acceptodds
Under review as a conference paper at ICLR 2027

Towards Unbiased Generative Recommendation

Abstract

Generative recommendation (GenRec) retrieves items by generating identifiers from user histories; semantic ID (SID)-based systems represent each item as a sequence of tokens. Imbalanced interactions repeatedly reinforce the token paths of popular items, potentially weakening the influence of user history. Balanced SID assignments need not balance interaction frequencies, while global token counts can hide imbalances among the valid continuations of a particular prefix. We propose UniGenRec, which explicitly factors historical frequency at each SID prefix into the training distribution alongside the history-conditioned generator. A gate conditioned on user history and SID prefix controls this factor during supervised fine-tuning; we extend the same factorization to policy optimization. At inference, the frequency branch and gate are removed; only the generator is used. Across three datasets, three SID-based GenRec configurations (TIGER, MiniOneRec, and LC-Rec), and three generative backbones, UniGenRec improves recommendation performance over the corresponding base models and leads the compared methods on MIND-small, KuaiRec-small, and Beauty. Recommendation performance is 3.38× the base model on MIND-small, 2.37× on KuaiRec-small, and 1.51× on Beauty. It also broadens catalog coverage and reduces head-item amplification relative to training clicks.

Primary area:
generative models
Keywords:
Generative Recommendation, Popularity Bias, Semantic IDs, Frequency Decomposition, Logit Adjustment

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 3 reviews, above 27% of the others.
Reviewer_AM9xOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a dibiase method for generative recommendation system, which focuses on the problem of overreliance on frequent token continuations that can drown out history-specific signals. The problem is quite an important one in modern LLM-based recommender system. UniGenRec proposes a simple yet effective factorization and training method. However, I still hold some concerns about the comparison with similar methods and its precise positioning.

Critical strengths

The paper and method are well motivated with clearly established research questions. The factorization is operationalized with a clean additive logit form where a learned gate can be easily applied. Overall, the proposed method is quite practical and easy to be applied.

Critical weaknesses

The logit additive training design and removing it at inference can be understood as the logit adjustment method, which is well learned in LLM token preference debias study. Specifically, the training objective is essentially Leanred-Mixin [1]. If the adaption from a general task to SID generation task is not trivial, the gap between should be more highlighted. [1] Don’t take the easy way out: Ensemble Based Methods for Avoiding Known Dataset Biases. EMNLP 2019. [1] also reports that the learned gate tends to collapse toward zero without an entropy penalty. What are the learned gate values across depths in UniGenRec, and does a similar collapse occur? On KuaiRec, uniform random retrieval gives roughly 1.5% R@50 (A uniform random ranker would achieve roughly 50/3,327 ≈ 1.5%), yet TIGER (1.34%) and MiniOneRec+Ghost (0.29%) fall below it. I am concerning they are relatively weaker baseline. Incorporating stronger methods can make the improvement impact of UniGenRec more convincing. Some latest SID/tokenizer works that address imbalance and prefix predictability through tokenization (for example ReSID GAOQ, CoST) should be discussed and compared. It is fine if authors think these works are not directly comparable if they can be well justified.

Questions for authors

Token-frequency bias is well studied in LLMs. What makes SID token preference fundamentally different, and why can’t these methods be applied with prefix conditioning? I am wondering what is the unique challenge here. How the hyperparameter is selected can be explicitly described. And it would be better to show the evolving pattern of the learned gate alpha value.

2 comments

Dear Reviewer,

Thank you for your careful review of UniGenRec. Your questions about the training objective, the learned gates, and the comparison with alternative methods helped us clarify our design and present additional evidence on how the gates behave during training.

W1 — Learned-Mixin and the SID setting

The logit additive training design and removing it at inference can be understood as the logit adjustment method, which is well learned in LLM token preference debias study. Specifically, the training objective is essentially Leanred-Mixin [1]. If the adaption from a general task to SID generation task is not trivial, the gap between should be more highlighted. [1] Don’t take the easy way out: Ensemble Based Methods for Avoiding Known Dataset Biases. EMNLP 2019.

Response

UniGenRec makes two design choices for SID generation.

