acceptodds
Under review as a conference paper at ICLR 2027

Role-Factorized Semantic IDs: Diagnosing and Mitigating Information Loss in Generative Recommendation

Abstract

Recent advances in generative modeling have led to growing interest in generative recommendation. Central to this paradigm is the semantic ID (SID), a short sequence of discrete tokens representing an item. Existing methods often construct SIDs by fusing content and behavior into a single continuous representation before quantization. However, our preliminary study shows that this early-fusion strategy can lead to a representational compromise, where the learned item geometry is less aligned with the relational structures of both sources. To address this issue, we propose Role-Factorized Semantic IDs (RF-SID), which assigns content and behavior to separate coordinate blocks within a unified SID. This design represents both relations without requiring every discrete coordinate to reconcile them. Experiments on three public benchmarks show that RF-SID consistently outperforms state-of-the-art recommendation methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Generative Recommendation, Semantic IDs, Representation Learning, Item Tokenization, Residual Quantization

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_Q5cjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes RF-SID for generative recommendation. It assigns separate Semantic ID coordinates to content and behavioral representations and uses a linear predictor to remove content-related information from the behavioral component. Experiments on three Amazon datasets report improvements over existing SID construction methods under sampled candidate ranking. The paper also presents a simplified rate-distortion analysis to motivate the proposed factorization.

Critical strengths

The separation of content and behavioral information in the semantic ID construction presented in this paper is intuitively sound, and the proposed method is easy to implement.

Critical weaknesses

W1. Separating semantic and collaborative signals through different codebooks, token groups, or generation streams has already been extensively explored, such as GenSAR , FlexCode, EAGER and DiscRec. The proposed method largely reduces to separately quantizing two representations, applying a standard linear residualization step, and concatenating their codes. The paper neither adequately distinguishes itself from these closely related approaches nor provides direct comparisons with the most relevant alternatives. It is difficult to identify a substantial new technical contribution.

W2: The evaluation protocol does not establish the claimed generative recommendation performance. All main results are based on ranking one positive item against 1000 sampled negatives, despite catalogs containing more than 150,000 items as in Table 1. So the reported superiority over existing methods may not hold under realistic retrieval.

W3: A more fundamental concern is that complete SID uniqueness is neither established nor evaluated. Different items sharing the same complete SID necessarily receive identical scores, yet the paper reports no collision statistics.

W4: Finally, the experimental evidence is limited to one training seed (line 658-667).

Questions for authors

see the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Q5cj

We thank the reviewer for the feedback. We conducted additional comparisons, full-catalog evaluations, collision analyses, and multi-seed experiments to address all four concerns.

W1: Novelty and Related Work

RF-SID differs from existing methods in how content and behavioral signals are organized within a unified, fixed-budget SID. Unlike GenSAR's shared quantization, EAGER's two-stream generation, FlexCode's dynamic token allocation, and DiscRec's gated embedding fusion, RF-SID uses content-first coordinates and content-adjusted behavioral residuals within a single autoregressive generation stream.

We additionally evaluated GenSAR, EAGER, and FlexCode on all three datasets (W2). RF-SID consistently outperforms these methods under full-catalog evaluation. Removing content adjustment also reduces NDCG@10 from 0.008998/0.009508/0.008240 to 0.008653/0.009345/0.008132 on Beauty/Sports/Toys.

Notably, FlexCode exhibited severe code collapse in our reproduction, with complete-SID collision rates of 44.15%/44.57%/36.10%. Its token allocation concentrated on zero or one behavioral token, and adjustments to pretraining coverage and learning rate did not resolve the issue. Since FlexCode's code is not publicly available, these findings raise substantial concerns about its reproducibility and robustness.

W2: Full-Catalog Evaluation

We conducted exhaustive full-catalog ranking over 207,649/156,235/162,035 items on Beauty/Sports/Toys, using the same 50,000 test contexts across methods, without sampled negatives or beam-search truncation.

Full-catalog Recall@10 / NDCG@10:

MethodBeautySportsToys
SASRec.01638/.008379.00828/.004034.00728/.003778
BERT4Rec.01844/.009596.01216/.006193.01154/.006012
Random SID.00916/.004675.00502/.002486.00474/.002146
MQL4GRec.01784/.009350.01670/.008681.01538/.008043
TIGER.01238/.006355.01344/.007215.01086/.005586
LETTER.01032/.005391.00508/.002505.00362/.001689
CCFRec.01856/.009700.00866/.004497.00756/.003864
UNGER.01370/.007282.01114/.005606.00832/.004276
GenSAR.01140/.006152.00742/.003706.00478/.002331
EAGER.01358/.008386.01650/.009188.01442/.007795
FlexCode.00642/.003586.00442/.002689.00540/.002779
RF-SID.01788/.008998.01846/.009508.01556/.008240

RF-SID achieves the highest NDCG@10 on Sports and Toys and outperforms all compared SID-based methods except CCFRec on Beauty. It consistently outperforms GenSAR, EAGER, and FlexCode across all three datasets, confirming that its advantages persist under exhaustive full-catalog ranking.

