acceptodds
Under review as a conference paper at ICLR 2027

MaLaCo: Masked Latent Completion for Efficient Generative Recommendation

Abstract

Semantic-ID (SID) recommenders retrieve items by generating short discrete identifiers, but masked-diffusion decoding repeatedly processes every beam hypothesis through the full history backbone. We introduce masked latent completion (MaLaCo), which factorizes retrieval into a single bidirectional history-completion pass followed by trie-constrained SID generation with a small autoregressive head. Beam-dependent computation is therefore confined to the head. On one A100, at beam width 50, MaLaCo reduces amortized generation cost from to ms/query at batch () and single-request latency from to ms (), despite a higher total parameter count relative to LLaDA-Rec, a leading generative retrieval baseline. Across four datasets, MaLaCo improves Recall@ over LLaDA-Rec at every reported cutoff \(k\geq10\), with larger gains at deeper cutoffs. It trades Recall@1 for stronger deep recall on Industrial & Scientific, Musical Instruments, and Video Games, while also improving Recall@1 on Arts, Crafts & Sewing. A four-seed \(2\times2\) architecture–encoder study confirms that this depth-dependent advantage persists under either item encoder, while the stronger encoder consistently benefits MaLaCo. We further introduce a collision-aware SID construction pipeline combining balanced residual quantization, last-level deduplication, learned candidate refinement, and catalog-level assignment. Across all four catalogs, it reduces the excess-collision rate from \(18.12%\)–\(31.04%\) to at most \(0.0122%\). On an item-arrival partition of Arts, the method inserts the newest quarter of the catalog without changing existing identifiers or introducing new collisions. After recommender retraining, the extended vocabulary obtains \(0.04526\) Recall@10 versus \(0.04598\) after complete vocabulary refitting, a relative shortfall. Together, these results position MaLaCo as an efficient, high-recall candidate generator with collision-aware and stably extensible semantic identifiers.

Primary area:
generative models
Keywords:
generative recommendation, semantic IDs, masked diffusion, efficient inference, catalog expansion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UDRTOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MaLaCo, a semantic-ID-based generative recommendation framework that separates history encoding from item-ID generation. A bidirectional Transformer completes a masked next-item latent representation in a single pass, and a small autoregressive head performs trie-constrained SID decoding, confining beam-dependent computation to the head. The paper also introduces a collision-aware SID construction pipeline and a procedure for adding items while preserving existing identifiers. Experiments on four Amazon datasets report improved Recall@k over LLaDA-Rec at all reported cutoffs k >= 10, with a trade-off in Recall@1 on three datasets. At beam width 50 on an A100, the reported speedup is 24.9x for amortized generation cost at batch size 20 and 2.4x for single-request latency. Additional validation includes a four-seed architecture–encoder comparison, collision measurements, and an item-arrival experiment in which catalog expansion preserves existing IDs and approaches the recall of full vocabulary refitting after recommender retraining.

Critical strengths
  • The paper reflects a substantial amount of implementation and experimental work.

  • It provides extensive implementation details, explanations, and hyperparameter configurations.

Critical weaknesses
  • Limited novelty. The core implementation is a systematic combination of masked item modeling, continuous item representations, autoregressive SID generation, trie-constrained beam search, contrastive learning, residual quantization, and collision reduction. I recognize the engineering complexity involved in integrating these components in an industrial setting. However, as an academic contribution, the work lacks a clear central insight that could be broadly useful to the research community. Components such as trie-constrained decoding, collision handling, and residual processing have already appeared in prior work such as OneRec. Although this paper is developed primarily in relation to LLaDA-Rec, I do not find the underlying techniques fundamentally different. Combining so many components also makes it difficult to attribute the reported gains to the central contribution.

  • A clear trade-off in recommendation quality. The authors acknowledge that MaLaCo generally underperforms LLaDA-Rec on Recall@1, while its advantages mainly emerge as kk increases. This suggests that the method improves coverage deeper in the candidate list but is less effective at placing the most relevant item first. In other words, MaLaCo improves recall at larger cutoffs at the cost of top-ranked accuracy. This appears to require a larger beam size to obtain the reported accuracy, raising questions about how this requirement interacts with the claimed inference-cost advantage.

  • Poor readability. The paper introduces too many concepts at once and moves quickly into equations and multiple technical branches without first presenting a concrete example of the complete inference process. In particular, SID construction contains four stages, but their relationship to recommender training is explained in a fragmented manner.

Questions for authors

Please refer to the weaknesses part.

4 comments

W1: central insight, prior components and attribution

What we do not claim. Trie-constrained beam search, residual quantization with collision handling, contrastive auxiliary losses and masked item modeling all appear in prior work (OneRec, TIGER, BERT4Rec). So does encoding the history once and decoding SIDs in a separate decoder (TIGER; concurrently, CoSID). The revised Sections 1–2 credit this work and state that we do not claim these components as new.

