acceptodds
Under review as a conference paper at ICLR 2027

SeqEditor: Editing User Behavior Sequence for Recommendation

Abstract

User Behavior Sequences (UBSs) are the data foundation for Sequential Recommender Systems (SRSs). Existing studies typically regard these sequences as high-quality behavioral features. However, UBSs in real-world scenarios often suffer from the **quality flaw problem**, manifested as noisy and missing items. This problem prevents SRSs from being fully optimized, hindering model performance. To address this issue, we propose **SeqEditor** to improve the quality of UBSs by editing sequences. Concretely, SeqEditor comprises a _Quality Assessor_ and a _Quality Editor_. The Quality Assessor evaluates UBS quality by estimating local training utility from the alignment between sequence gradients and the prediction-loss gradient on held-out reference users. The LLM-based Quality Editor is optimized by performing supervised fine-tuning (SFT) and reinforcement learning (RL) with interest-aware rewards. This enables it to understand user behavior patterns and identify noise. During the inference stage, SeqEditor effectively improves UBS quality by performing potential missing item supplement and noise removal. SeqEditor can be applied to various SRSs, while its text-based Quality Editor supports cross-domain generalization. Extensive empirical results demonstrate the effectiveness of SeqEditor. Our anonymous code is available at https://anonymous.4open.science/r/SeqEditor_2027.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Sequential Recommendation, User Behavior Sequence, Quality Flaw, Large Language Models

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

The paper argues that the user behavior sequences fed to sequential recommenders contain noisy and missing items, and proposes SeqEditor, a model-agnostic procedure that improves sequential recommendation through sequence editing. A Quality Assessor selects seed sequences using normalized gradient alignment with a held-out reference-user objective, averaged across auxiliary SASRec checkpoints. An LLM-based Quality Editor is trained on synthetically corrupted seeds through supervised fine-tuning and DAPO with an interest-aware reward. The editing pipeline applies denoising, directed-PMI-based predecessor insertion, and a second denoising pass. Experiments report improvements across five recommendation backbones (GRU4Rec, BERT4Rec, SASRec, TIGER, HSTU) on three Amazon datasets, together with ablations, cross-domain transfer, and efficiency analyses. The paper also provides a link to an anonymous code repository.

Critical strengths

1.The problem is relevant, and the separation between sequence editing and downstream recommendation makes the framework potentially reusable. Table 1 reports consistent improvements across both conventional and generative recommendation backbones, with five-run averages and significance tests against the corresponding unedited baselines.

2.The assessor has a coherent local-utility interpretation. Equation (5) reasonably connects normalized gradient alignment to a first-order reduction in reference loss. The interest-aware reward also provides a sensible mechanism for discouraging erroneous deletions for users with broader interests.

3.The evaluation includes useful controls beyond the main comparison: random and loss-based seed selection, single-checkpoint scoring, SFT-only editing, denoising/supplementation ablations, denoising baselines, cross-domain editor transfer, and an LLM-embedding comparison. These experiments provide meaningful evidence that the approach is practically useful.

Critical weaknesses

W1.The connection between the operational quality score and actual noise/missingness remains insufficiently established.

  • Positive gradient alignment indicates local usefulness to an auxiliary model and reference distribution; it does not establish that a sequence is free of noisy interactions. Nevertheless, seed interactions are implicitly treated as clean when constructing synthetic labels. Likewise, successful recommendation after PMI-based insertion does not establish recovery of genuinely missing behavior. The ablations and Appendix C.7 support usefulness of the pipeline, but do not fully distinguish behavioral repair from task-oriented augmentation or suppression of atypical preferences. This matters because the paper repeatedly attributes its gains to resolving sequence-quality flaws.

W2. Ambiguity in Test-Time Evaluation Protocol and Prohibitive Online Inference Latency.

  • Section 3.3 and Figure 2 describe editing as SeqEditor's inference stage, while Algorithm 1 (Lines 10–11) specifies editing training data. Although Section 4.5 and Figure 5 suggest editing test histories, whether this applies to the main evaluation remains unclear. If test histories are edited, is editing performed online or precomputed/cached? The pipeline requires two denoising passes, but Table 6 does not clarify whether its reported 45ms per sample covers one pass or the full pipeline, leaving serving costs uncertain. If test histories remain unedited, please clarify how Figure 5 relates to the main evaluation and discuss the potential mismatch between edited training histories and raw test inputs.