  1. Estimate a separate historical frequency distribution at each full prefix. Consider the same second-level SID code 42. After prefix (7), it may lead to items that account for a large share of historical interactions among the valid choices. After prefix (3), it may lead to a different set of items and account for a much smaller share. We estimate these two distributions separately from earlier training interactions. The frequency branch can therefore account for a strong historical skew after (7) without applying the same adjustment to 42 after (3).
  2. Use the same decomposition in GRPO. Reward-based training can reinforce the paths the generator already favors. If popular paths dominate the rollout group, alternatives have fewer opportunities to receive credit. UniGenRec-PO carries the frequency decomposition into GRPO so that reward-weighted updates act on the frequency-adjusted policy. Rollouts come from the generator alone to avoid adding an extra preference for frequent continuations during sampling. Rewards give credit for matching the target's prefixes as well as retrieving the target item.

W2 — Gate values and collapse

[1] also reports that the learned gate tends to collapse toward zero without an entropy penalty. What are the learned gate values across depths in UniGenRec, and does a similar collapse occur?

Response

The gate outputs α∈(0,1)\alpha\in(0,1) at each SID depth; one scalar is shared by all legal next codes at that prefix. Each depth-specific gate is initialized near 0.50.5 using a zero output bias and small final-layer weights. The table reports epoch-2 mean ±\pm standard deviation of α\alpha after teacher-forced routing over the full test split (three decimal places; not sigmoid(bias)\mathrm{sigmoid}(\text{bias})). At the epoch-2 checkpoint, the mean gate value is well above zero at every depth on both datasets. One possible explanation is that the prefix-conditioned frequency estimate already captures an easy-to-fit part of the training signal. When the target code is more frequent than its alternatives, increasing the gate can directly raise its probability and reduce the training loss. Adjusting the gate gives the optimizer a simple way to fit this part of the signal and reduces how much of it the backbone needs to learn.

Datasetα\alpha at depth 1α\alpha at depth 2α\alpha at depth 3α\alpha at depth 4
MIND-small0.483±0.1560.483 \pm 0.1560.673±0.0740.673 \pm 0.0740.634±0.1420.634 \pm 0.1420.526±0.0890.526 \pm 0.089
KuaiRec-small0.684±0.0830.684 \pm 0.0830.769±0.1540.769 \pm 0.1540.780±0.0710.780 \pm 0.0710.736±0.1380.736 \pm 0.138

W3 — Baselines and random retrieval on KuaiRec

On KuaiRec, uniform random retrieval gives roughly 1.5% R@50 (A uniform random ranker would achieve roughly 50/3,327 ≈ 1.5%), yet TIGER (1.34%) and MiniOneRec+Ghost (0.29%) fall below it. I am concerning they are relatively weaker baseline. Incorporating stronger methods can make the improvement impact of UniGenRec more convincing.

Response

  1. TIGER and MiniOneRec are representative SID-based generative recommenders. MiniOneRec+Ghost additionally applies a method designed to address popularity bias. On KuaiRec, TIGER and MiniOneRec+Ghost obtain R@50 of 1.3408% and 0.2911%, respectively, compared with the uniform reference of approximately 1.50%.
  2. Figure 1 shows the concentration of MiniOneRec's recommendations on head items on MIND. Table 3 reports MIND head amplification of 10.82× for MiniOneRec and 15.51× for MiniOneRec+Ghost. On KuaiRec, Figure 2b shows that MiniOneRec gives the most popular items roughly three times their cumulative training-click share. A strong popularity bias can make a recommender repeatedly retrieve the same items that were popular during training. If those items seldom match the test targets, Recall@50 can fall below uniform random retrieval.

W4 — Relationship to SID tokenizers

Some latest SID/tokenizer works that address imbalance and prefix predictability through tokenization (for example ReSID GAOQ, CoST) should be discussed and compared. It is fine if authors think these works are not directly comparable if they can be well justified.

Response

ReSID/GAOQ and CoST intervene at tokenization. They change which codes are assigned to catalog items, addressing properties such as code quality and prefix predictability. UniGenRec addresses a different question: given a SID mapping, how can we reduce the tendency of the retriever to internalize historical interaction imbalance during training? As discussed in Related Work, even balanced code assignments can yield imbalanced training targets when popular items contribute more interactions. Replacing the tokenizer changes the item groups, their legal continuations, and the interaction frequencies attached to them.

Q1 — What makes SID token-frequency bias different?

Token-frequency bias is well studied in LLMs. What makes SID token preference fundamentally different, and why can’t these methods be applied with prefix conditioning? I am wondering what is the unique challenge here.

