acceptodds
Under review as a conference paper at ICLR 2027

COMPOTE: Collapse-Resistant Multiple-Hypothesis Learning For Sequential Recommendation

Abstract

Next-item prediction is inherently ambiguous: a single interaction history often admits multiple plausible continuations across distinct catalogue regions. Stan- dard sequential recommenders compress these alternatives into a single prediction, conflating user intents. We introduce COllapse-resistant Multiple hyPOThesis lEarning (COMPOTE), a multi-hypothesis transformer trained with an annealed objective that transitions from uniform supervision to a winner-takes-all assign- ment. This mechanism forces hypotheses to specialize into distinct catalogue re- gions, while inference relies on an efficient, normalizer-free max-similarity rule. Theoretically, we show that in the population case, the optima solutionspartition the conditional next-item distribution according to a max-entropy quantization principle, while pooled scores preserve the Bayes-optimal ranking. We also prove a generalization bound scaling logarithmically with the number of hypotheses. Across three Amazon Reviews 2023 benchmarks, COMPOTE outperforms strong ID-based, generative, and multi-interest baselines using item IDs alone. Matched- capacity ablations confirm that performance gains stem directly from the annealed objective, preventing the head collapse observed in conventional multi-head archi- tectures, rather than added capacity, ensembling, or routing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
sequential recommendation, multi-hypothesis prediction, multiple-choice learning, winner-takes-all learning

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

The paper proposes COMPOTE, a sequential recommender that outputs M hypotheses per user history. A shared causal transformer feeds M heads, each producing a cosine/temperature softmax over the catalogue. Training uses the annealed loss, which interpolates between uniform supervision (high τ) and winner-takes-all (τ→0). An auxiliary confidence head is trained with a BCE loss. At inference, items are ranked by max-pooling over heads. They show this can be done on raw cosine similarities, without the per-head softmax normalizer. Empirically, on three Amazon-2023 subsets, COMPOTE is best in 11 of 12 columns against ID-based baselines. Ablation studies were performed to argue the gains come from the annealed objective. Diagnostics are performed to argue that heads specialize into a near-static tiling of the catalogue.

Critical strengths
  1. Controlled evidence that the multi-head parameterization matters, plus a clear collapse diagnosis. Table 6 shows EnsCE and Router heads collapse completely (S=1.00), while COMPOTE’s heads do not (S≈0.6). Table 7 shows the oracle gap is zero for collapsed models and large for COMPOTE. This is a convincing demonstration that equal-weight and routed multi-head training collapses in this setting, and that the COMPOTE objective yields distinct heads. It is the paper’s most valuable empirical contribution.

  2. A practical inference rule, honestly validated, with transparent limitations. Dropping the per-head softmax normalizer costs nothing (Table 2). Appendix C quantifies why: the cross-head spread of log Z_m is small relative to the cosine gaps that matter. Appendix M openly reports that the learned tiling is nearly static, which the theory does not predict. This candor makes the empirical story easier to trust and gives the community a concrete open question.

Critical weaknesses
  1. A large share of the improvement over SASRec comes from the backbone and loss recipe, not from multiple hypotheses. For example, Game N@5 goes 3.31 (SASRec) → 3.78 (MCT-SASRec) → 4.01 (COMPOTE). Multi-interest baselines trained by the authors (e.g., MIND 3.22 H@5 on Game) land far below SASRec, which is atypical and suggests they were under-tuned or trained with a weaker (sampled-softmax) objective than COMPOTE’s full softmax.

2.The deployment claims (sublinear ANN, “key deployment result”) are never tested. Training requires a full-catalogue softmax per head, and no cost or latency numbers are given.

Questions for authors
  1. Key experimental details are missing from the main text: the annealing schedule (initial τ, decay, final τ), β, architecture size, tuning budget, seeds, and whether model selection is on validation or test.
  2. The proof of Thm 7 cites “a direct application of Theorem 7” (circular).
  3. S(x) in Sec. 5.1 sums over j,k=1..M including j=k, while normalizing by M(M−1).
  4. Appendix J is an empty heading.
1 comment

Dear Reviewer bFPG, thank you for your thoughtful and constructive feedback on our paper.

W1: We would like to clarify that we do not claim that the entire improvement over SASRec comes from the multiple-hypothesis formulation: the backbone and training recipe also contribute to performance. Our claim is that COMPOTE provides additional gains beyond these components.

