acceptodds
Under review as a conference paper at ICLR 2027

ATLASREC: LANGUAGE-GROUNDED MULTI-HORIZON SUCCESSOR DYNAMICS FOR SEQUENTIAL RECOMMENDATION

Abstract

Language-enhanced sequential recommenders improve how items and histories are represented, but next-item prediction also requires direction: knowing what a sequence means is not the same as knowing where it is going. We formulate this task as a state–dynamics–decision problem and introduce ATLASREC. Frozen language representations estimate a recency-weighted semantic state; training interactions estimate sparse, popularity-relative successor dynamics; and a pretrained recommender provides the behavioral decision reference. An event \(r\) positions back queries the \((r + 1)\)-step operator, so every source is interpreted at its exact distance from the target. We show that repeating a one-step operator assumes horizon-invariant dynamics and otherwise incurs a discounted conditional-KL mismatch. A minimum-information closure incorporates the two evidence fields while deviating minimally from the behavioral prior, and finite-sample successor uncertainty yields an explicit robust ranking condition. On three same-setting Amazon benchmarks, ATLASREC is best in 11 of 12 comparisons with prior methods. Against its exact anchor, it improves all 21 full-catalog metrics, with relative gains up to 18.64%. Component, horizon, state, popularity, encoder-transfer, and efficiency results validate the complete formulation.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Sequential Recommendation, Language-Enhanced Recommendation, Multi-Horizon Dynamics, Semantic Representation, Large Language Models, Recommender Systems

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hvJNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ATLASRec, a language-grounded multi-horizon successor dynamics framework for sequential recommendation. The method combines three sources of information: a recency-weighted semantic state derived from frozen language embeddings, sparse horizon-specific successor statistics estimated from training interactions, and a pretrained MPT recommender serving as the behavioral prior. The main methodological idea is to associate each historical interaction with a successor operator corresponding to its exact temporal distance from the prediction target. The resulting semantic and successor scores are integrated with the behavioral prior through additive logit correction. The paper also presents theoretical analyses based on conditional KL divergence, minimum-information updating, and finite-sample uncertainty. Experiments on five datasets report consistent improvements over the MPT anchor, with competitive results against existing sequential recommendation methods on three Amazon benchmarks.

Critical strengths
  1. Clear motivation and interpretable framework. The separation of semantic compatibility, temporal successor information, and behavioral prior is intuitive. In particular, the proposed target-distance alignment offers a reasonable way to incorporate position-dependent transition statistics into sequential recommendation.
  2. Consistent improvements over the matched backbone. Table 2 reports improvements on all 21 evaluated metrics across the three Amazon datasets. The use of a matched MPT anchor and full-catalog evaluation makes these improvements more informative than comparisons using inconsistent evaluation protocols. The additional encoder experiments also suggest that the approach is not restricted to a single language embedding model.
  3. Lightweight and modular implementation. The proposed evidence layer does not require additional end-to-end training of the backbone. The sparse successor representation and cached language embeddings make the approach computationally attractive. The reported inference overhead is relatively small.
Critical weaknesses
  1. Limited methodological novelty beyond combining existing mechanisms. The core components are relatively straightforward: the integration of semantic states and successor operators is logical; however, the technical distinction between this approach and existing methods—such as "feature augmentation" and "transition-based recommendation"—is not clearly defined. Furthermore, the incremental contribution of "horizon-aligned successor retrieval" compared to "one-step control" is relatively modest, particularly on the "Beauty" and "Toys" datasets (see Table 4).
  2. The theoretical analysis does not fully justify the central methodological claim. Theorem 1 characterizes the discrepancy between the "successor distribution for a specific horizon" and the "direct reuse of a single-step distribution." However, this comparison should be distinguished from the process of "multi-step propagation via a single-step transition operator." The analysis—grounded in conditional KL divergence—does not demonstrate that estimating a "horizon-specific operator" is superior to methods that appropriately combine single-step dynamics. Furthermore, Theorem 2 is derived from a standard variational characterization of exponentially tilted distributions, while the conclusions regarding ranking stability rely primarily on bounded score perturbations; consequently, the work offers limited novelty and has a marginal direct impact on recommendation quality.
  3. Important ambiguities in the ablation experiments. An unexplained correspondence exists between different ablation configurations. For the "Beauty" and "Sports" datasets, the results for "One-step history" in Table 4 are identical to those for "Last item" in Tables 5 and 11 across all five reported ranking metrics. The paper should clarify why these results coincide and confirm whether the experiments involved modifications solely to the intended components.
  4. Incomplete experimental evidence for generalization and individual design choices. Table 3 presents performance results on the Yelp and OnlineRetail datasets; however, for these two datasets, no comparisons using the “matched-anchor” configuration—the primary evaluation setup—are provided. Although Appendix C.2 includes additional comparisons for OnlineRetail, these rely on a separate, fixed “encoder-transfer” configuration. Consequently, the evidence supporting the claim that performance improvements can be achieved across different recommendation domains remains incomplete.
  5. Limited practical validation of the theoretical uncertainty guarantees. The finite-sample concentration analysis in Appendix A relies on conditional sampling assumptions, positive probability floors, and an oracle defined over the empirically selected sparse support. The paper does not evaluate how frequently these assumptions are satisfied or whether the resulting bounds are informative on the reported sparse datasets. Therefore, the theoretical guarantees do not yet provide convincing practical evidence for the reliability of the estimated successor dynamics.
Questions for authors