Response

A next-code decision in SID retrieval selects which catalog items are still reachable. For example, choosing 42 after (7) keeps the items whose SIDs start with (7, 42). Choosing 42 after (3) keeps a different set. The first set might account for most of the training interactions among the choices after (7), while the second accounts for very few after (3). The same code can therefore inherit a popularity advantage in one decision and have little such advantage in another.

A frequency correction can be adapted to SID prefixes. Adding the prefix to the model input still leaves a global or depth-level frequency correction for 42 unchanged. The challenge is to redefine what is being counted. At each prefix, we use the catalog to find the valid next codes, sum earlier training interactions over every item each code can reach, and compare those counts among the valid choices. This matters most at early steps, where pruning one continuation drops every item behind it.

Q2 — Hyperparameter selection

How the hyperparameter is selected can be explicitly described.

Response

The gate is learned during training. Its sigmoid output lies in (0, 1), and we initialize it near 0.5, the midpoint of that range. The coefficient stays at 0.5 only in the fixed-gate ablation. For frequency estimation, we use add-one smoothing so that every legal continuation has a defined probability, including those with no earlier interactions. The remaining training settings are listed in the SFT and policy optimization appendices.

Q3 — Gate dynamics during training

And it would be better to show the evolving pattern of the learned gate alpha value.

Response

We tracked the mean gate value on MIND training routes at each of the four SID depths over the five epochs shown in Appendix Figure 4. All four curves start near 0.50, then follow different trajectories. The means at depths 2 and 3 rise to approximately 0.67 and 0.63, respectively. The mean at depth 4 rises slightly to 0.53, while that at depth 1 decreases slightly to 0.48. The curves flatten toward the end of the run. This shows that the frequency contribution evolves differently across depths. At all four depths, the mean gate value stays well above zero during training.

Thank you again for your constructive feedback. We hope these clarifications and the additional gate analysis address your concerns, and we would appreciate your taking them into consideration in your final assessment.

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

The paper proposes UniGenRec, aiming to alleviate the popularity bias issue in semantic ID-based generative recommendation. It proposes to add a logit adjustment term during model training. Concretely, the next token logits during training are the sum of two terms: (1) next token prediction logits predicted by the generative model (2) a frequency factor branch that calculates the legal next-token's frequency prior to a certain timestamp.

Based on this idea, the paper describes how to extend both the SFT and RL training methods. Extensive experiments are conducted on four public datasets (three in the main paper, an additional one only in the appendix).

Critical strengths
  1. There have been papers discussing similar idea in alleviating popularity bias for conventional recommendation, but the proposed method is a technically reasonable extension of these method to next-token prediction in the generative recommendation era. It focuses on next legal continuations, rather than all tokens in the vocabulary.
Critical weaknesses
  1. Experimental setting could be questionable.
    1. In appendix C, the authors provide details about how their semantic IDs were constructed and aligned with the backbone LLMs. However, it also mentions "LC-Rec, CRAB, Ghost, and FSGR each use their own original item-to-SID procedure and do not reuse this RQ-VAE map". It's quite uncommon in generative recommendation that most of the baseline methods are not even use the same item <> semantic ID mappings, which is one of the most important component in generative recommendation.
    2. The authors processes the Amazon Beauty dataset in an unusual way. Only 3,000 items and 30,000 events are remained in the catalog after the data processing, which in my opinion, far from an amount that can support training a reasonable enough recommendation model.
    3. No code is available during the reviewing phase before rebuttal. Although I understand providing code is just recommended but not mandatory, given how large the proposed debiasing methods improve against the biased baseline TIGER on ranking metrics (9x on MIND, 3x on KuaiRec), providing code would be very helpful in understanding where the huge gains come from.
  2. Writing can be improved.
    1. A lot of terms seem to be defined by the paper itself (likely not widely used in other papers in this field), but lack explanations. For example, "dev click" in Figure 1; "frequency factor", "local fit", "item subtree", ...
    2. The title sounds a bit overclaimed. The paper focuses on alleviating a certain bias in generative recommendation (popularity), but the title sounds the proposed model is universally unbiased.
  3. The idea does not sound very new to me. The overall idea is to add a logit adjustment term calculated by next token frequency. However, the general idea has been widely explored in debiasing methods for recommendation or general machine learning, for example, Menon et al. "Long-tail learning via logit adjustment." ICLR 2021, and Zhang et al. "Causal Intervention for Leveraging Popularity Bias in Recommendation." SIGIR 2021.
Questions for authors
  1. The paper mentions that, frequency factor branch calculates frequencies on "legal continuations". How the legal continuations are calculated in the experiments?
3 comments

Dear Reviewer,

Thank you for your careful review of UniGenRec and for recognizing the value of studying popularity bias in SID-based generative recommendation. We appreciate your constructive feedback on the experimental setup, presentation, and novelty of the method. We address each concern below.