W3. Unrealistic Uniform Synthetic Noise and Noise-Free Seed Assumption.

  • In Section 3.2.1, sequences in DH\mathcal{D}_H (where Q(S)>0Q(S) > 0) are treated as completely clean ground truth, and only the 0 to KK (K=5K=5) artificially injected items are labeled as noise YY. However, a positive sequence-level gradient inner product Q(S)>0Q(S) > 0 only indicates positive aggregate local training utility for the sequence as a whole and does not guarantee that every individual interaction vi∈Sv_i \in S is noise-free. Treating original noisy items inside DH\mathcal{D}_H as clean labels introduces false-negative supervision during SFT and RL.

  • More importantly, the NOISE(S)\text{NOISE}(S) operator in Equation (7) injects items uniformly at random from V∖S\mathcal{V} \setminus S. As illustrated in Appendix Figure 7, this frequently inserts completely unrelated cross-category items (e.g., inserting a GPS/Altimeter wristwatch v11v_{11} into a clothing sequence containing men's pajama pants v17472v_{17472} and women's boots v15364v_{15364}). Consequently, the LLM can achieve high validation F1F_1 (Figure 6a) simply by performing trivial category-outlier detection, rather than learning to identify realistic, fine-grained noise such as within-category comparison clicks or popularity-biased accidental clicks.

W4. Context-Insensitive First-Order PMI Supplementation.

  • In Equations (13)–(17), the supplementation function fsup(vi,Vtr)f_{sup}(v_i, \mathcal{V}_{tr}) inserts a predecessor item vk∗v_k^{\ast} before viv_i solely based on the global pairwise transition score svk→viPMI>ηs_{v_k \rightarrow v_i}^{PMI} > \eta, completely ignoring the actual preceding item vi−1v_{i-1} already present in SdenS_{den}.

  • Even when the existing transition vi−1→viv_{i-1} \rightarrow v_i is already semantically coherent, the rule in Equation (17) blindly inserts vk∗v_k^{\ast} between vi−1v_{i-1} and viv_i whenever a globally frequent predecessor exceeds η=0.8\eta = 0.8. Because this operation is applied independently to every item in SdenS_{den}, it risks nearly doubling the sequence length and repeatedly injecting globally popular items.

Questions for authors
  • Q1. Test-Time Protocol and Serving Latency. During evaluation on the test set (Tables 1, 2, and 7), are the test user sequences also edited by SeqEditor (QE→fsup→QEQE \rightarrow f_{sup} \rightarrow QE) prior to next-item prediction? If yes, how can a two-pass 4B LLM pipeline (~90ms per sequence according to Table 6) be practically deployed in low-latency recommendation serving? If no (only training sequences are edited), how do you reconcile the train–test distribution discrepancy and the test-time editing shown in Figure 5?

  • Q2. Robustness to Realistic/Hard Noise Injection. How does SeqEditor perform if NOISE(S)\text{NOISE}(S) in Equation (7) samples harder negative items—such as popularity-sampled items or items within the same coarse category as the user's existing interactions—rather than uniform random items from V∖S\mathcal{V} \setminus S?

  • Q3. Clarification on DH\mathcal{D}_H Selection Ratio: What exact percentage of training sequences satisfy Q(S)>0Q(S) > 0 on Beauty, Sports, and Clothing respectively? Please clarify whether the Quality Editor was trained on S:Q(S)>0\\{S : Q(S) > 0\\} (Section 3.1) or the top 20% subset (Appendix B.1, Equation 18).

  • Q4. Context-Aware Supplementation: In Equation (17), what is the average number of supplemented items per sequence before and after the second denoising pass? Would conditioning supplementation on the transition from the preceding item vi−1v_{i-1} (e.g., only supplementing when svi−1→viPMIs_{v_{i-1} \rightarrow v_i}^{PMI} is low) prevent disrupting already coherent transitions?

7 comments
AuthorsOpenReview ↗

Response to Reviewer RftV

Thank you for your careful review and constructive comments; we hope our responses below resolve your concerns.


Q1: Test-Time Protocol and Serving Latency.


We thank the reviewer for the question. We clarify the Test-Time Protocol and the Serving Latency of SeqEditor separately.