1.Horizon alignment: How does ATLASRec compare against multi-step propagation obtained from powers of the estimated one-step transition operator, or against a horizon-agnostic aggregation of multi-step transition statistics? Can the authors clarify the distinction between direct one-step reuse and repeated transition-operator composition in Theorem 1? 2.Ablation consistency: Why do the "One-step history" and "Last item" configurations produce identical results across the reported metrics on Beauty and Sports? Please specify the exact changes made to the semantic-state and successor-history components in each experiment. 3.Popularity correction: How much of the successor contribution comes from the popularity-relative log-ratio versus the rank-based reliability term? Could the authors provide a controlled experiment with the mixing coefficient set to zero? 4.Generalization: Could the authors provide matched-anchor and representative baseline comparisons on Yelp and OnlineRetail using the same configurations as the main results in Table 3? 5.Theoretical guarantees: What are the empirical values or distributions of the probability floor, transition-row sample sizes, and resulting uncertainty bounds? Are these bounds sufficiently tight to support meaningful ranking guarantees on the evaluated datasets?

9 comments

Dear Reviewer,

Thank you again for your valuable comments and feedback. As the rebuttal deadline is approaching, we would like to kindly follow up on our responses, in which we have addressed your questions and concerns.

We would greatly appreciate it if you could revisit our responses and consider whether they warrant an upward revision of your score. If you have any remaining questions or concerns, please do not hesitate to let us know. We would be more than happy to provide further clarification.

Thank you for your time and consideration.

Best regards,
The Authors

Thank you for the authors’ response. However, I believe that two key concerns remain unresolved.

  1. Limited methodological novelty beyond combining existing mechanisms. The integration of semantic states and successor operators is reasonable, but the response does not clearly establish what is technically novel beyond this combination. In particular, the distinction from existing approaches, such as feature augmentation and transition-based recommendation, remains insufficiently articulated. This is my primary concern, as it directly affects the assessment of the paper’s methodological contribution. The authors should clarify whether the proposed approach introduces a substantively new mechanism or formulation, and explain precisely how it differs from these established approaches.

  2. The theoretical analysis does not adequately support the central methodological claim. Theorem 1 characterizes the discrepancy between a horizon-specific successor distribution and the direct reuse of a one-step distribution. However, directly reusing a one-step distribution is distinct from performing multi-step propagation using a one-step transition operator. The conditional KL divergence analysis therefore does not establish that estimating a horizon-specific operator is superior to approaches that properly compose one-step dynamics. A theoretical comparison addressing this distinction, or a more carefully scoped claim, is needed.

I would be willing to raise the score if the authors can convincingly address these two concerns, particularly the first one.

Dear Reviewer,

We sincerely thank you for your further evaluation and for identifying the two remaining concerns regarding methodological novelty and theoretical justification. We have strengthened our response through a more explicit mathematical formulation of the proposed mechanism, controlled comparisons with learned feature fusion, and a separate theoretical treatment of direct multi-horizon estimation versus Markovian composition. We first address methodological novelty, which you identified as your primary concern.

