acceptodds
Under review as a conference paper at ICLR 2027

LOTUS-UCC: UNITARY SEMANTIC COCYCLES FOR SINGLE-MODEL SEQUENTIAL RECOMMENDATION

Abstract

LLM-enhanced recommendation has improved semantic representation, but largely leaves semantic evolution unspecified: knowing what an item means does not explain how an interaction moves intent or when that motion changes a ranking decision. We introduce LOTUS-UCC (Language-Oriented Transport with Unitary Semantic Cocycles), a single-model semantics-to-decision dynamical system in which language defines the coordinate space of user intent. Causal spherical forecasting and persistent particle transport maintain multiple latent semantic states; a unitary cocycle carries the resulting representation without distorting its geometry; and realification produces one dense full-catalog score. Our analysis follows perturbations through the complete chain from semantic history and persistent state to particles, semantic endpoints, catalog scores, and Top-K decisions. It also quantifies the curvature error of Euclidean updates and proves that CNDB upper-controls cutoff failure risk. Under chronological leave-one-out and exhaustive ranking, LOTUS-UCC establishes new best results on 19 of 20 metrics across Beauty, Sports, Toys, Yelp, and Online Retail, with a 25.08% average relative gain over the strongest published result in each setting. Structural ablations provide evidence for the geometric and decision-aware mechanisms, and the proposed dynamics generalize consistently across four language representation spaces.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Sequential Recommendation, LLM-enhanced Recommendation, Semantic Transport, Unitary Semantic Cocycle, Spherical Geometry, Particle Transport, Full-Catalog Ranking, Top-K Recommendation

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

This paper proposes LOTUS-UCC, a single-model framework for language-enhanced sequential recommendation. It aims to model how user intent evolves in a language-defined semantic space and how this evolution affects ranking decisions. The method combines causal spherical forecasting, persistent particle transport, a unitary cocycle, and full-catalog scoring, with CNDB for cutoff-aware ranking. Experiments on five benchmarks show that LOTUS-UCC achieves the best results on 19 of 20 metrics, supported by structural ablations and cross-LLM evaluations.

Critical strengths
  1. The proposed framework combines semantic forecasting, persistent particle transport, and unitary state evolution, with theoretical analysis connecting semantic changes to ranking decisions.
  2. The experiments cover five benchmarks with full-catalog evaluation, structural ablations, and cross-LLM tests, showing strong recommendation performance.
Critical weaknesses
  1. Insufficient motivation for explicit semantic evolution. The paper argues that existing language-enhanced recommendation methods do not explicitly model how user intent evolves in semantic space. However, existing semantic-enhanced sequential recommendation and generative recommendation methods already learn history-dependent transitions using semantic or collaborative representations. It is unclear why explicitly modeling semantic movement is necessary compared with directly learning these transitions through next-item prediction. Although the ablation experiments show improvements from semantic forecasting, they do not establish its advantage over simpler methods that directly predict future semantic representations.
  2. Insufficient motivation for the proposed components and their integration. The paper combines several complex components, including spherical forecasting, persistent particle transport, Householder transformations, and unitary cocycles. However, the motivation for these specific designs is not sufficiently explained. For example, it is unclear why spherical geometry is necessary for modeling semantic transitions, why optimal transport is suitable for maintaining multiple user intents, and why unitary transformations are needed for recommendation. Although the paper provides theoretical properties and ablation results for these components, these results do not fully explain why these designs are preferable to simpler alternatives.
Questions for authors
  1. Why is explicit semantic transport necessary when existing semantic-enhanced sequential and generative recommendation methods already learn history-dependent transitions?
  2. What specific challenges motivate the use of spherical forecasting, persistent particle transport, and unitary cocycles, and why do these components need to be combined in a single framework?
Reviewer_K5UnOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a language-enhanced sequential recommendation framework that models user intent evolution through geometric and dynamical transformations in a language-defined semantic space. The proposed architecture integrates several components: causal spherical semantic forecasting, persistent particle transport through entropy-regularized optimal transport, semantic-conditioned Householder transformations, a unitary cocycle for complex-valued representation evolution, and realified full-catalog scoring. The authors further introduce a cutoff-aware ranking objective, CNDB, to optimize recommendation decisions at specific ranking cutoffs. The paper provides theoretical analyses concerning geometric consistency, perturbation propagation, compositional stability, and ranking cutoff guarantees. Experiments on five sequential recommendation benchmarks report state-of-the-art results on 19 of 20 evaluated metrics. Additional analyses investigate component effectiveness, alternative geometric operations, language encoder generalization, statistical significance, and computational efficiency.