W3: Complete-SID Collisions

We measured complete-SID collisions over the entire catalog before collision resolution, using CollisionRate=100(1−U/N)%\mathrm{CollisionRate}=100(1-U/N)\%, where UU is the number of unique complete SIDs.

MethodBeauty (%)Sports (%)Toys (%)
MQL4GRec7.098513.618611.0785
TIGER12.89725.07956.8942
LETTER0.77820.36360.7393
CCFRec0.00000.23360.2074
UNGER0.04430.00190.0148
GenSAR0.00870.00060.0031
EAGER0.00000.00000.0000
FlexCode44.150944.572036.1002
RF-SID0.15410.11390.1018

RF-SID achieves approximately 99.85%–99.90% unique complete SIDs without any collision-resolution procedure, with at most three items sharing one SID.

During evaluation, all items remain separate ranking candidates, including those sharing a SID. Tied scores are handled consistently across methods using the average rank r=G+(E+1)/2r=G+(E+1)/2, where GG and EE count candidates with greater and equal scores, respectively. Thus, the reported results do not benefit from item deduplication or favorable tie-breaking.

W4: Multi-Seed Stability

We completed three independent runs (seeds 2026, 2027, 2028), each covering tokenizer training, AR training, and full-catalog ranking over 50,000 test contexts.

Full-catalog results (mean ± sample standard deviation):

DatasetRecall@10NDCG@10
Beauty0.017480 ± 0.0005290.008910 ± 0.000342
Sports0.017933 ± 0.0004610.009251 ± 0.000237
Toys0.016053 ± 0.0006280.008462 ± 0.000281

The NDCG@10 standard deviations are only 3.84%/2.56%/3.32% of the corresponding means, demonstrating stable performance across independent training runs.

Thank you to the authors for their rebuttal. The response has partially alleviated my concerns about W2, W3, and W4, but it has not fully addressed my concerns about the novelty of W1. The distinction between RF-SID and existing methods still requires more thorough justification.

Furthermore, the current revision still uses 1000 negative samples and adds neither new comparisons nor collision results, so I do not consider these issues formally resolved. In addition, the latest full-catalog ranking results in the rebuttal differ substantially from the performance reported in the original submitted paper, especially in the comparisons between methods, making them highly misleading.

Therefore, I still believe maintaining my score is appropriate.

Reviewer_vaa6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RF-SID for generative recommendation (GR). The method separately quantizes item content and content-adjusted behavior, and combines their tokens into one fixed-length semantic ID. Experiments on three Amazon datasets show improvements over the compared methods under sampled evaluation. The authors also study the effects of content adjustment, token allocation, and token order.

Critical strengths
  1. This paper provides a simple way to combine content and behavioral information within a fixed-length semantic ID, without changing the downstream generative recommender.

  2. The ablation studies show the contributions of role factorization and content adjustment.

  3. The experiments on token allocation and order provide further analysis of the SID design.

Critical weaknesses
  1. The evaluation ranks each target against only 1,000 sampled negatives. It is unclear whether the improvements hold when all items are considered. If the sampled negatives contain few similar items, the setting may favor RF-SID's content-prefix design without sufficiently testing fine-grained ranking.

  2. The novelty seems limited, as the method combines separate SID blocks for content and behavior with standard residualization and residual quantization. The baseline comparison also lacks related methods, such as EAGER [1], DiscRec [2], and UNGER [3].

  3. The theoretical analysis assumes that early fusion averages the two sources, which already removes their differences before quantization. It is unclear whether the conclusion applies to more general fusion methods. The reconstruction probes also use only two SID positions, which may disadvantage joint codes that distribute information across all four positions.

  4. Some implementation details are missing. How are the projection matrices obtained and how is the ridge regularization parameter selected? Without these details, it is difficult to reproduce the method.

  5. This paper does not explain how the uniqueness of SIDs is guaranteed. Since these items have identical sequence scores, please clarify how they are mapped back to individual items and how ties are handled during evaluation.

[1] EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration

[2] DiscRec: Disentangled Semantic-Collaborative Modeling for Generative Recommendation

[3] UNGER: Generative Recommendation with A Unified Code via Semantic and Collaborative Integration

Questions for authors
  1. Are the SASRec behavioral embeddings trained using only the training interactions?