1. Methodological Novelty: Target-Distance-Constrained Successor Evidence

1.1. A distinct source-to-horizon mechanism

ATLASRec's contribution is not the independent use of semantic embeddings, higher-order transitions, or additive score correction. It is the explicit construction of candidate-specific successor evidence under a source-to-target distance constraint. Each historical event is interpreted using the conditional transition statistics corresponding to its actual distance from the current prediction target.

For an interaction sequence i1:ti_{1:t} and candidate item jj, the successor field is

dt(j)=∑r=0Hdyn,t−1ηrτ^r+1(it−r,j)d_t(j)=\sum_{r=0}^{H_{\mathrm{dyn},t}-1}\eta^r\hat{\tau}_{r+1}(i_{t-r},j).

Here η∈(0,1]\eta\in(0,1] controls temporal discounting, and τ^r+1(it−r,j)\hat{\tau}_{r+1}(i_{t-r},j) denotes directly estimated, popularity-relative, support-modulated evidence at horizon r+1r+1. Since the event it−ri_{t-r} is precisely r+1r+1 interaction steps away from the target it+1i_{t+1}, ATLASRec imposes the correspondence it↦T^1i_t\mapsto\hat T_1, it−1↦T^2i_{t-1}\mapsto\hat T_2, and it−2↦T^3i_{t-2}\mapsto\hat T_3.

Under the default three-event history, the field becomes

dt(j)=τ^1(it,j)+ητ^2(it−1,j)+η2τ^3(it−2,j)d_t(j)=\hat{\tau}_1(i_t,j)+\eta\hat{\tau}_2(i_{t-1},j)+\eta^2\hat{\tau}_3(i_{t-2},j).

The terms have different conditional meanings: immediate succession from the latest event, two-step succession from the preceding event, and three-step succession from the oldest retained event. All contribute evidence about the same future candidate. The resulting field is combined with the language-grounded semantic compatibility at(j)a_t(j) and behavioral-prior logit bt(j)b_t(j) through ft(j)=bt(j)+αat(j)+βdt(j)f_t(j)=b_t(j)+\alpha a_t(j)+\beta d_t(j). The novel statistical assignment is made before this conventional additive fusion.

1.2. A position–horizon selection formulation

To formalize the difference from generic multi-horizon feature fusion, consider the position–horizon evidence matrix for candidate jj:

Zt(j)r,n=τ^n(it−r,j)Z_t(j)_{r,n}=\hat{\tau}_n(i_{t-r},j),

where 0≤r<Hdyn,t0\le r<H_{\mathrm{dyn},t} indexes historical position and 1≤n≤Hop1\le n\le H_{\mathrm{op}} indexes the transition horizon. Every matrix entry corresponds to a specific historical source and an independently estimated prediction horizon.

A general multi-horizon fusion family can be written as dtfusion(j)=F(Zt(j))d_t^{\mathrm{fusion}}(j)=F(Z_t(j)), where FF is an aggregation or learned mapping over the available position–horizon evidence. Access to all entries, however, does not itself prescribe which source–horizon relationships correspond to the current prediction target.

ATLASRec instead defines a fixed selection mask

Mr,n=I[n=r+1]M_{r,n}=\mathbb{I}[n=r+1],

and evaluates

dt(j)=∑r=0Hdyn,t−1∑n=1HopηrMr,nZt(j)r,nd_t(j)=\sum_{r=0}^{H_{\mathrm{dyn},t}-1}\sum_{n=1}^{H_{\mathrm{op}}}\eta^r M_{r,n}Z_t(j)_{r,n}.

Thus, among all possible position–horizon pairs, only the entries Zt(j)r,r+1Z_t(j)_{r,r+1} contribute. This is a target-distance-constrained selection operator, not an unconstrained mixture over transition features. It supplies a transparent inductive bias: the meaning of each selected statistic follows from the prediction horizon required by the observed source position.

Importantly, the distinction is not that a sufficiently expressive fusion network could never learn the same mapping. Rather, ATLASRec explicitly imposes it and thereby makes the relevant evidence identifiable, inspectable, and independently testable. This formulation pinpoints the technical mechanism whose contribution is examined by the controlled comparisons in Part 2.

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

This paper studies how to model semantic information and behavioral transitions in language-enhanced sequential recommendation. Existing methods mainly use language models to enrich the semantic representations of items and interaction histories, but semantic similarity is not enough to capture the direction of future behavior. In addition, different historical events sit at different time distances from the prediction target, and applying a one-step transition relation uniformly may overlook this difference.