Critical strengths
  • The paper addresses an interesting research problem by modeling how item semantics evolve over time, rather than relying on static language representations. Its combination of spherical geometry, multi-hypothesis particle transport, and unitary transformations offers a different perspective on semantic modeling. The proposed scoring framework, which connects continuous semantic evolution with discrete recommendation decisions, is also an interesting aspect of the work.

  • The authors evaluate LOTUS-UCC on five recommendation benchmarks and report consistent improvements over a diverse set of sequential recommendation baselines. The experiments follow a chronological leave-one-out protocol with full-catalog ranking, which avoids the potential biases introduced by sampled-negative evaluation. Notably, the performance gains are particularly large on Yelp and Online Retail. The paper also includes additional analyses of architectural components, geometric operators, language embedding models, statistical significance, and computational efficiency, providing a fairly comprehensive empirical evaluation.

Critical weaknesses
  • The methodology and theoretical analysis do not convincingly support the claimed semantic forecasting and compositional dynamics. First, it is unclear how the causal spherical forecast in Eq. (4) contributes to the subsequent model operations. Section 3.2 constructs a semantic trajectory from historical tangent velocities and uses the spherical exponential map to predict a semantic endpoint. However, the particle transport mechanism in Section 3.3 instead builds its anchors from the current semantic state sts_t, the Euclidean displacement dt=st−st−1d_t=s_t-s_{t-1}, and the behavioral hidden state, without an explicit connection to the forecast endpoint. A similar concern applies to the proposed unitary cocycle dynamics. Section 3.5 introduces a sequence of unitary operators and their compositional path operator, while Theorem 1 shows that their products preserve norms and satisfy the cocycle composition property. Although these properties are mathematically valid, they follow naturally from the definition of unitary matrices. More importantly, the model does not appear to update its representations through the accumulated cocycle operator Ut←t−1U_{t\leftarrow t-1}, and it is unclear whether the path composition in Eq. (24) is actually used during scoring. As a result, the connection between the theoretical formulation and the implemented model remains insufficiently explained.

  • There are notable inconsistencies in the reported results, and the ablation studies provide limited evidence for the necessity of the proposed geometric components. Table 1 reports the three-seed average performance of LOTUS-UCC, which agrees with the results in Appendix Table 5. However, the Full LOTUS-UCC results in Table 3 are consistently higher across all 12 dataset-metric combinations on the Amazon benchmarks. This discrepancy raises concerns about the comparability of the main experiments and ablation results. Beyond this issue, the ablations do not fully justify the complexity of the proposed framework, which incorporates spherical forecasting, optimal transport, particle persistence, Householder transformations, unitary phase evolution, and cutoff-aware optimization. While Table 3 shows that removing individual components generally leads to lower performance, the differences are often relatively small. It therefore remains unclear whether these geometric mechanisms are essential to the observed improvements or whether a simpler architecture could achieve comparable results.

Questions for authors
  • Could the authors clarify how the spherical forecast in Eq. (4) is used in the subsequent computations (Eqs. (5)–(35))? In particular, Eq. (14) seems to derive the final semantic endpoint directly from the particle representations, without incorporating the forecast from Eq. (4).

  • Could the authors explain why Full LOTUS-UCC reports different performance in Tables 1, 3, and 4, despite apparently using similar experimental settings?