Test-Time Protocol:


SeqEditor adopts the protocol of “editing sequences offline and directly using the edited results at test time to predict the held-out item.” The recommendation model is trained on edited sequences, and at test time it directly reads the corresponding edited sequences for prediction.Therefore, the main experiments use the same sequence processing for training and test, and SeqEditor does not need to be called again during test-time prediction.


Serving Latency:


In deployment, SeqEditor runs offline and is not part of the real-time serving path. Online requests directly use the saved edited sequences for recommendation; user's new interactions are appended in chronological order for prediction and edited only in the next offline update. Thus, the ~90 ms per sequence in Table 6 is an offline cost, not added latency per online recommendation.


We have added this clarification to Appendix C.5, and have also revised Section 4.5 to state explicitly that the editing shown in Figure 5 is performed offline and that the edited sequence is used for next-item prediction.

Q2: Robustness to Realistic/Hard Noise Injection.


Thank you for raising this concern. Eq. (7) is not meant to exactly simulate every real accidental click. Instead, it constructs a large and diverse set of self-supervised perturbations that cover common noise patterns. For the two harder negatives you mention—popularity-sampled items and items within the same coarse category—we respond from both Methodological Rationale and Experimental Validation.


Methodological Rationale:


Related Studies: Real accidental clicks may come from exposure to popular items, confusion between similar products, or unintended touches. Since ground-truth noise labels are unavailable, random perturbation is a practical self-supervised way to build diverse training signals; this strategy has also been used in STEAM (WWW 2023)[1] and LLM4DSR (TOIS 2025)[2].


  [1] A Self-Correcting Sequential Recommender, WWW 2023.
  [2] Llm4dsr: Leveraging large language model for denoising sequential recommendation, TOIS 2025.

Training Setup: During training, SeqEditor varies the injected items, their number, and their positions, and resamples them from the full item space each epoch. This helps the model learn stable behavioral patterns instead of memorizing fixed noise.


Experimental validation.


We compare three noise injection strategies, all evaluated with SASRec as the backbone:


  • Random Noise (RN): the random injection used in the paper.
  • Popularity-based Noise (PN): items are sampled by occurrence frequency and injected, simulating accidental clicks caused by exposure to popular items.
  • Semantic Noise (SN): Items are sampled with probabilities proportional to their textual semantic similarity to items in the sequence and then injected, simulating accidental clicks caused by confusion between similar products.

Table 1: Recommendation performance with different noise injection strategies.


Noise InjectionBeauty R@10Beauty N@10Sports R@10Sports N@10Clothing R@10Clothing N@10
PN0.06400.03350.04000.02040.02990.0134
SN0.06280.03280.03900.02000.02930.0132
RN0.06580.03410.04140.02140.03080.0142

As shown in Table 1, SeqEditor trained with random noise consistently achieves the best recommendation performance. This result suggests that diverse random perturbations capture a broad range of real-world noise patterns.


We have added these experiments to Appendix C.11 of our paper.

Q3: Clarification on DH\mathcal D_H Selection Ratio.


Thank you for the question. We clarify two points: the seed ratio for Quality Editor, and the role of the top 20% in Appendix B.1.


Training seeds and selection ratio of Quality Editor. The training seeds used by Quality Editor in both SFT and RL are selected by the Q(S)>0Q(S)>0 criterion in Section 3.1, not by a fixed top 20%. On Beauty, Sports, and Clothing, sequences satisfying this criterion account for 34.18%, 27.69%, and 31.65% of the training sequences, respectively. The denominator is the candidate training set D=Dtr∖R\mathcal{D} = \mathcal{D}_{tr} \setminus \mathcal{R}, i.e., the training sequences after excluding the reference set.


Role of the top 20% in Appendix B.1. This rule is used only for the high- vs. low-quality comparison in Section 1 and 4.4. Specifically, we select the top and bottom 20% by score, train SASRec-HQ and SASRec-LQ, and compare the effect of sequence quality under the same data scale. This split is not used to construct Quality Editor's training data. Our manuscript reuses DH\mathcal{D}_H in both places, which may cause confusion.


We have renamed the seed set as Dseed\mathcal{D}_{seed} in Sections 3.1–3.2 and Eq. (7), and added a note in Appendix B.1 clarifying that the top-20% split DH\mathcal{D}_H and DL\mathcal{D}_L is used only for the comparison experiments and is not the training seed set.

Q4: Context-Aware Supplementation.