W1.1 — SID construction across baselines

Why do some baselines use different item-to-SID mappings?

Response

  1. Some baselines include their own SID construction as part of the method. We use their native pipelines so each baseline can be evaluated as designed and retain the benefits of its own codebook construction.
  2. The TIGER and MiniOneRec base models and their UniGenRec variants use the same SID construction pipeline and the same item-to-code mapping on each dataset. These comparisons hold SID indexing fixed and directly evaluate the effect of our training method.

W1.2 — Amazon Beauty dataset size

Is the retained Beauty catalog too small for training and evaluation?

Response

As reported in the dataset statistics table, the Amazon Beauty catalog contains 29,180 items. At this catalog scale, the dataset supports training the recommender and evaluating it with full-catalog retrieval.

W1.3 — Code availability

Can the authors provide code to help explain the retrieval gains over TIGER?

Response

We have included the training and evaluation code in the supplementary material.

W2.1 — Terminology and Figure 1

Please define the frequency factor, local fit, and item subtree, and explain the “dev click” label.

Response

We could have defined these terms more clearly on first use. The frequency factor is the share of earlier training interactions associated with each valid next SID code at the current prefix, with add-one smoothing. An item subtree is the set of catalog items whose SIDs start with a given sequence of codes. By local fit, we mean that at a fixed user history and SID prefix, the combined training distribution approximates the observed distribution of the next code. We will also specify the data split and click counting convention behind “dev click” in the Figure 1 caption.

W2.2 — Scope of the title

The title appears to imply a broader claim of unbiasedness than the paper supports.

Response

We appreciate the concern about the scope implied by the title. UniGenRec targets prefix-conditioned SID token-frequency bias: repeated interactions with popular items can make the model over-rely on frequent SID continuations. At the item level, we evaluate whether addressing this mechanism reduces the overexposure of historically popular items.

W3 — Novelty of the frequency adjustment

How does UniGenRec go beyond existing frequency-based logit adjustment and popularity debiasing?

Response

UniGenRec makes two contributions to SID-based retrieval.

  1. Estimate frequency among the valid choices at each full prefix. Consider the same second-level SID code 42. After prefix (7), it may lead to items that account for a large share of historical interactions among the valid choices. After prefix (3), it may lead to a different set of items and account for a much smaller share. We estimate these two distributions separately from earlier training interactions. The frequency branch can therefore account for a strong historical skew after (7) without applying the same adjustment to 42 after (3). The history- and prefix-dependent gate controls how much the frequency factor contributes during training.
  2. Carry the same decomposition into reinforcement learning. Reward-based training can reinforce the paths the generator already favors. If popular paths dominate the rollout group, alternatives have fewer opportunities to receive credit. UniGenRec-PO carries the frequency decomposition into GRPO so that reward-weighted updates act on the frequency-adjusted policy. Rollouts come from the generator alone to avoid adding an extra preference for frequent continuations during sampling. Rewards give credit for matching the target's prefixes as well as retrieving the target item.

Q1 — Legal SID continuations

How are legal next codes identified and their frequencies computed?

Response

A code is legal after a prefix if appending it still matches the beginning of at least one SID in the retained catalog. For example, suppose the catalog contains three SIDs, (7, 42, 1), (7, 42, 2), and (7, 53, 1), with 8, 2, and 4 training interactions strictly before the current target event, respectively. After (7), only 42 and 53 are legal. Code 42 gets 8 + 2 = 10 interactions because it can lead to either of the first two items. Code 53 gets 4. We add one to each legal count and normalize, giving 11/16 and 5/16. A code absent after (7) receives no probability, even if it appears elsewhere in the catalog. A legal code with zero earlier interactions still receives a smoothed count of one.

Thank you again for your time and constructive feedback. We hope these clarifications address your concerns, and we would appreciate your taking them into consideration in your final assessment.

Thank you to the authors for taking time in addressing my concerns.

Re. W1.1, I understand SID construction can be part of the baseline methods themselves. My concern is that there seem to be no debias baseline that applied to the base model TIGER and MiniOneRec that use the same SIDs as the proposed method.