Central insight. Beam search in a generative recommender does not need to re-read the history. One target state, trained by masked completion over continuous item representations, carries enough information for a small head to decode the whole beam. This makes two testable claims: (i) giving the head the whole encoded history adds nothing over that single state; (ii) masked completion over continuous inputs, rather than the surrounding components, is what makes the state good. The identifier table, to which new items can be added without changing any existing identifier, is a separate contribution (App. E).

Attribution. Each row changes one factor and keeps the rest fixed. R@10, four seeds:

FactorChangeSciInst
History model (new)TIGER → MaLaCo backbone; same SID table, both decoders read the whole history0.0415 → 0.04870.0538 → 0.0614
Head input (new)MaLaCo (whole history) → single state (Table 1)0.0487 → 0.04990.0614 → 0.0621
SID table (Table 5)random collision-free / RQ-VAE + last-level Sinkhorn → ours0.0426 / 0.0481 → 0.0499–
SID Stages 3–4 (new)Stages 1–2 only → all four0.0497 → 0.0499–
  • Masked completion is the main source of the gain. It is the largest single factor (+0.024 / +0.028). With identifiers and decoding aligned, continuous masked completion (inputs and objective together) is also what separates the MaLaCo backbone from TIGER (+0.0072 / +0.0076).
  • Bidirectional attention matters only through masked completion. With next-item training it makes no difference. Under masked completion, a causal backbone loses 14% / 17%.
  • The single state is sufficient, claim (i). Letting the head cross-attend to every history state does not help and doubles the cost per query (1.03 vs 0.526 ms at batch 20).
  • Identifiers. The gain over RQ-VAE tables comes from the semantic hierarchy of Stages 1–2. Stages 3–4 add no recall; they serve contribution 2.
  • InfoNCE (App. D) helps at every cutoff, from +0.0010 at R@1 to +0.0062 at R@50. App. D's rows use different test subsets, so we rely only on the direction.
  • Item encoder (Table 2). A stronger encoder raises MaLaCo's R@10 by +0.0027 / +0.0016 and LLaDA-Rec's by +0.0000 / −0.0001, consistent with MaLaCo's continuous inputs passing the encoder's information through to the model.

Relation to LLaDA-Rec. Both mask and complete, but they differ in what is completed and where beam search runs. LLaDA-Rec completes discrete SID tokens; MaLaCo completes continuous item representations. A history item enters MaLaCo through its own text-derived vector, not its quantized ID. A newly added item can therefore appear in users' histories as soon as its vector is computed, and a stronger item encoder helps MaLaCo but not LLaDA-Rec (above). LLaDA-Rec re-runs its full backbone for every hypothesis at every denoising step, whereas MaLaCo runs its backbone once and beam search runs only in the 4-layer head. Its cost is therefore nearly flat in beam width: from beam 10 to 50, LLaDA-Rec's cost per query grows 4.6× while MaLaCo's barely changes (Comment 2). The system-level gain is 24.9× at batch 20. Against TIGER, an optimized KV-cached encoder–decoder that also encodes the history once, MaLaCo is 3.4–3.5× faster and more accurate.

W2: the Recall@1 trade-off and beam width

The trade-off is real. Table 1, LLaDA-Rec / MaLaCo:

SciInstGameArts
R@10.0097 / 0.00890.0122 / 0.01110.0198 / 0.01630.0081 / 0.0085
R@100.0466 / 0.04990.0615 / 0.06210.0939 / 0.09510.0433 / 0.0460
R@500.1081 / 0.11730.1420 / 0.14670.2122 / 0.22440.0986 / 0.1089

MaLaCo is below LLaDA-Rec at R@1 on three datasets (by most on Game) and above it on Arts. It leads at R@10 and R@50 on all four. NDCG@10 is mixed: higher on Scientific and Arts, lower on Instrument and Game. We position MaLaCo as a candidate generator whose list feeds a ranking stage, where R@10–R@50 matter most. We do not claim top-1 gains. The abstract already states this trade-off, and the revision states it in the introduction as well.

Does the accuracy require a large beam? We re-decoded our trained checkpoints at beam 5–100 (Scientific and Instrument, four seeds, full test set), and LLaDA-Rec's App. F checkpoint at beam 10 and 20. Each entry is the fraction of that model's own beam-50 value retained:

BeamMaLaCo R@1, Sci / InstMaLaCo R@10, Sci / InstLLaDA-Rec R@1, SciLLaDA-Rec R@10, Sci
598% / 97%–––
10100% / 100%92.5% / 89.6%99%90.3%
20100% / 100%98.4% / 97.9%100%98.4%
100100% / 100%100.3% / 100.2%––
  • R@1 does not depend on the beam. It is unchanged from beam 10 to 100 for MaLaCo and from beam 10 to 50 for LLaDA-Rec. The R@1 gap therefore does not come from the beam width, and a wider beam neither causes nor closes it.
  • MaLaCo does not need a wider beam than LLaDA-Rec. Both lose almost the same fraction of R@10 at beam 10 and 20, and MaLaCo gains nothing beyond beam 50.