Thank you for the question. We respond to two points: (i) the number of supplements before and after the second denoising pass; (ii) the conditional supplementation experiment.


(i) Supplemented items before and after the second denoising pass.


We count the average number of supplement items before the second denoising pass and the average number of newly added items that remain after it. The statistics cover all sequences to be edited, including those with no supplementation. After denoising, we count only the newly added items that are still kept.


As shown in Table 2, SeqEditor does not keep all supplemental items produced by Eq. (17). It first supplements items and then filters them with full context.


Table 2: Average number of supplements per sequence before and after the second denoising pass.

DatasetBefore 2nd denoising: avg. supplements per sequenceAfter 2nd denoising: avg. kept supplements per sequence
Beauty1.140.62
Sports1.210.71
Clothing0.910.58

(ii) Conditional supplementation experiment


We implement the suggested variant: supplement only when the original transition satisfies svi−1→viPMI<τs_{v_{i-1}\rightarrow v_i}^{\text{PMI}} < \tau. We set τ∈[0.2,0.5,0.8]\tau \in [0.2,0.5,0.8 ], with SASRec as the backbone. Results are shown in Table 3.


Table 3. Recommendation performance under conditional item supplementation with different thresholds.

SettingBeauty (R@10 / N@10)Sports (R@10 / N@10)Clothing (R@10 / N@10)
SeqEditor0.0658 / 0.03410.0414 / 0.02140.0308 / 0.0142
Conditional, τ\tau=0.20.0625 / 0.03300.0378 / 0.02030.0296 / 0.0137
Conditional, τ\tau=0.50.0646 / 0.03370.0405 / 0.02110.0301 / 0.0139
Conditional, τ\tau=0.80.0657 / 0.03430.0417 / 0.02130.0304 / 0.0140

Results show that tightening τ\tau consistently lowers performance. Eq. (17) targets high recall at candidate supplementation generation, leaving precision to the second Quality Editor pass over the full context. Conditioning supplementation adds an extra hyperparameter τ\tau that pre-filters candidates. This trades recall for precision: some unreasonable candidates are filtered out early, but reasonable ones removed at this stage cannot be recovered later.


We thank the reviewer for this valuable direction. This two-stage design (high recall in candidate generation, precision from the second Quality Editor pass) achieves an effectiveness–efficiency balance in our experiments.

Dear Reviewer RftV,

We hope this message finds you well.

Thank you for your insightful and constructive comments on our manuscript. They have been invaluable in helping us improve the quality and clarity of our work. We have carefully responded to each of your comments and revised the manuscript accordingly.

We would be grateful if you could kindly review our response and the revised manuscript at your earliest convenience. Please let us know if any further clarification or additional information is needed; we would be happy to provide it.

Thank you again for your time and thoughtful review. We look forward to hearing from you.

Sincerely,

The Authors

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

The authors leverage the augmentation pipeline in sequential recommendation.

Specifically, for a particular given recommender system, they obtain the gradient-based signal that quantifies the consistency of a particular item within a sequence.

Then, they refine the sequence based on the signal using the RL-based refinement pipeline.

The authors demonstrate the effectiveness under several backbone recommender systems and datasets.

Critical strengths
  • S1. The key motivation of the research topic, refinement of the user sequences, is clear and sound.

  • S2. Experiments are conducted under diverse backbone recommender systems.

Critical weaknesses
  • W1. Generalizability across recommendation models. The informativeness of a user sequence may depend on the backbone recommender. However, the proposed quality score is computed using an auxiliary SASRec model, and it remains unclear whether the resulting quality rankings and editing decisions transfer consistently to other architectures. The authors should examine this dependence and clarify why edits guided by one model should benefit others.

  • W2. Insufficient discussion of related work. The paper does not adequately discuss existing studies on sequence augmentation for sequential recommendation [1, 2]. These works are closely related to the proposed sequence-editing framework. The authors should cover them and clarify the methodological differences.

  • W3. Computational overhead and deployment practicality. It is unclear whether sequence editing is also required when processing user histories at recommendation inference time. If so, the two LLM-based denoising passes could introduce substantial overhead in latency-sensitive recommendation systems. Although Appendix C.5 reports inference costs, the authors should clarify whether these measurements cover the complete editing pipeline and quantify its impact on end-to-end recommendation latency and throughput.