Following the reviewer's suggestion, we conducted additional experiments on Yelp, comparing COMPOTE with SASRec, TIGER, and several multi-hypothesis ablations (MCT-SASRec, MCT-Router, and MCT-EnsCe). COMPOTE achieves the best results across all four metrics, reaching 3.40% Hit@5, 4.80% Hit@10, 2.47% NDCG@5, and 2.92% NDCG@10. Relative to the strongest non-COMPOTE baseline for each metric, these correspond to improvements of 34.4%, 17.9%, 25.4%, and 35.8%, respectively.

ModelHit@5Hit@10NDCG@5NDCG@10
SASRec1.83%2.96%1.16%1.52%
TIGER2.53%4.07%1.64%2.13%
MCT-SASRec1.53%2.62%1.36%1.60%
MCT-Router1.94%2.73%1.43%1.67%
MCT-EnsCe2.39%3.46%1.97%2.15%
COMPOTE3.40%4.80%2.47%**2.92%
Improvement+34.4%+17.9%+25.4%+35.8%

These additional results support our central claim: while the backbone and training recipe matter, COMPOTE's multiple-hypothesis learning and retrieval approach provides further improvements over the considered alternatives. We will include these results in the revised paper at appendix Q.

W1 bis: We have checked our implementations, the runs of MIND and we can confirm that there is no implementation errors and that they have been trained using full softmax and for a sufficiently large number of iterations.

We also want to note that there exists other works in the literature that report worse performance of MIND and COMIREC compared to SASRec, albeit in a bit different context than ours.

[1] Nguyen et al, BACH: A Bayesian Admixture of Contrastive Heads for Multi-Interest Two-Tower Retrieval, https://arxiv.org/pdf/2607, 2026.

[2] Zhang et al, Controllable Generative Recommendation via Guided Token Refinement, https://arxiv.org/pdf/2610.28642, 2026.

[3] Si et al, Generative Retrieval with Semantic Tree-Structured Identifiers and Contrastive Learning, https://arxiv.org/pdf/2309.13375, 2023

[4] Zhang et al Divide and Conquer: Towards Better Embedding-base, Retrieval for Recommender Systems from a Multi-task Perspective, https://arxiv.org/pdf/2302.02657, 2023

W2: COMPOTE does require a full-catalogue softmax for each hypothesis during training. At inference, however, as described in Section 4.1 (Fast Retrieval for COMPOTE), our max-similarity scoring rule avoids the full-catalogue softmax and enables retrieval through MM independent approximate nearest-neighbor (ANN) searches, one per hypothesis. These searches can be executed in parallel, yielding sublinear query time in the catalogue size under standard ANN indexing, although the total computational work still scales linearly with the number of hypotheses MM. Thus, our sublinearity claim concerns inference latency, not total FLOPs or training complexity.

To make it clearer, we have updated section 4.1.

Q1: We thank the reviewer for noting that these details should be more accessible. We have added Appendix H to provide a consolidated description of the model architecture and training hyperparameters. In particular, we report the embedding dimension d=128d=128, the annealing schedule \tau$: 1.0 $\rightarrow$ 0$, with linear decay), the learning-rate schedule ($10^{-3}$), with linear decay), weight decay ($10^{-3}$), catalogue temperature \beta=0.07=0.07, and score-loss weight $$\alpha=0.1=0.1.

We also clarify in appendix G the experimental protocol and validation-based model selection, to make the setup fully reproducible.

Q2: Thanks for pointing this out. The correct reference was "Th. 1 from Ruppert et al."". The proof has been edited.

Q3: We have corrected the typo and removed the case j=k.

Q4 The section contains the two tables about the head collapse (Tables 6 and 7 in the submitted version). We have now added some context to this section, as well as some insights from the tables.

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

This paper studies ambiguity in next-item prediction for sequential recommendation. The authors propose COMPOTE, which uses a shared Transformer with multiple prediction heads and an annealed objective that approaches winner-takes-all training, while the items are ranked by their maximum cosine similarity across heads. The paper provides theoretical analyses of head specialization, ranking, and generalization. Experiments on three Amazon datasets show ranking improvements and complementary predictions across heads.

Critical strengths
  • The method learns complementary recommendation heads without an explicit diversity penalty.
  • The experiments show ranking improvements over conventional multi-head training.
Critical weaknesses
  • The generalization bound in Theorem 4 does not hold under the stated assumptions.
  • The motivation and contribution beyond related preference-decomposition methods are insufficiently established.
Questions for authors

I have several questions that need clarification:

  • Motivation. The claim that a single query vector mixes different interest modes is not directly validated. The experiments do not establish this failure of the single-head model. Please analyze whether single-head embeddings concentrate on a dominant interest or mix distinct interests, and whether COMPOTE recovers other relevant interests.