How beam width interacts with cost. This is where the design pays off. Beam width multiplies LLaDA-Rec's whole backbone, but only MaLaCo's small head. Per query, batch 20, Scientific:

BeamMaLaCo (App. F)LLaDA-RecRatio
100.503 ms2.84 ms5.6×
200.498 ms5.41 ms10.9×
500.526 ms13.10 ms (App. F)24.9×
  • MaLaCo's cost is flat in beam width (+5% from beam 10 to 50), while LLaDA-Rec's grows 4.6× and its peak memory from 626 to 3,013 MiB.
  • MaLaCo at beam 50 costs less than LLaDA-Rec at beam 10 (0.526 vs 2.84 ms). At beam 50 it is also more accurate at R@10 and R@50 on all four datasets (Table 1). The large beam is cheap for MaLaCo, so the reported accuracy does not depend on an expensive setting.
  • TIGER, which also encodes the history once, is likewise flat in beam width (1.80 / 1.80 / 1.79 ms) but 3.4–3.6× slower than MaLaCo at every width.
  • The LLaDA-Rec runs at beam 10 and 20 use App. F's timing hook. Its beam-50 run in the same setup reproduces App. F (13.11 vs 13.10 ms).

The revised App. F will add both tables.

W3. The revision opens Section 3 with a short end-to-end inference example and an overview figure before any equations. It states that SID construction runs once per catalog, before recommender training, which uses the resulting table only as SID targets and as the decoding trie.

Thanks for your replay. I will keep my score.

Thank you for reading our response. Could you please let us know if there is any specific point that we did not address clearly, especially about novelty, the ablation results, or the difference from prior work? We would be happy to clarify it briefly. Thank you again for your time and feedback.

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

MaLaCo combines bidirectional completion of continuous item-history representations with a small autoregressive semantic-ID head. It also proposes collision-aware SID construction and vocabulary extension that preserves existing identifiers. Experiments show stronger recall at deeper cutoffs and lower inference cost than LLaDA-Rec, alongside top-rank trade-offs and the need for recommender adaptation when adding new items.

Critical strengths
  1. Useful efficiency results. The method reduces inference cost relative to LLaDA-Rec. Appendix F further isolates the throughput benefit of moving beam-dependent computation out of the history backbone, distinguishing it from end-to-end savings.

  2. Meaningful experimental controls. The architecture–encoder factorial examines the role of the text encoder, while Appendix G compares three masking policies with validation-selected checkpoints on two datasets.

  3. A useful study of cold start. ( A useful distinction between naming and retrieval.) The catalog-growth study evaluates identifier stability and recommendation separately, showing that assigning new items stable SIDs does not by itself enable a frozen recommender to retrieve them.

Critical weaknesses
  1. The architectural advance needs a closer comparison. TIGER already separates history encoding from autoregressive SID generation [1]. The efficiency analysis in Appendix F compares MaLaCo with LLaDA-Rec and a repeated-backbone control, rather than a trained encoder–decoder that reuses its history encoding. The autoregressive SID head is not, by itself, the main novelty; the question is what continuous completion and the compact target-state interface add beyond this established separation. A comparison with aligned item information, SID construction, comparable capacity, and a shared decoding protocol would better establish that accuracy–cost benefit. The existing encoder and masking controls are valuable, but do not resolve this architectural question.

    [1] Rajput, S., Mehta, N., Singh, A., Keshavan, R. H., Vu, T., Heldt, L., Hong, L., Tay, Y., Tran, V. Q., Samost, J., Kula, M., Chi, E. H., & Sathiamoorthy, M. (2023). Recommender systems with generative retrieval. Advances in Neural Information Processing Systems, 36.

  2. The benefit of bidirectional history contextualization remains unclear. Appendix G supports the masking strategy, but all its variants retain bidirectional attention; SASRec/BERT4Rec comparisons also change other design choices. Since a causal model's terminal prediction position can access the observed history, the remaining question is what bidirectional contextualization adds. A focused causal-versus-bidirectional comparison under otherwise matched terminal-only training would help answer this. This concerns the mechanism behind the gains, rather than whether the full system improves on the evaluated baselines.

Minor Issues

  • Table 1 presents RPG and LLaDA-Rec alongside MaLaCo, while the broader comparisons appear in Appendix B. Although the main text points to the appendix, this layout makes the experimental coverage appear narrower than it is. A compact main-text summary of the additional baselines would improve the presentation.

  • Section 2 could better position the SID pipeline relative to SaviorRec's Sinkhorn-based quantization [2] and HiD-VAE's collision-aware representation learning [3].

    [2] Yao, Y., Li, Z., Xiao, S., Du, B., Zhu, J., Zheng, J., Kong, X., & Jiang, Y. (2025). SaviorRec: Semantic-behavior alignment for cold-start recommendation (arXiv:2508.01375). arXiv.

    [3] Fang, D., Gao, J., Zhu, C., Li, Y., Zhao, X., & Chang, Y. (2025). HiD-VAE: Interpretable generative recommendation via hierarchical and disentangled semantic IDs (arXiv:2508.04618). arXiv.

  • important Section 2's LIGER citation points to Lan et al.'s LLM linearization paper, rather than the intended recommendation work [4]. Please correct the entry.