  2. What are the complete-SID collision rates, and how are collisions and tied scores handled for each method?

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

The paper studies how item content and interaction behavior should be represented within a fixed-length semantic ID. RF-SID quantizes the two sources separately and assigns them dedicated positions in one identifier. It also removes the component of the behavioral representation predicted from content before constructing the behavior tokens. An autoregressive recommender predicts these fixed identifiers. The experiments cover three Amazon Reviews 2023 categories and examine representation alignment, source reconstruction, token allocation, and token order.

I find the problem well motivated and the proposed design useful. Content similarity and behavioral similarity capture different item relationships, so giving each source dedicated positions is a reasonable choice. The main value of the paper is the empirical evidence that this choice improves recommendation quality under a controlled identifier budget and downstream architecture.

Critical strengths
  1. A simple representation design with a clear practical motivation. The method offers an understandable way to preserve complementary content and behavioral signals within one compact identifier. It can be used without redesigning the downstream recommender. This is a useful contribution for practitioners deciding how to represent heterogeneous item information. (Section 4.)

  2. The central comparison controls the identifier budget and downstream model. SID-based methods use the same four-token budget and the same recommendation architecture. This makes the comparison informative about tokenizer design and reduces the risk of attributing gains simply to a larger identifier or a different decoder. (Section 5.1.)

  3. The ablation provides credible evidence for source separation. Role factorization without content adjustment improves NDCG@10 over the tested early-fusion baseline from 0.1409 to 0.1740 on Beauty, 0.1087 to 0.1241 on Toys, and 0.1048 to 0.1311 on Sports. The complete method improves further to 0.1848, 0.1262, and 0.1470, respectively. The consistent improvement before adding content adjustment is particularly useful evidence that dedicated source positions contribute to the result. (Table 3.)

  4. The additional experiments make the design actionable. The paper studies how to allocate positions between sources and how to order the resulting tokens. Content-first ordering performs best among the tested arrangements, while allocation sensitivity differs across domains. The alignment and held-out reconstruction analyses offer further insight into how the representation behaves. (Tables 4-6 and Appendix C.)

Critical weaknesses
  1. The fusion comparator needs a more precise specification. The exact operation used to combine the sources before quantization is not sufficiently clear. This is the most important remaining clarification because early fusion is the central ablation comparator. Concat-RQ is specified in Appendix C but evaluated only for reconstruction; its recommendation results would strengthen the comparison. UNGER, discussed in Section 2.2 as a joint semantic-collaborative tokenizer, is also a relevant comparator whose omission should be explained. The current results support an advantage over the evaluated fusion implementation, without establishing superiority over the full family of fusion methods.

  2. The information-loss interpretation should be narrower. CKA measures alignment, and the two-position reconstruction probes in Appendix C measure how readily each source can be recovered from part of the identifier. A fused representation may distribute information across all four positions. These results therefore support source accessibility more directly than a general claim of lower total information loss. The simplified rate-distortion model provides intuition under its stated fusion assumption. I would distinguish that intuition from the demonstrated recommendation gains and either add a full-identifier reconstruction probe or qualify the broader interpretation.

  3. The evaluation scope and variability need qualification. The shared evaluation ranks each target against 1,000 sampled negatives and uses one training seed (Section 5.1 and Appendix A.2). It is a valid controlled comparison of candidate scoring, but does not establish full-catalog decoding quality or efficiency. Some margins over ID-based models are small: on Toys, NDCG@10 is 0.1262 versus BERT4Rec's 0.1240. Repeated runs for the principal comparisons would improve confidence. Full-catalog decoding is a valuable extension; the acceptance case here rests on the demonstrated tokenizer contribution.

Questions for authors
  1. What exact fusion operation and source preprocessing are used for the early-fusion ablation? Can the projection-fitting procedure and ridge regularization setting be specified?

  2. How does Concat-RQ perform in the recommendation task under the same budget and decoder, and why is UNGER omitted?

  3. How stable are the main tokenizer comparisons across training seeds? Please distinguish the current sampled-ranking evidence from any claims about full-catalog generative retrieval.

1 comment
AuthorsOpenReview ↗

Response to Reviewer NsMA

Thank you for the constructive review. We added Concat-RQ, UNGER, full-four-position probes, and three-seed full-catalog ranking.

Weakness 1: Fusion comparison

The added Concat-RQ control concatenates separately normalized 128-dimensional content and behavior projections, then applies four-level residual quantization. Each level uses ten full Lloyd KMeans restarts, selected by training quantization error, with at most 300 iterations each. Complete-SID collision rates are 0.0026%–0.0116%.