  • Closely-Related Work. Representing heterogeneous preferences with multiple components is closely related to reward factorization [R1], latent reward mixtures [R2], and specialized expert models [R3]. Please discuss the differences in supervision, specialization, and aggregation, and clarify the contribution beyond aMCL.

  • Generalization Bound. Theorem 4 does not hold under its stated assumptions. In the second term in line 967, Appendix D.2, the winner mi∗m_i^* can differ across samples, whereas max⁡m∑i\max_m\sum_i selects one head for the entire sum. The proof also drops the supremum over head tuples: the residual vectors vary with the chosen heads and cannot be treated as MM fixed candidates. Consequently, the subsequent finite-class bound does not establish the claimed logarithmic dependence on MM. Please provide a valid bound and revise the associated claims.

Reference:

  • [R1] Language Model Personalization via Reward Factorization. COLM 2025.
  • [R2] MiCRo: Mixture Modeling and Context-aware Routing for Personalized Preference Learning. EMNLP 2025.
  • [R3] Sparse Mixture-of-Experts Reward Models Learn Interpretable and Specialized Experts for Personalized Preference Modeling. arXiv 2026.
2 comments

Dear Reviewer parV, thank you for your thoughtful and constructive feedback on our paper. We have carefully addressed your questions and suggestions below.

W1 & Q3: We thank the reviewer for pointing out that our previous application of Massart's finite class lemma was incorrect. We agree with this assessment and have revised our bound accordingly. Specifically, we decoupled the complexity of the shared encoder from that of the MM heads. Using classical arguments, we now provide a rigorous generalization bound that scales with M/n\sqrt{M/n}. While this bound is not as sharp as our previous erroneous result, it still maintains a nice sublinear scaling with respect to MM, preserving the core claims of our analysis.

The paper has been modified accordingly ( Section 3 and Appendix D.2)

W2 Thanks for pointing this out. We have discussed the difference between those methods and our approach in the introduction. In addition, we have also shown, in Table 10 and Appendix P, that linearly combining specialized representations into a single query vector (as in standard Annealed MCL) degrades performance compared to our method.

Q1 We thank the reviewer for highlighting the need to validate this motivation. For this purpose, we conducted an additional analysis on Amazon Scientific (54 subcategories), comparing COMPOTE and SASRec across different target types. We distinguish:

  • Single-category histories: all historical items belong to the same subcategory.

  • Multi-category histories:

    • Dominant-category targets: the target belongs to the most represented historical category.

    • Secondary-category targets: the target belongs to a non-dominant category already present in the history.

    • Novel-category targets: the target category is absent from the history.

SASRec performs slightly better on single-category and dominant-category targets (11.68% vs. 11.31% and 7.98% vs. 7.85% Recall@10, respectively). However, COMPOTE improves Recall@10 by 13.1% on secondary-category targets (3.45% vs. 3.05%) and 14.0% on novel-category targets (1.14% vs. 1.00%). Consistently, only 33.2% of COMPOTE's top-10 recommendations belong to the dominant historical category, compared with 42.3% for SASRec.

This suggests that, while single-head prediction favors dominant-interest continuations, COMPOTE better recovers alternative plausible continuations, improving coverage of secondary and previously unseen categories.

We added those results to the manuscript in Appendix O.

Q2: We thank the reviewer for pointing us to these related works. While R1–R3 use multiple components to model heterogeneity in human preference modeling, COMPOTE addresses ambiguity in next-item prediction from a single interaction history.

Supervision, specialization, and aggregation. R1 learns shared reward functions combined through user-specific coefficients, while R2 and R3 learn preference components from pairwise comparisons, using context-dependent routing (and sparsity/diversity regularization in R3). In contrast, COMPOTE uses only observed next-item interactions and specializes its predictive hypotheses through annealed competitive assignment. At inference, it uses parameter-free max-similarity aggregation, allowing items matched by different hypotheses to coexist in the final ranking, without learned routing or mixture weights.

Contribution beyond aMCL. While COMPOTE builds on the aMCL principle of learning multiple hypotheses, it introduces a recommendation-specific, parameter-free max-similarity aggregation that allows different hypotheses to contribute to the same recommendation list. Its empirical importance is reflected in the performance gap with a direct aMCL application to our backbone on Amazon Games (NDCG@5: 3.01 vs. 4.01; HR@5: 4.50 vs. 6.06).