    [4] Yang, L., Paischer, F., Hassani, K., Li, J., Shao, S., Li, Z. G., He, Y., Feng, X., Noorshams, N., Park, S., Long, B., Nowak, R. D., Gao, X., & Eghbalzadeh, H. (2024). Unifying generative and dense retrieval for sequential recommendation (arXiv:2411.18814). arXiv.

I view this as a borderline submission and currently lean against acceptance. The efficiency results and controls have value, but the architectural advance and the contribution of bidirectional contextualization are not yet sufficiently established. These reservations concern the evidence for the methodological contribution, not a fundamental flaw in the approach. A convincing rebuttal addressing these two issues could change my recommendation to acceptance; a modest, well-supported improvement would be a worthwhile contribution.

Questions for authors
  1. Comparison with history-reusing generative recommenders. Could you compare MaLaCo with TIGER or a representative trained encoder–decoder that encodes history once and batches candidate-beam expansion in its SID decoder? I would like to see both recommendation quality and inference cost under aligned item information, SID construction, comparable capacity, and matched decoding conditions, including beam width, batch size, hardware, and caching policy.

  2. Isolating bidirectional contextualization. Could you add a causal-versus-bidirectional comparison with the remaining architecture and supervision held fixed, for example using terminal-only training in both variants? The terminal prediction position can access all observed history in either case, so this comparison would isolate the value of bidirectional interaction within the history.

5 comments
AuthorsOpenReview ↗

what MaLaCo adds beyond TIGER's separation (quality)

Thank you for the careful review and for stating exactly what evidence would settle each question.

We agree with the premise. TIGER already encodes the history once and expands the beam in a separate decoder. The revision will credit TIGER in Sections 1–2 and ask the question as you do: what do continuous completion and the compact target-state interface add beyond this separation?

Design. Three trained systems, each differing from the next in one respect:

History inputObjectiveSID decoder readsParams
A: TIGERSID tokens (3 per item)next itemall encoded tokens (cross-attention)4.81M
B: MaLaCo, cross-attention headcontinuous item representationsmasked completionHH + all backbone states7.32M
C: MaLaCocontinuous item representationsmasked completionHH only6.19M

(A) and (B) are fully aligned. They share the item representations (TIGER sees them through the SID table), the same 3 × 256 SID table, the same 20-item windows and splits, trie-constrained beam 50 over the same catalog, validation-selected checkpoints and seeds 2024–2027. Both decoders read the whole encoded history; only the backbone's input and objective differ.

(B) is MaLaCo's architecture with one change: its head also cross-attends to every backbone state, the information TIGER's decoder has. (C ) is MaLaCo as reported in Table 1. TIGER variants. (A) keeps TIGER's published architecture and optimizer (4+4 layers, dmodeld_{model} 128, dffd_{ff} 1024, Adafactor, 200k steps). (A′) widens dffd_{ff} to 1698, matching MaLaCo's 6,194,560 parameters. TIGER's official code is not publicly available, so we also implemented its full published pipeline (Sentence-T5 embeddings, RQ-VAE identifiers with an appended disambiguation digit). Dropout for (A′) and the full pipeline was tuned on validation R@10.

Four seeds, mean ± std, full test set:

ModelR@1R@10R@50N@10
SciTIGER as reported by Shi et al. (2025)0.00840.0446–0.0236
(A) TIGER0.0071±.00020.0415±.00120.1009±.00180.0214±.0006
(A′) TIGER, 6.19M, dropout tuned0.0071±.00020.0411±.00060.1009±.00110.0213±.0004
TIGER, full published pipeline0.0061±.00030.0371±.00050.0942±.00160.0190±.0003
(B) MaLaCo, cross-attention head0.0088±.00010.0487±.00060.1151±.00110.0256±.0004
(C) MaLaCo (Table 1)0.0089±.00040.0499±.00110.1173±.00180.0261±.0008
ModelR@1R@10R@50N@10
InstTIGER as reported by Shi et al. (2025)0.01050.0566–0.0300
(A) TIGER0.0099±.00030.0538±.00080.1314±.00100.0283±.0005
(A′) TIGER, 6.19M, dropout tuned0.0100±.00020.0558±.00100.1337±.00160.0293±.0004
TIGER, full published pipeline0.0091±.00020.0511±.00130.1247±.00190.0267±.0006
(B) MaLaCo, cross-attention head0.0113±.00060.0614±.00110.1464±.00160.0322±.0009
(C) MaLaCo (Table 1)0.0111±.00020.0621±.00080.1467±.00120.0325±.0003
  • Continuous masked completion is the source of the gain (A --> B). With aligned inputs, identifiers and decoding, and a decoder that reads the whole encoded history in both, (B) exceeds TIGER by +0.0072 R@10 on Scientific and +0.0076 on Instrument, and by +0.014 and +0.015 R@50.
  • The compact interface costs no accuracy (B --> C). Reading only HH instead of every history state leaves every metric within seed variation (R@10 0.0499 vs 0.0487; 0.0621 vs 0.0614), while halving the cost per query (Comment 2).
  • Capacity and tuning do not explain it. At MaLaCo's exact parameter count, with dropout tuned, TIGER does not improve (0.0411 / 0.0558). Its validation recall peaks early (epochs 35–60) and then declines, whatever the width or dropout.
  • MaLaCo exceeds every TIGER row on every metric, including TIGER's reported numbers on these splits.
AuthorsOpenReview ↗

Inference cost under matched decoding conditions

inference cost under matched decoding conditions

Matched conditions. Every system of Comment 1 was timed on the GPU of App. F and with its protocol: one A100-SXM4-40GB, fp32, TF32 off, CUDA-synchronized calls, warm-up discarded, median of repeated calls, no other job on the GPU, over batch {1, 4, 20, 64, 128, 256} × beam {10, 20, 50}.

