@32 than the full method. The text also says plain OneRec traces hurt TIGER and HSTU, but Table 3 shows they improve both.

(3) Table 5 is used to argue that policies trained in GenSim transfer better to real users. However, the supplementary README says the nDCG@10 numbers do not come from the trained A2C/SA2C agents. Instead, the rankings come from a fixed combination of scores plus random noise, so this table does not measure policy transfer.

Why it matters: the paper’s main claims are that synthetic traces help downstream training and that policies trained in GenSim transfer to real users. The current experiment results fail to support them.

Questions for authors
  1. The supplementary README says the nDCG@10 numbers come from a fixed mix of scores and noise, not from the trained A2C/SA2C agents. Is this how the reported numbers were produced? If you can report nDCG@10 for the trained agents on held-out users, this would directly affect my view of the RL claim.

  2. In RecIF-Bench, 20% of users are held out for testing. The released data (96M interactions, 160K users) is the training portion, and the standard OneRec checkpoints were trained on it. Which users are evaluated in the downstream and RL experiments, and were their interactions in the backbone’s training data? If they were, the reported gains may overstate the benefit.

  3. How are labels assigned to synthetic traces in the label-prediction task? With gains of only 0.004–0.006 AUC, the answer determines whether the improvements are meaningful.

Reviewer_DazzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether pretrained generative recommenders can serve as user data simulators. The proposed framework, GenSim, builds on OpenOneRec and introduces two capabilities: long-horizon trajectory generation through chunk-wise candidate selection and GSPO-based optimization of behavioral proxy rewards, and synthetic user construction through profile and history composition. The paper evaluates response prediction, distributional alignment, downstream recommendation training, and reinforcement-learning applications. It reports that synthetic data improves several downstream models, although the benefits vary across architectures and tasks.

Critical strengths

The research direction is useful and practically motivated. Reusing a pretrained generative recommender as a synthetic-data engine is a promising way to exploit learned behavioral representations. The paper goes beyond unconstrained sampling by explicitly addressing repetition, interest consistency, transition plausibility, and suffix degradation. Defining these criteria makes the generation process more inspectable and controllable.

The downstream augmentation results provide encouraging evidence of utility. The comparison with vanilla generation under the same backbone and synthetic-data budget is a meaningful control. Improvements across several downstream architectures suggest that controlled generation can produce useful supervision, while the reported negative results for TIGER also reveal that the benefits are not universal.

Critical weaknesses

The paper does not sufficiently examine what makes a synthetic interaction informative, particularly under exposure bias and heterogeneous user histories. Recommendation logs are strongly shaped by the historical exposure policy. Consequently, matching logged item transitions, co-consumption statistics, and embedding-based interests does not establish that a simulator captures users’ latent preferences; it may primarily reproduce the patterns induced by previous recommendations. The proposed rewards provide an explicit operational definition of quality, which is useful, but the paper does not adequately validate whether optimizing these rewards improves independently measured behavioral fidelity. This also leaves an important question about the purpose of synthesis unresolved. For users with rich histories, generation may infer valuable additional interactions from well-supported preferences. For cold-start or sparse-history users, the need for augmentation is greater, but the conditioning information is weaker and synthetic behaviors may mostly reflect population priors. The paper does not sufficiently distinguish these regimes or establish where its synthetic supervision adds information beyond similar-user transfer, history composition, and teacher soft labels. The cold-item results in Table 1 do not answer this cold-user question. I do not consider an online A/B test mandatory for this work. However, more direct offline evidence connecting the proposed quality scores to held-out user responses, together with a focused analysis across user-history regimes, is needed to substantiate the simulator interpretation and explain the augmentation gains.

Questions for authors

What evidence establishes the external validity and incremental value of the proposed generation-quality criteria? In particular, do higher proxy scores correspond to better held-out response prediction or other independent behavioral measurements? Please clarify whether the gains concentrate on rich-history users or sparse-history users, and what evidence distinguishes the benefit of controlled generation from similar-user/history transfer and supervision from the frozen response teacher. This clarification would materially affect my assessment of the methodological contribution.

Please reconcile the reported experiments with the supplementary implementation. For Table 5, which exact code evaluates saved A2C/SA2C actor checkpoints on a disjoint real evaluation set? For Tables 3–4, please specify the actual split, training and synthetic-data sizes, checkpoint-selection procedure, and final evaluation set. Please also identify the profile-embedding and composition settings used to obtain the reported results. If the supplied pipeline differs from the experimental version, please explain the differences and provide the corresponding implementation and per-seed results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.