  • [1] Data Augmentation for Sequential Recommendation: A Survey

  • [2] Sequential Data Augmentation for Generative Recommendation

Questions for authors

See Weakness

6 comments
AuthorsOpenReview ↗

Responses to Reviewer QjZE

Thank you for your careful review and constructive comments; we hope our responses below resolve your concerns.


Q1:Generalizability Across Different Recommendation Models.


We thank the reviewer for this comment. We response auxiliary model choice and its cross-architecture transferability via Methodological Analysis and Experimental Validation.


Methodological Analysis:


SeqEditor evaluates the training value of a sequence by whether it helps predict the reference user's behavior. For an auxiliary model mm, the score is (Eq. (4) in our paper):


qm(S;θm)=∇θmLR(m)(θm)⊤∇θmlm(S;θm)∥∇θmlm(S;θm)∥2+ϵ.q_m(S; \theta_m)=\frac{\nabla_{\theta_m}L_R^{(m)}(\theta_m)^{\top} \nabla_{\theta_m}l_m(S;\theta_m)}{\left\|\nabla_{\theta_m}l_m(S;\theta_m)\right\|^{2}+\epsilon}.


Here, LR(m)L_R^{(m)} is the next-item prediction loss on the reference user. dm(S)d_m(S) denote the normalized sequence gradient. By Eq.(5) in our paper, a small training update satisfies:


LR(m)(θm−αdm(S))−LR(m)(θm)=−αqm(S)+O(α2).L_R^{(m)}(\theta_m-\alpha d_m(S)) - L_R^{(m)}(\theta_m) = -\alpha q_m(S)+O(\alpha^2).


A positive score means the sequence's training update helps reduce the reference user's prediction loss. Importantly, this score is mainly determined by the data, not by the specific reference model. Different recommenders use different architectures but learn user preferences and item transitions from the same data. The score selects training sequences around this shared objective, and SeqEditor learns to denoise noisy labels, improving the behavioral data that different backbones rely on. This supports SeqEditor's cross-architecture transferability.


Experimental Validation:


Our experiments show that SASRec-based quality scoring generalizes across downstream backbones. To further examine the sequence quality scores from other auxiliary models, we test three additional quality scoring schemes. Specifically, we use the scores produced by GRU4Rec, BERT4Rec, and an equally weighted GRU4Rec+BERT4Rec combination to optimize SeqEditor, all with SASRec as the downstream recommender.


Table 1: Recommendation performance with different auxiliary quality scoring models.

Auxiliary Scoring ModelBeauty R@10Beauty N@10Sports R@10Sports N@10Clothing R@10Clothing N@10
SASRec0.06580.03410.04140.02140.03080.0142
GRU4Rec0.06550.03420.04090.02110.03100.0143
BERT4Rec0.06520.03370.04060.02100.03060.0141
GRU4Rec + BERT4Rec0.06600.03430.04120.02130.03120.0145

As shown in Table 1, different auxiliary scoring models give comparable results, all close to SASRec-based scoring, indicating that the quality signal does not depend on the specific auxiliary architecture. In addition, the edited data improves all five downstream backbones in Table 1 of the paper (GRU4Rec, SASRec, BERT4Rec, TIGER, and HSTU), showing that the editing decisions also transfer across downstream architectures. Together, these results support SeqEditor's cross-architecture transferability.


We have added the above experimental setup, results, and related analysis to Appendix C.12 of our paper.

Q2:Discussion of Related Work on Sequence Augmentation.


We thank the reviewer for the time and for the valuable suggestion on the related work discussion. We have carefully studied these two works [1, 2] and have expanded the related work in Appendix A, adding their main methods and their differences from SeqEditor.

  [1] Data Augmentation for Sequential Recommendation: A Survey
  [2] Sequential Data Augmentation for Generative Recommendation.

Q3:Computational Overhead and Deployment Practicality.


We thank the reviewer for this comment. We clarify SeqEditor's Testing Protocol and Serving Latency as follows.


  • Testing Protocol: SeqEditor follows an "offline editing, test-time reuse" protocol: the recommender is trained on edited sequences, and at test time it directly reads the corresponding edited history for prediction. Training and testing therefore use consistent history processing, and SeqEditor is not invoked again at test time.

  • Serving Latency: In deployment, SeqEditor runs as an offline batch process, synchronized with the periodic updates of the downstream recommender, and does not sit in the real-time request path. Edited sequences are stored and reused for model training and later prediction. Real-time recommendation is therefore handled by the downstream recommender alone.