  • Caching. TIGER's encoder runs once per query. Its decoder uses a self-attention KV cache, and cross-attention keys and values are computed once per query and reused at every step. MaLaCo's backbone runs once per query; its head recomputes its prefix (at most 4 tokens) at each step without a KV cache, which if anything disadvantages MaLaCo.
  • Implementation. We time TIGER with HuggingFace generate, as its evaluation calls it, and with a hand-written beam loop that returns identical ranked lists with less per-step overhead. We report the faster.
  • Trie. Every 3-digit system applies the catalog trie as a GPU table lookup. TIGER's per-hypothesis Python prefix function is replaced by an equivalent tensorized one that admits exactly the same tokens on every prefix tested.

Where the cost differs structurally. In each of its 4 decoder layers, TIGER cross-attends to the 62 encoder tokens (a user token, 60 SID tokens and EOS) for every hypothesis at every step. After its single backbone pass, MaLaCo's head reads only HH, so its per-hypothesis cost does not grow with history length. (B) adds that cross-attention back, over 20 states.

Scientific, beam 50ms/query, b=1b=20b=128peak MiB, b=128
LLaDA-Rec (App. F)24.4513.1012.6219,131
(A) TIGER35.31.790.4515,880
(A′) TIGER, 6.19M35.91.790.4715,884
TIGER, full published pipeline (4 digits)43.12.140.6867,695
(B) MaLaCo, cross-attention head19.11.030.554896
(C) MaLaCo (App. F)9.980.5260.239326
  • Speed. MaLaCo is 3.5× faster than TIGER at batch 1 and 3.4× at batch 20. The gap narrows to 1.9× at batch 128, where both systems keep the GPU busy. The ratios are the same against the 6.19M TIGER (3.6×, 3.4×, 2.0×), and larger against HuggingFace generate (4.6× and 4.8×) and against the full published pipeline (4.3×, 4.1×, 2.9×).
  • Memory. MaLaCo needs 18× less memory than TIGER at batch 128 (326 vs 5,880 MiB) and 19× less at batch 256 (599 vs 11,661 MiB), because TIGER keeps per-hypothesis caches.
  • The single-state interface. (B) roughly doubles MaLaCo's time per query and nearly triples its memory, with no gain in accuracy (Comment 1). That is the price of reading the history once per hypothesis, and it is the cost the single-state interface removes.
  • Consistency. The MaLaCo and LLaDA-Rec rows are App. F's. Re-timed in the session that timed TIGER and (B), both agree with App. F within 4%. The full batch × beam sweep goes in the appendix.

App. F's control. As you note, App. F's repeated-backbone control is not a trained decoder: it has no accuracy to go with its speed, and it performs no token scoring, valid-code masking or beam selection. The revision uses it only to illustrate how the gap scales with batch size, beam width and history length. The TIGER comparison above answers your question with trained systems that both run a complete beam search.

AuthorsOpenReview ↗

Isolating bidirectional contextualization

We agree the question was open. App. G varies only which positions are masked, and every arm is bidirectional; SASRec and BERT4Rec change other design choices as well.

Design. We added a causal-attention option to MaLaCo's backbone: each history position attends only to itself and earlier positions. The prediction slot comes last, so, as you note, it sees the whole observed history in both modes. The option is a single flag. With it off, the backbone's output is bit-identical to the original; with it on, earlier positions are unaffected by later ones.

Everything else is copied from the App. G arms:

  • item representations and SID table;
  • backbone and head sizes;
  • SID and InfoNCE terms with the same coefficients;
  • optimizer, batch size and epoch budget;
  • beam 50 and seeds;
  • App. G's checkpoint rule, best validation Recall@10. This is why App. G's MaLaCo figures differ slightly from Table 1, which selects by validation loss.

Within each row of the table, only the attention mask differs. Entries are R@10 / R@50, four seeds:

Sci, bidirectional (App. G)Sci, causal (new)Inst, bidirectional (App. G)Inst, causal (new)
Terminal only0.0263±.0006 / 0.0714±.00070.0267±.0006 / 0.0719±.00010.0349±.0011 / 0.0929±.00240.0350±.0001 / 0.0936±.0010
Random subset0.0500±.0011 / 0.1173±.00090.0430±.0001 / 0.1044±.00050.0627±.0009 / 0.1478±.00070.0521±.0009 / 0.1276±.0013
  • Terminal-only training, the comparison you proposed. Causal and bidirectional backbones are within seed variation (0.0267 vs 0.0263 on Scientific; 0.0350 vs 0.0349 on Instrument). As you anticipated, the prediction slot sees the whole history either way.
  • Random-subset masking. The causal backbone loses 0.0070 R@10 on Scientific (−14%) and 0.0106 on Instrument (−17%), with no overlap between seeds. Only bidirectional attention lets a masked interior slot be completed from both sides. Under a causal mask it sees only earlier items, so the objective reduces to next-item prediction at sampled positions.
  • The two factors interact. Masked completion helps even a causal backbone (+0.0163 R@10 on Scientific, +0.0171 on Instrument), and bidirectional context adds to it only when interior positions are masked.