Re. W1.2, from the updated PDF, we can see 38,579 events happening on the claimed "29,180" catalog size. I assume the authors have carefully double checked, does "one event for each item on average" still make sense to the authors?

Reviewer_GQSVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies token-frequency bias in SID-based generative recommendation, arguing that frequencies among the valid continuations of a full prefix better reflect the competition faced during decoding than global token counts do. UniGenRec introduces a gated historical-frequency factor into the training distribution and extends this factorization to policy optimization; inference uses only the generator. The authors report retrieval accuracy, history-sensitivity diagnostics, and exposure-distribution results across multiple SID configurations and generative backbones.

Critical strengths
  • Prefix-local competition is a useful perspective, and the method requires no additional frequency branch at inference. It makes a concrete adjustment to the structure of SID decoding, giving the contribution substance beyond a new debiasing label.
  • The empirical support is fairly strong. Table 4 shows that SFT itself accounts for the main gains, and positive results hold across several configurations and backbones. The authors are also reasonably candid about the limits of content-transductive evaluation, coverage metrics, and the PO approximation.
Critical weaknesses
  1. The additional contribution over standard training-time logit adjustment has not been sufficiently isolated. Table 5 primarily varies the gate rather than comparing global, layer-level, and full-prefix frequency statistics under the same gate. The central claim about prefix-conditioned frequency therefore still lacks a direct test.
  2. The interpretations of debiasing and personalization go beyond what these measurements identify. The frequency decomposition is not unique, and exposure and empty-history diagnostics do not establish recovery of true preferences. The mixed Industrial results also indicate meaningful limits to the settings in which the large gains hold.
Questions for authors
  • With the gate, training budget, and inference held fixed, how do global, layer-level, and full-prefix-conditioned frequency statistics compare? How much does the learned gate contribute relative to a tuned fixed coefficient?
  • Clarification: Why do the gains shrink substantially on Industrial, with some metrics falling below Ghost? Can existing gate statistics and head/tail breakdowns clarify the frequency regimes or distribution shifts in which the method is effective? I am not asking for a proof of causal unbiasedness, but I do need a precise account of the method's scope.
2 comments

Dear Reviewer,

Thank you for your careful review of UniGenRec and for recognizing the value of our work on popularity bias in SID-based generative recommendation. We appreciate your questions about the contribution of the frequency estimator, the role of the learned gate, and the scope of the empirical results.

W1 and Q1 — Frequency estimation and the learned gate

With the gate, training budget, and inference held fixed, how do global, depth-level, and full-prefix frequency estimates compare? How much does learning the gate add?

Response

We estimate frequency at each full SID prefix because the generator is choosing among the codes that can extend that prefix. The same code can lead to different groups of items after different prefixes. Global counts pool these decisions, and depth-level counts still pool prefixes at the same depth.

Frequency estimator comparison. Under MiniOneRec, the three frequency-assisted variants use the same learned gate design, training budget, and beam retrieval. Each uses SFT followed by UniGenRec-PO. Base is MiniOneRec-SFT followed by standard GRPO. Full-prefix estimation gives the best result on all four metrics on both datasets.

MIND-small (107,968 events)

SID frequency estimateRecall@10nDCG@10Recall@50nDCG@50
Base (MiniOneRec-SFT + GRPO)0.12410.05950.82150.2038
Global0.48070.23231.81160.5407
By SID depth0.55300.29042.26460.6218
By full prefix (UniGenRec)1.10960.57052.77860.9242

KuaiRec-small (20,957 events)

SID frequency estimateRecall@10nDCG@10Recall@50nDCG@50
Base (MiniOneRec-SFT + GRPO)0.37220.14761.61760.4281
Global0.51060.22802.50510.7485
By SID depth0.58690.28503.13470.8608
By full prefix (UniGenRec)1.17380.80473.84121.3452

Gate comparison. Table 5 fixes the full-prefix frequency estimator and compares the gate settings at the SFT stage. The full learned gate improves Recall@50 over the fixed coefficient of 0.5 from 1.1781% to 2.7860% on MIND and from 2.0136% to 2.5624% on KuaiRec. It also performs best on all four reported metrics on both datasets.

W2 — Interpreting exposure and history sensitivity

What do the exposure and empty-history diagnostics establish about debiasing and personalization?