To address this, the authors propose ATLASRec, which combines a semantic state, successor dynamics, and a behavioral baseline to make recommendations. The method uses a frozen language encoder to build a time-weighted semantic state, estimates sparse successor transition operators for different step lengths from training interactions, and selects the corresponding operator according to the distance between a historical event and the prediction target. Finally, the semantic and transition information is fused into the prediction scores of a pre-trained MPT model. The paper also provides related theoretical analysis, and reports performance comparisons, a component ablation, and an efficiency analysis on the Amazon, Yelp, and OnlineRetail datasets.

Critical strengths
  1. The motivation is clear, and the method has a certain degree of novelty. The paper points out the difference between semantic similarity and the direction of behavioral transitions, and builds successor information from the actual distance between a historical event and the prediction target. The semantic state, successor dynamics, and behavioral baseline have clearly separated roles, which gives the method good interpretability. The related theoretical analysis also offers some support for the method design.

  2. The experimental validation is fairly thorough. ATLASRec achieves the best result in 11 of the 12 main comparisons on the three Amazon datasets. Table 2, built on the same MPT checkpoint, shows improvements on all 21 full-candidate-set metrics and includes a paired significance analysis. Table 4 also provides a component ablation and a one-step transition control, and the other experiments cover state construction, popularity correction, language encoder transfer, parameter sensitivity, and efficiency analysis, which together give fairly solid support for the method's effectiveness.

Critical weaknesses
  1. The existing ablation does not yet adequately separate the independent contribution of exact time-distance alignment. Theorem 1 explains the distributional gap that can arise from reusing one-step transitions, but this result mainly concerns the idealized transition distribution and does not directly guarantee that the empirical multi-step operators are effective. The One-step history control in Table 4 removes both the higher-order transition information and the exact time-distance alignment at once, so it is hard to tell the contribution of each. Moreover, the NDCG@10 gains of the full method over One-step history are only 0.0008, 0.0018, and 0.0009 on Beauty, Sports, and Toys. Although the results support the effectiveness of the full method, the independent value of the exact alignment mechanism still needs stronger validation, for example a comparison with a scheme that uses the same higher-order transition information but does not align by the true distance. The existing ablation also lacks information on how this incremental gain varies across random seeds.

  2. There is still some gap between the statistical design of the sparse successor operators and the theoretical guarantees. Eq. (4) uses popularity-relative transition information, but Eq. (5) then introduces a positive count-rank weight, so an observed transition whose relative lift is zero can still receive a positive score. In addition, Top-KsK_s is selected by raw counts, which may miss low-frequency transitions with a high relative lift. The reference weights in Theorem 4 in the appendix use the same support set and ranking as the empirical operator, so the theorem does not directly control the error caused by support selection and ranking instability. This deserves particular attention for longer prediction steps and sparse transition scenarios, and the authors need to explain further what the actual successor weights mean statistically and what range the theoretical result applies to.

Questions for authors
  1. The One-step history control in Table 4 removes both the higher-order transition information and the exact time-distance alignment. Can the authors separate the contribution of each? For example, if the same multi-step transition operators are used but their correspondence to historical positions is changed, would a similar gain still be obtained?

  2. Why does Eq. (5) need to add a count-rank term on top of the popularity-relative transition information? When the relative lift in Eq. (4) is zero, what is the basis for still giving a positive weight? In addition, could using raw counts to determine the Top-KsK_s support set miss low-frequency transitions with a high relative lift? What is the scope of Theorem 4 with respect to this kind of support selection error?

  3. The default parameters in Table 8 are α=2.4, β=0.24, and H=3, but in Table 6 the NDCG@10 for α=2.4 is 0.0727, while the results for β=0.24 and H=3 are both 0.0730. Can the authors explain whether the other parameters, random seeds, and evaluation settings are the same across the three sensitivity experiments? Were all of these hyperparameters chosen on the validation set? Also, why do the Anchor metrics differ across the language encoders in Table 9, and was the same MPT checkpoint used?

  4. Why does Table 3 not provide same-setting comparisons with other recommendation methods on Yelp and OnlineRetail? In addition, do the efficiency statistics in Table 6 cover only online inference after caching, and how is the offline cost of generating language representations and building the multi-step successor operators accounted for?