Bidirectional attention therefore matters because of the masked-completion objective, not on its own. The revision will state the mechanism this way, attributes the gain to masked completion with bidirectional context rather than to attention direction alone, and reports this 2×2 in a new appendix.

minor issues

Table 1 coverage. Agreed. Table 1 gains a compact block with SASRec, BERT4Rec and LC-Rec at R@10, plus TIGER: the reported numbers on these splits and the runs of Comment 1. Full results remain in App. B.

SaviorRec and HiD-VAE. Section 2 now positions the SID pipeline against both:

SaviorRec (Yao et al., 2025) replaces the arg-min assignment of an RQ-VAE with Sinkhorn optimal transport to improve codebook utilization and prevent collapse, and uses the resulting IDs as features of a CTR ranking model rather than as generation targets. HiD-VAE (Fang et al., 2025) supervises each quantization level with hierarchical item tags and adds a uniqueness loss that penalizes overlap between the pre-quantization latents of items whose full IDs coincide within a training batch. MaLaCo also builds its hierarchy with Sinkhorn-balanced assignment, so balancing itself is not our contribution; it differs in how residual collisions are removed. Last-level transport is re-run only within each group of colliding items, leaving coarser codes unchanged by construction, and the conflicts that remain are handled by catalog-level assignment over learned candidates rather than by a soft penalty. The same assignment step inserts new items without changing any existing SID.

LIGER. Thank you. The revision will cite the correct work.

Reviewer_cpamOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

This paper proposes MaLaCo for efficient Semantic ID based generative recommendation. It separates history encoding from SID generation by using a bidirectional history encoder and a lightweight autoregressive prediction head. The paper also introduces a multi-stage SID construction pipeline to reduce collisions and support catalog extension.

Critical strengths

The paper addresses relevant efficiency and scalability challenges in generative recommendation. The reported throughput improvements and stronger recall@KK at larger KK could be useful for candidate generation.

Critical weaknesses

W1. Encoding user history separately from autoregressive SID generation is already a standard design in encoder-decoder recommenders such as TIGER. The closely related concurrent work CoSID also uses a single history encoding pass followed by lightweight SID decoding. The additional masked training and contrastive objectives are based on established techniques, and the paper does not provide convincing evidence that their combination constitutes a substantial new contribution. More importantly, there is no direct comparison with an optimized autoregressive encoder-decoder baseline.

W2: The reported 24.9x speedup is obtained against a diffusion-based model with different backbone depths, sequence lengths, and decoding procedures. In the authors' own matched decoding control, the speedup drops to 3.26x at batch size 20, while single-request latency is slightly worse. This substantially weakens the argument that the large efficiency gain comes from the proposed architectural design.

W3: The additional SID construction pipeline combines residual quantization, Sinkhorn assignment, an autoencoder, reinforcement learning, and catalog-level conflict resolution, but the necessity of this complexity is not sufficiently demonstrated. Simpler collision-resolution approaches are not adequately compared in terms of recommendation quality, construction cost, and catalog maintenance.

W4. There are also reproducibility and presentation issues. The authors do not release source code, checkpoints, or intermediate artifacts, making independent verification of the complex pipeline and efficiency measurements difficult. The related work contains an incorrect citation for LIGER, referring to Lan et al. rather than the generative retrieval work by Yang et al (Unifying Generative and Dense Retrieval for Sequential Recommendation).

Questions for authors

see the weaknesses.

4 comments
AuthorsOpenReview ↗

W1: novelty and a direct encoder–decoder comparison

Separating history encoding from SID decoding is not our contribution. We agree it is standard: TIGER encodes the history once and decodes SIDs in a separate decoder. The revision will say so in Sections 1–2 and credits TIGER and CoSID. We now claim two things, each tested below and in Comment 3:

  1. Continuous masked completion as the training signal for a single target state, which a small head decodes without reading the history again so the outpu can be cached.
  2. Stable catalog extension. A three-digit semantic identifier table with near-zero collisions (0.0039% excess on Scientific, Table 4) to which new items can be added without changing any existing identifier and without creating new collisions. A deployed recommender therefore keeps a valid output vocabulary as the catalog grows. Inserting the newest 22,490 Arts items changed no existing SID and added no collision (see Table 6 and App. E).

Direct comparison with an optimized encoder–decoder (new). TIGER's official code is not publicly available, so we implemented it ourselves in three variants. Every TIGER row uses a KV cache, trie-constrained beam 50, the best-validation checkpoint and the same splits as Table 1. Rows are four-seed means ± std; costs are in Comment 2.