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

This paper introduces LOTUS-UCC (Language-Oriented Transport with Unitary Semantic Cocycles), a unified framework for sequential recommendation that explicitly models the evolution of user intent in a language-defined semantic space. The authors argue that existing language-enhanced recommendation methods primarily leverage static semantic representations without explicitly modeling how user intent evolves across successive interactions. To address this limitation, LOTUS-UCC combines several mechanisms:

  • Causal spherical semantic forecasting: Item embeddings from a pretrained language model define a spherical semantic space, where historical interactions determine geodesic directions for forecasting future intent.
  • Persistent particle transport: Multiple latent semantic states are maintained and updated using entropic optimal transport, allowing the model to represent competing intent hypotheses.
  • Householder transformations and unitary semantic cocycles: Orthogonal and unitary transformations are used to preserve representation geometry while modeling event-conditioned semantic evolution.
  • Full-catalog ranking with CNDB: A cutoff-aware ranking objective supplements full-catalog likelihood to improve Top-K recommendation performance. The paper further develops theoretical results concerning geometric consistency, perturbation propagation, compositional stability, and ranking-boundary behavior. Experiments are conducted on five sequential recommendation datasets: Amazon Beauty, Sports, Toys, Yelp, and Online Retail. Under chronological leave-one-out evaluation and full-catalog ranking, the authors report improvements over existing baselines on 19 of 20 HR and NDCG metrics, with an average relative improvement of 25.08%. Additional experiments investigate component ablations, alternative geometric operations, different language encoders, parameter sensitivity, and computational efficiency.
Critical strengths

S1. Novel geometric perspective on semantic evolution in sequential recommendation. The paper proposes a conceptually interesting formulation of sequential recommendation as a dynamical process in a language-defined semantic space. Rather than treating language embeddings as static auxiliary features, LOTUS-UCC explicitly models semantic transitions and uses them to influence subsequent recommendation decisions. This perspective is potentially valuable because it connects semantic representation learning with temporal preference modeling, two aspects that are often addressed separately. S2. Comprehensive integration of geometric and dynamical modeling mechanisms. The proposed framework integrates spherical forecasting, persistent particle transport, Householder transformations, and unitary dynamics into a single recommendation model. Although these mathematical operations are individually established, their integration into a unified semantics-to-decision pipeline is technically interesting. The use of unitary transformations provides exact norm preservation, while the particle transport mechanism offers a structured approach to maintaining multiple latent preference hypotheses.

Critical weaknesses

W1. Theoretical stability guarantees are conditional and their practical relevance is not sufficiently demonstrated. A central contribution of the paper is the theoretical connection between geometric representation evolution and stable recommendation decisions. However, the strongest stability results depend on assumptions whose satisfaction is not established empirically. For example, the Top-K invariance condition depends on the score margin exceeding a perturbation-dependent threshold. Without empirical measurements of these margins and perturbation bounds, it is difficult to determine whether the theoretical guarantees are informative in practical recommendation settings. The mathematical analysis therefore establishes conditional properties of the framework, but does not yet demonstrate that these properties explain the observed recommendation improvements.

Questions for authors

Q1. Experimental consistency Could the authors explain why the full LOTUS-UCC results in Table 3 differ from those in Table 1 and Appendix Table 5? Do these results correspond to different seeds, checkpoints, hyperparameter configurations, or evaluation procedures? If different experimental settings were used, could the authors provide a consistent comparison and report the variance of the ablation results? Q2. Practical validity of the theoretical stability guarantees Does the trained model satisfy the contraction condition L_P<1 required by Corollary 1? Q3. Contribution of individual components The ablation results suggest that several components provide relatively small improvements. Could the authors report three-seed means and standard deviations, or paired statistical tests, for all component ablations rather than primarily the UCC removal? Q4. Computational scalability How does training and inference complexity scale with catalog size, sequence length, number of particles, and Sinkhorn iterations? Can the method efficiently support catalogs containing millions of items, especially given its use of full-catalog likelihood and hard-negative selection?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.