RF-SID uses the learnable content-prefix/conditional-behavior-suffix tokenizer. Both methods use four 256-way positions and the same AR trained for five epochs. Results below are mean ± sample SD over seeds 2026, 2027, and 2028.

DatasetMethodRecall@10NDCG@10
BeautyConcat-RQ0.013227 ± 0.0006850.007187 ± 0.000443
BeautyRF-SID0.017480 ± 0.0005290.008910 ± 0.000342
SportsConcat-RQ0.011407 ± 0.0002200.005862 ± 0.000202
SportsRF-SID0.017933 ± 0.0004610.009251 ± 0.000237
ToysConcat-RQ0.007973 ± 0.0003720.004064 ± 0.000235
ToysRF-SID0.016053 ± 0.0006280.008462 ± 0.000281

RF-SID improves mean NDCG@10 by 23.97%, 57.82%, and 108.20%. Tokenizer budgets are reported separately: RF-SID trains for 80 epochs; Concat-RQ performs 40 Lloyd fits per run.

Weakness 2: Information-loss interpretation

We added probes reading all four SID positions. Methods share raw normalized content/behavior targets, a 70/15/15 item split, and a 854,528-parameter MLP trained for 30 epochs with validation selection. The table reports content / behavior cosine similarity; RF-SID uses its seed-2026 encoding.

MethodBeautySportsToys
RF-SID0.8677 / 0.57600.8639 / 0.46660.8520 / 0.4453
TIGER0.8666 / 0.55710.8631 / 0.42500.8470 / 0.4013
MQL4GRec0.8516 / 0.53010.8281 / 0.40550.8155 / 0.3784
Average fusion0.8017 / 0.58570.7880 / 0.47870.7711 / 0.4238
UNGER0.7303 / 0.59540.7308 / 0.50400.7249 / 0.4801

RF-SID reconstructs content more accurately, while UNGER reconstructs behavior more accurately. We will frame these results as finite-probe recoverability and content–behavior trade-offs. CKA measures alignment; the theoretical argument applies under its stated fusion assumptions.

Weakness 3: Evaluation scope and variability

Each domain uses the same 50,000 test contexts, scoring all 207,649 Beauty, 156,235 Sports, or 162,035 Toys items. Ties receive their average rank.

RF-SID and Concat-RQ independently retrain the tokenizer and AR for each seed, keeping teachers, preprocessing, validation candidates, and test queries fixed. Their three-seed results appear above; SD uses n−1n-1. These supplementary statistics correspond to full-sort evaluation.

Question 1: Exact fusion, projection, and ridge settings

Let norm⁡\operatorname{norm} denote L2 normalization:

u=norm⁡(c)Pc,v=norm⁡(b)Pb.u=\operatorname{norm}(c)P_c,\qquad v=\operatorname{norm}(b)P_b.

Each projection outputs 128 dimensions. Entries are sampled from N(0,1/128)\mathcal{N}(0,1/128) with seed 2026 and frozen. Ordinary average fusion uses norm⁡((u+v)/2)\operatorname{norm}((u+v)/2).

Set C=norm⁡(u)C=\operatorname{norm}(u) and B=norm⁡(v)B=\operatorname{norm}(v). Conditional behavior is

W=(C⊤C+10−3I)−1C⊤B,R=norm⁡(B−CW).W=(C^\top C+10^{-3}I)^{-1}C^\top B,\qquad R=\operatorname{norm}(B-CW).

Ridge uses summed squared error and no intercept. The average-fusion probe uses norm⁡((C+R)/2)\operatorname{norm}((C+R)/2); Concat-RQ uses norm⁡([C;B])\operatorname{norm}([C;B]). Learnable RF-SID encodes raw normalized content and the fixed residual separately.

Question 2: Concat-RQ and UNGER recommendation results

Concat-RQ results are given above. We also added UNGER using shared frozen teachers, semantic adaptation, cross-modal alignment, and next-item learning. Its last-history-query DIN-style module trains for 20,000 steps, followed by four-level residual KMeans and the shared AR. These implementation choices replace its original input and Stage-II pipelines.

UNGER’s single-run full-sort results are:

DatasetRecall@10NDCG@10
Beauty0.013700.007282
Sports0.011140.005606
Toys0.008320.004276

Question 3: Seed stability and retrieval scope

RF-SID exceeds Concat-RQ in Recall@10 and NDCG@10 in all nine matched dataset–seed comparisons, with lower cross-seed SD in each domain. History-filtered ranking gives the same ordering.

The original experiments measure sampled ranking; the added experiments measure exact full-catalog ranking. Beam-search recall and end-to-end decoding efficiency are separate evaluation targets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.