Sci R@10Sci R@50Inst R@10Inst R@50
TIGER, published architecture, SIDs from our pipeline (4.81M)0.0415±.00120.1009±.00180.0538±.00080.1314±.0010
TIGER, widened to MaLaCo's 6.19M, dropout tuned0.0411±.00060.1009±.00110.0558±.00100.1337±.0016
TIGER, full published pipeline (Sentence-T5, RQ-VAE, 4th digit)0.0371±.00050.0942±.00160.0511±.00130.1247±.0019
TIGER as reported by Shi et al. (2025), same splits0.0446–0.0566–
MaLaCo, head cross-attends to all history states0.0487±.00060.1151±.00110.0614±.00110.1464±.0016
MaLaCo (Table 1)0.0499±.00110.1173±.00180.0621±.00080.1467±.0012
  • MaLaCo exceeds every TIGER row on both datasets. With the same splits, decoding and SID construction pipeline, the margin is R@10 +0.0084 on Scientific and +0.0083 on Instrument, and R@50 +0.016 and +0.015.
  • Capacity is not the explanation. At MaLaCo's exact parameter count, with dropout tuned on validation, TIGER does not improve (0.0411 / 0.0558).
  • What the decoder reads is not the explanation. Letting MaLaCo's head cross-attend to every history state, as TIGER's decoder does, does not help (0.0487 vs 0.0499; 0.0614 vs 0.0621).
  • Our implementation of TIGER's full pipeline. We followed the published configuration. It reaches 0.0371 / 0.0511, below the 0.0446 / 0.0566 reported by Shi et al. Fewer than 0.01% of its decoded identifiers are invalid, so decoding is not the cause. We report the numbers we obtained; MaLaCo also exceeds the reported numbers.

Masked training and the contrastive term are established techniques. We agree and claim neither as new. Our claim is about the role of masked completion over continuous item states: it is what trains the single state the head decodes. With the architecture and data fixed:

  • Masking (App. G). Terminal-only (next-item) training reaches R@10 0.0263 on Scientific and 0.0349 on Instrument. Random-subset masking reaches 0.0500 and 0.0627.
  • Attention direction (new). We added a causal-attention option to the backbone.
    • With terminal-only training, causal and bidirectional backbones match (0.0267 vs 0.0263 on Scientific; 0.0350 vs 0.0349 on Instrument).
    • With random-subset masking, the causal backbone loses 0.0070 on Scientific (0.0430 vs 0.0500) and 0.0106 on Instrument (0.0521 vs 0.0627), with no overlap across seeds.
    • Bidirectional context therefore pays off only through masked completion.

CoSID. We were not aware of CoSID when we submitted. It appeared afterwards (RecSys 2026, 27 Sep 2026), and we found no public code, so a direct comparison is not possible. We will cite it as concurrent work.

Like MaLaCo, CoSID encodes the history once and lets a light decoder run the beam search. The two differ in how the history state is learned and in the identifiers:

  • History state. CoSID uses a causal backbone over SID-based item tokens, trained for next-item prediction with help from a teacher model. MaLaCo uses a bidirectional backbone over continuous item representations, trained from scratch by masked completion.
  • Identifiers. CoSID keeps standard residual-quantized SIDs and does not address catalog growth, which our identifier table is designed for.
AuthorsOpenReview ↗

W2, where the efficiency gain comes from

We agree that 24.9× is a system-level comparison with the released LLaDA-Rec, whose depth, history length and decoding all differ from ours. Both systems run a complete beam search (width 50, invalid codes masked), and both are timed on the generation call only, on the same GPU. The comparison takes each system at its validated configuration, where MaLaCo is also more accurate (Table 1). It is therefore the practically relevant number, and the revised abstract and introduction label it as system-level. Against an optimized KV-cached encoder–decoder (TIGER, below), MaLaCo is 3.4–3.5× faster and more accurate.

What the matched-decoder control measures (correction). The control runs MaLaCo's backbone four times, once per denoising step, over batch × beam inputs: as many sequences as LLaDA-Rec processes after copying each history once per beam hypothesis. It does nothing else. It performs none of the token scoring, valid-code masking or beam selection that LLaDA-Rec performs after every step, whereas MaLaCo's timing includes its complete beam search. App. F stated that the two arms differ only in where beam search runs. They also differ in that the control performs no search at all. The control therefore measures only the backbone cost of that decoding pattern, and every ratio against it is a lower bound:

  • Batch 20 and above. MaLaCo is at least 3.26× faster at batch 20 and at least 7.43× faster at batch 256. The rest of the 24.9× reflects LLaDA-Rec's deeper backbone, longer sequence and implementation, and the search the control omits.
  • Batch 1. The control appears faster (0.91×) only because it skips search. Its four passes take 9.1 ms, about 2.3 ms each since the GPU is not saturated. MaLaCo spends 2.25 ms in its backbone and 7.7 ms in its beam search. Against systems that do run a complete beam search, MaLaCo is faster at batch 1: 2.4× against LLaDA-Rec and 3.5× against TIGER (below).