Beyond this adaptation, we provide theoretical results characterizing hypothesis specialization, establish that the aggregated scores preserve Bayes-optimal ranking, and derive a generalization bound with sublinear dependence on the number of hypotheses.

We have clarified these distinctions and contributions in the introduction of the revised paper.

Thank you for the detailed responses and additional experiments. I appreciate the effort put into the revision. However, my main concerns remain insufficiently addressed.

  • Motivation. The category-based subgroup analysis alone does not sufficiently validate the claim that a single query vector conflates distinct interest modes. Further representation-level evidence, including spectral analysis of the learned embeddings, would help substantiate this motivation. Moreover, the reported results suggest that COMPOTE performs slightly worse when the target belongs to the dominant historical category. This effect warrants deeper analysis to explain why it occurs and how it relates to the proposed specialization mechanism.
  • Novelty. I still consider the methodological novelty relative to aMCL and ComiRec [R4] to be limited. While the rebuttal clarifies the intended distinctions, it does not sufficiently establish the incremental contribution beyond aMCL's annealed multiple-choice learning and ComiRec's max-over-interest aggregation.

Given these remaining concerns, I maintain my original rating.

References:

  • [R4] Controllable multi-interest framework for recommendation. KDD 2020.
Reviewer_9qj5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COMPOTE, a multiple-hypothesis learning framework for sequential recommendation. The method employs a shared transformer encoder with multiple prediction heads and introduces an annealed winner-takes-all objective to encourage hypothesis specialization and mitigate head collapse. At inference time, the model aggregates predictions using a max-similarity rule that avoids computing catalogue-wide softmax normalizers. The paper provides theoretical analyses of the generalization behavior of the WTA objective, the structure of population-optimal hypotheses, and the ranking properties of max-pooling aggregation. Experiments on three Amazon Reviews 2023 datasets show improvements over several sequential recommendation and multi-interest baselines. Additional ablations investigate the contributions of the training objective, hypothesis diversity, and aggregation strategy.

Critical strengths

1.The paper proposes an annealed WTA objective to encourage hypothesis specialization without explicit diversity regularization, offering a simple approach to modeling multimodal user preferences. 2.COMPOTE achieves competitive performance across three Amazon Reviews 2023 datasets. The capacity-matched ablations provide useful evidence for the effectiveness of the proposed training objective.

Critical weaknesses

1.The paper reuses the symbols α, S, and g to denote different quantities across sections, leading to ambiguity and making the mathematical formulations and proofs harder to follow. 2.Theorem 7 extends the Voronoi characterization of Rupprecht et al. (2017) to discrete categorical distributions under a WTA log-loss. However, the original result relies on regularity assumptions that do not directly carry over to the discrete setting. The current theorem does not address ties, zero-mass cells, or degenerate optima, and therefore requires additional assumptions or a more precise formulation. 3.The proof in Appendix D.2 appears to incorrectly apply Massart's finite-class lemma to sample-dependent head assignments, treating them as a selection among only M fixed candidates. This overlooks the combinatorial complexity of pointwise head selection and leaves the claimed O(logM/n) dependence unjustified. The theorem may require additional structural assumptions or a substantially revised bound.

Questions for authors

1.Could the authors clarify how the proof accounts for sample-dependent winning-head assignments when applying Massart's finite-class lemma? What justifies the claimed logarithmic dependence on the number of hypotheses? 2.Does the partition characterization require additional assumptions regarding zero-probability items, empty cells, or ties between optimal heads? Could the authors clarify how these degenerate cases are handled in the proof?

1 comment

Dear Reviewer 9qj5, Thank you for your thoughtful and constructive feedback on our paper. We have carefully addressed your questions and suggestions below.

W1: We thank you for highlighting these notation collisions across sections. We have revised the manuscript to ensure every symbol has a unique definition:

  1. α\alpha: The simplex vector in Corollary 3 has been changed to w∈ΔM−1\mathbf{w} \in \Delta^{M-1} (with components wmw_m). The symbol α\alpha is now strictly reserved for the auxiliary score-loss weight Lτ,α\mathcal{L}_{\tau,\alpha}.
  2. SS: The pairwise head similarity metric in Section 5.1 has been renamed to Sim(x)\mathrm{Sim}(\mathbf{x}) and Sim\mathrm{Sim}, reserving SS exclusively for the training dataset/sample in statistical learning statements.
  3. gg: The 1D scalar function in the proof of Proposition 5 (Appendix B.1) has been renamed to ξ(t)\xi(t), reserving gg uniquely for the recommender ranking model (g=argsort∘ω∘fg = \mathrm{argsort} \circ \omega \circ f).