3 comments

Dear Reviewer,

Thank you again for your valuable comments and feedback. As the rebuttal deadline is approaching, we would like to kindly follow up on our responses, in which we have addressed your questions and concerns.

We would greatly appreciate it if you could revisit our responses and consider whether they warrant an upward revision of your score. If you have any remaining questions or concerns, please do not hesitate to let us know. We would be more than happy to provide further clarification.

Thank you for your time and consideration.

Best regards,
The Authors

Thanks for the additional experiments. The new alignment ablations and baseline comparisons address some of my main concerns. I will maintain my current rating.

We are truly grateful for your encouraging feedback and continued support. We sincerely appreciate the time and effort you have devoted to carefully evaluating our work, as well as your insightful and constructive suggestions, which have been invaluable in strengthening our manuscript. We are particularly thankful for your recognition of our additional experiments and improvements. Your thoughtful engagement throughout the review process means a great deal to us, and we deeply appreciate your willingness to maintain your rating. Thank you once again for your generosity, consideration, and valuable contributions to improving our work.

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

This paper proposes ATLASRec, a language-grounded framework for sequential recommendation that combines semantic state estimation, multi-horizon successor dynamics, and a pretrained behavioral prior. It uses frozen language representations to capture users' current semantic interests and horizon-specific transition statistics to model future interactions. These components are integrated with a pretrained recommender through an additive score combination, which is also characterized by a KL-based formulation. The paper also provides theoretical analyses of horizon alignment and ranking stability, along with extensive experiments demonstrating improvements over existing methods.

Critical strengths
  1. The paper addresses a meaningful problem in sequential recommendation by distinguishing semantic similarity from directional transition dynamics. The motivation for incorporating horizon-specific transition information is intuitive and well aligned with the task.

  2. Although the proposed method is relatively straightforward, the paper provides substantial theoretical analysis of horizon mismatch, KL-based evidence integration, and ranking stability.

  3. The paper compares ATLASRec with various sequential recommendation baselines and provides controlled ablations using the same behavioral backbone to isolate the contributions of individual components. The results show consistent improvements over the behavioral anchor and competitive performance against existing methods.

Critical weaknesses
  1. ATLASRec is evaluated only with MPT as its behavioral backbone. Although other sequential recommenders, such as SASRec and HSTU, are included as baselines, the proposed method is not evaluated with these models as alternative backbones. In addition, Table 3 reports only the absolute performance on Yelp and OnlineRetail without corresponding baseline comparisons.

  2. Horizon alignment seems to be one of the main components of ATLASRec, but its improvements over the one-step history variant in Table 4 are relatively modest compared with the overall gains over the behavioral anchor. This raises some questions about the extent to which horizon alignment contributes to the overall effectiveness of the framework.

  3. The paper frequently uses abstract and elaborate terminology, particularly in the Introduction, which makes the core ideas harder to follow than necessary. For example, expressions such as "language-grounded atlas" and "minimum-information closure" are introduced without sufficiently concrete explanations. In addition, MPT, which serves as the behavioral backbone, is introduced without explaining what the acronym stands for or providing sufficient background.

Questions for authors

NA

3 comments

Dear Reviewer,

Thank you again for your valuable comments and feedback. As the rebuttal deadline is approaching, we would like to kindly follow up on our responses, in which we have addressed your questions and concerns.

We would greatly appreciate it if you could revisit our responses and consider whether they warrant an upward revision of your score. If you have any remaining questions or concerns, please do not hesitate to let us know. We would be more than happy to provide further clarification.

Thank you for your time and consideration.

Best regards,
The Authors

Thank you for the detailed response. The additional experiments have addressed most of my concerns. I will maintain my current score.

Dear reviewer,

We are truly grateful for your encouraging feedback and continued support. We sincerely appreciate the time and effort you have devoted to carefully evaluating our work, as well as your insightful and constructive suggestions, which have been invaluable in strengthening our manuscript. We are particularly thankful for your recognition of our additional experiments and improvements. Your thoughtful engagement throughout the review process means a great deal to us, and we deeply appreciate your willingness to maintain your rating. Thank you once again for your generosity, consideration, and valuable contributions to improving our work.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.