The control, however, is a synthetic model: MaLaCo's backbone run in a diffusion pattern but never trained for that role, so it has no accuracy to go with its speed. We use it only to illustrate how the gap scales with batch size, beam width and history length. The better architectural comparison is the KV-cached encoder–decoder below (TIGER), where both systems are trained and run a complete beam search.

New: comparison with a KV-cached encoder–decoder. We timed the TIGER models of Comment 1 with the harness, GPU (A100-SXM4-40GB) and settings of App. F: fp32, CUDA-synchronized calls, beam 50. TIGER's encoder runs once per query and its decoder uses a KV cache. We time it with HuggingFace generate and with a hand-written beam loop that returns identical ranked lists, and report the faster of the two.

ms/query, beam 50batch 1batch 20
TIGER, 4.81M35.31.79
TIGER, 6.19M (= MaLaCo)35.91.79
TIGER, full published pipeline (4 digits)43.12.14
MaLaCo (App. F)9.980.526
  • Speed. MaLaCo is 3.5× faster than TIGER at batch 1 and 3.4× faster at batch 20, and equally so against the 6.19M TIGER. The ratio narrows to about 2× at batch 128, where both systems keep the GPU busy.
  • Batch 1. Against an optimized encoder–decoder, MaLaCo is faster at batch 1 too. TIGER's decoder cross-attends to the encoded history for every hypothesis at every step; MaLaCo's head reads only the single target state.
  • Consistency. Re-timing MaLaCo and LLaDA-Rec in the session that timed TIGER reproduces App. F within 4%.

Memory. At batch 128, TIGER peaks at 5,880 MiB, MaLaCo at 326 MiB and LLaDA-Rec at 19,131 MiB (App. F). At batch 256, LLaDA-Rec runs out of memory on the 40 GB GPU, TIGER uses 11,661 MiB and MaLaCo 599 MiB.

AuthorsOpenReview ↗

W3, necessity of the SID pipeline; W4, reproducibility and citation

W3. Is the pipeline necessary? Scientific (25,848 items), one A100:

StageWhat it buysWall-clock
1–2: balanced residual k-means + last-level Sinkhorn dedupsemantic hierarchy; 0.68% excess collisions remain (Table 4)10 s
3: SID autoencoder refined with REINFORCEalternative semantic candidate SIDs per item83 min
4: margin-ordered catalog assignment0.0039% excess collisions (Table 4)6 s

1. Recommendation quality. Table 5 already compares simpler output vocabularies, changing only the SID table. A direct catalog softmax reaches R@10 0.0412, random collision-free SIDs 0.0426, and RQ-VAE + Sinkhorn on all levels or on the last level only 0.0455 and 0.0481. Ours reaches 0.0499. We now add the comparison within our own hierarchy: the same recommender trained on two simpler tables built from the Stages 1–2 output (Scientific, four seeds).

SID tableDigitsR@10R@50N@10
Stages 1–2 only (some items share a SID)30.0497±.00080.1159±.00150.0260±.0006
Stages 1–2 + TIGER-style appended disambiguation digit40.0499±.00100.1166±.00070.0263±.0005
All four stages (Table 1)30.0499±.00110.1173±.00180.0261±.0008

All three give the same recall within seed variation. Stages 3–4 do not improve recall, and the paper do not suggest that they do. The gain over RQ-VAE-based tables comes from the balanced residual k-means hierarchy of Stages 1–2.

3. Catalog maintenance. New items arrive continually. Refitting the tokenizer can reassign the identifiers of existing items, which invalidates the trained recommender's output vocabulary and every stored SID. Our table avoids this. Insertion reuses Stage 4 only and takes seconds: existing SIDs are reserved, and each new item receives an unclaimed semantic candidate. On Arts, we inserted the newest quarter of the catalog (22,490 items) into a table fitted on the older items only (App. E, Table 6):

  • every existing SID is unchanged, and insertion adds no collision;
  • 78.1% of new items share their first code with their nearest existing neighbour, against 0.83% by chance;
  • after retraining, the extended table reaches R@10 0.04526, against 0.04598 after refitting the whole vocabulary (1.6% relative);
  • a recommender never trained on the new items keeps existing-item R@10 nearly unchanged (0.05788 vs 0.05803).

The appended digit also keeps existing identifiers fixed. The difference lies in the identifiers. With the appended digit, a new item that lands on an occupied SID is told apart only by a non-semantic counter. Every identifier also gets a fourth decoding step, which makes decoding about 30% slower (0.687 vs 0.526 ms/query at batch 20). Stages 3–4 therefore trade 83 offline minutes per catalog for fully semantic three-digit identifiers and faster decoding. The revision will present them as an option on that basis and adds the table above.

W4a. Code. As stated in App. A, institutional policy prevents us from releasing code, checkpoints or intermediate artifacts, and we acknowledge this as a limitation

W4b. LIGER. Thank you. The revision will cite the correct work

Thank you for the authors’ response. It has resolved some of my concerns, and I believe it would be more appropriate to raise the score to 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.