W2 & Q2: We now provide the detailed proof of this Theorem in Appendix L.1, largely inspired by the line of proof of Rupprecht et al. (2017). We reformulated the statement to

  1. fix the case when supp(q⋆)⊊Y\mathrm{supp}(q^\star) \subsetneq \mathcal{Y} (that was indeed not covered, thank you for that),
  2. make it clear that ties, zero-mass cells cannot happen at optima (without additional assumptions).

Some intuition about this second point.

  • Intuitively, zero-mass cells do not exist because of the entropy-maximizing property of Ym\mathcal{Y}_m.
  • Ties between hypotheses are not optimal. Even though it is a consequence of the theorem's proof, the proof brings no intuition as to why. Here is a semi-formal intuition: assume p⋆p^\star is such that for some mm, m′m' and yy, pm,y⋆=pm′,y⋆p^\star_{m,y} = p^\star_{m',y} (a tie). Let p~\tilde{p} be equal to p⋆p^\star almost everywhere, just zeroing p~m,y\tilde{p}_{m,y} and renormalizing the rest of the row (p~m,⋅\tilde{p}_{m,\cdot}). The contribution of sample yy to the WTA loss is unchanged (as p~m’,y=pm’,y⋆\tilde{p}_{m’,y} = p^\star_{m’,y}). However, for all y’∈Ym∖yy’\in\mathcal{Y}_{m}\setminus {y}, p~m,y’>pm’,y’⋆\tilde{p}_{m,y’} > p^\star_{m’,y’} (because of the renormalization). Thus their contribution to the WTA loss, that is −log⁡(p~m,y’)-\log(\tilde{p}_{m,y’}), decreased (compared to p⋆p^\star) and thus p~\tilde{p} has a lower WTA loss than p⋆p^\star (contradiction).

W3&Q1: We thank the reviewer for pointing out that our previous application of Massart's finite class lemma was unjustified. We agree with this assessment and have revised our bound accordingly. Specifically, we decoupled the complexity of the shared encoder from that of the MM heads. Using classical arguments, we now provide a rigorous generalization bound that scales with M/n\sqrt{M/n} times the complexity of an individual head, but additive w.r.t. the complexity of the shared encoder. While this bound is not as sharp as our previous erroneous result, it still maintains a nice sublinear scaling with respect to MM which is additive w.r.t. the main complexity (shared encoder), preserving the core claims of our analysis.

The paper has been modified accordingly ( Section 3 and Appendix D.2)

Comments on the paper

2 comments

Dear Area Chair and Reviewers,

Thank you for your time, thoughtful reviews, and constructive feedback. We have uploaded a revised version of our manuscript, with all modifications and new additions highlighted in blue for your convenience.

In addition to addressing your individual comments, we conducted new experiments on the Yelp dataset to demonstrate the effectiveness and generalizability of COMPOTE beyond the Amazon Reviews 2023 benchmarks.

As shown in the table below, COMPOTE consistently outperforms all baselines on Yelp across every evaluation metric, achieving relative improvements ranging from +17.9% to +35.8% over the strongest baseline:

ModelHit@5Hit@10NDCG@5NDCG@10
SASRec1.83%2.96%1.16%1.52%
TIGER2.53%4.07%1.64%2.13%
MCT-SASRec1.53%2.62%1.36%1.60%
MCT-Router1.94%2.73%1.43%1.67%
MCT-EnsCe2.39%3.46%1.97%2.15%
COMPOTE3.40%4.80%2.47%2.92%
Improvement+34.4%+17.9%+25.4%+35.8%

These results have been added to Appendix Q of the revised manuscript. Below, we address each reviewer's specific questions and comments individually.

As the rebuttal discussion period approaches its conclusion, we would like to kindly follow up on our response and the additional revisions and experiments we provided.

In particular, we further clarified and strengthened the motivation behind our approach, corrected the generalization bound, revised the proof of Theorem 7 (Theorem 6 in the revised version of the paper), and conducted additional experiments on the Yelp dataset to further demonstrate the robustness of our methods.

We would be happy to address any remaining concerns or provide further clarifications if needed. If our responses and the additional theoretical and experimental evidence have satisfactorily addressed your concerns, we would sincerely appreciate your reconsideration of the rating.

Thank you again for your time, thoughtful feedback, and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.