Response

Figure 1 shows how UniGenRec mitigates excessive concentration on popular items. On MIND at K=50, MiniOneRec and SASRec concentrate exposure on a small set of head items, while UniGenRec distributes exposure more broadly and its cumulative curve is closer to the dev-click reference (Figure 1a–d). Table 3 quantifies this effect under MiniOneRec. Head exposure amplification relative to training clicks falls from 10.82× to 0.56×, while catalog coverage increases from 3.34% to 8.57%. The history intervention gives complementary evidence. Adding the observed history instead of an empty history increases Recall@50 by 2.49 percentage points with UniGenRec and 0.47 points with MiniOneRec. Together, these results show reduced amplification of historical popularity and a larger retrieval gain from using the observed history.

Q2 — Scope of the Industrial results

Why are the gains smaller on Industrial, and what do the subgroup results tell us about the method’s scope?

Response

Shorter user histories are one plausible explanation for the smaller gains on Industrial. Under MiniOneRec on Industrial, UniGenRec is slightly below Ghost at Recall@10 (1.2342 vs. 1.2469) and nDCG@10 (0.6150 vs. 0.6478), and slightly above Ghost at Recall@50 (3.0189 vs. 2.6799) and nDCG@50 (1.0397 vs. 0.9647). Ghost also leads TIGER at Recall@10 and nDCG@10.

Head, mid, and tail are the top 10%, the next 40%, and the remainder of items that appear as training targets, ranked by training-target count; cold items never appear as training targets. The median is the dataset-level length of the user history.

DatasetGroupEventsMedian history length
IndustrialHead9,3513
IndustrialMid8,1183
IndustrialTail9,5533
IndustrialCold43,8753
MINDHead12,33624
MINDMid28,49524
MINDTail11,71724
MINDCold55,42024
KuaiRecHead97050
KuaiRecMid1,02150
KuaiRecTail1,51250
KuaiRecCold17,45450

The median history contains only 3 interactions on Industrial, compared with 24 on MIND and 50 on KuaiRec. Shorter histories give the generator less user-specific evidence to use alongside the frequency signal, which could limit the benefit of modeling frequency separately on Industrial.

The retrieval gains are substantial on datasets with longer user histories. Under MiniOneRec, explicitly modeling historical frequency raises Recall@50 from 0.82% to 2.78% on MIND and from 1.62% to 3.84% on KuaiRec. This is particularly relevant to news and short-video recommendation, where content can lose relevance quickly and past popularity may be less useful for predicting what a user wants next.

Thank you again for your constructive feedback. We hope these comparisons and clarifications address your concerns and that you will take them into consideration in your final assessment.

Comments on the paper

2 comments
Area_Chair_dqDfOpenReview ↗

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

AuthorsOpenReview ↗

Response to the Area Chair and All Reviewers

Dear Area Chair and Reviewers,

Thank you for your careful reviews and constructive feedback. We appreciate your recognition of the importance of studying popularity bias in generative recommendation. UniGenRec makes three contributions:

  1. It identifies where popularity bias enters SID retrieval. A model generates an item one code at a time. Popular items can repeatedly reinforce particular choices, and an early choice can rule out whole groups of items. UniGenRec models the historical frequency of the codes that actually compete at each step.
  2. It brings the same approach to supervised learning and reinforcement learning. Reward-based training can further reinforce paths the model already favors. UniGenRec accounts for historical frequency in both stages and fits into existing SID recommenders using their current item index and serving pipeline.
  3. It improves recommendation performance across several generative recommendation frameworks. It also reduces excessive exposure of popular items in our MIND evaluation. These results show that recommendations can become less concentrated while retrieval improves.

Our point-by-point responses address the main concerns raised in the reviews:

  1. Frequency estimation and the learned gate. Under matched settings, full-prefix frequency estimates outperform global and depth-level estimates on MIND and KuaiRec. The gate ablation shows an improvement over a fixed coefficient of 0.5. We also added a figure showing how the learned gate changes during training.
  2. Novelty and related work. We clarified why the same SID code can have different frequencies after different prefixes and how UniGenRec incorporates this distinction into GRPO. We also explained how training with a fixed SID mapping addresses a different problem from redesigning the tokenizer.
  3. Experimental comparisons and scope. We clarified which baselines share a SID mapping and which use their own indexing. We also explained what the exposure and user-history comparisons measure and discussed the smaller gains on Industrial.

The new material in the revised manuscript is highlighted in dark blue for easy reference. Thank you again for your constructive feedback. We hope our responses address your concerns, and we would appreciate it if you would reconsider your ratings after reading the updated manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.