In addition, when an online request arrives, the recommender directly uses the stored edited sequence. New user interactions are simply appended and are not edited on the fly; they are handled in the next offline update. Thus, even if two calls take about 90 ms per sequence (Table 6, Appendix C.5), this is an offline cost and does not add to online request latency. Online throughput depends only on the recommender, not on the editing pipeline.


We have added this clarification to Appendix C.5

We thank you again for your time and thoughtful feedback, which has helped improve our manuscript. We hope our responses address your concerns and would be glad to clarify any further points.

Dear Reviewer QjZE,

We hope this message finds you well.

Thank you for your insightful and constructive comments on our manuscript. They have been invaluable in helping us improve the quality and clarity of our work. We have carefully responded to each of your comments and revised the manuscript accordingly.

We would be grateful if you could kindly review our response and the revised manuscript at your earliest convenience. Please let us know if any further clarification or additional information is needed; we would be happy to provide it.

Thank you again for your time and thoughtful review. We look forward to hearing from you.

Sincerely,

The Authors

Reviewer_aXB8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SeqEditor, a model-agnostic user behavior sequence editing framework for sequential recommendation systems. The authors point out that real-world user behavior sequences often contain noisy items, such as accidental clicks, and may also suffer from missing items due to exposure bias or offline purchases, which can negatively affect the training of recommendation models. To address these issues, the authors introduce SeqEditor, which consists of two core modules: Quality Assessor and Quality Editor. The Quality Assessor is designed to select high-quality user behavior sequences, while the Quality Editor learns to identify noisy items and edit user behavior sequences to improve data quality before they are used to train recommendation models. The authors evaluate SeqEditor on three Amazon Reviews datasets using five sequential recommendation models, reporting an average performance improvement of approximately 10.55%.

Critical strengths
  1. The paper addresses the quality issues of user behavior sequences in sequential recommendation, presenting a clear research motivation and practical significance. Unlike existing methods that primarily mitigate the impact of noise through attention mechanisms and other techniques, SeqEditor directly edits user behavior sequences at the data level, addressing both noise removal and missing item supplementation.
  2. The proposed Quality Assessor demonstrates a certain degree of novelty. Unlike traditional sample selection methods based on prediction loss, it evaluates the local impact of individual behavior sequences on the prediction loss of reference users through gradient alignment, thereby connecting sequence quality assessment with the optimization objective of recommendation models.
  3. The authors combine SFT and RL to optimize the LLM and introduce an Interest-Aware Reward function to account for user interest diversity. This reward mechanism imposes a higher penalty for false positive noise predictions for users with diverse interests, reducing the risk of incorrectly removing legitimate interactions and helping the model better distinguish diverse user preferences from noisy behaviors.
Critical weaknesses
  1. The proposed SeqEditor seems to be a form of data augmentation framework. Given the strong semantic understanding capabilities of LLMs, it seems feasible to directly leverage an LLM for sequence editing instead of using the proposed framework. The authors should further clarify the advantages of SeqEditor over such a straightforward approach. Additionally, I suggest including comparisons with existing data augmentation methods to better demonstrate the effectiveness of the proposed framework.

  2. Is it possible that the sequence editing process incorrectly removes genuine but infrequent user interactions? Such erroneous removals may introduce biases into user preference modeling, particularly in capturing diverse or long-tail interests. The authors should further investigate this potential issue.

  3. The training data for the Quality Editor is constructed by randomly inserting noisy items into user behavior sequences. However, these randomly inserted items may be semantically quite different from the original sequences, making them relatively easy to identify. I suggest conducting additional experiments with semantically similar noisy items to evaluate how different noise types affect sequence editing quality and downstream recommendation performance.

Questions for authors
  1. Given the strong semantic understanding capabilities of LLMs, could the authors clarify the advantages of SeqEditor over directly using LLMs for sequence editing? Would it be possible to include comparisons with simpler LLM-based editing approaches and existing data augmentation methods?

  2. Could the authors investigate whether SeqEditor may incorrectly remove genuine but infrequent user interactions? How does the proposed framework ensure that such removals do not negatively affect the modeling of diverse or long-tail user preferences?

  3. Since the Quality Editor is trained using randomly inserted noisy items, how robust is the proposed method to more challenging noise, such as semantically similar items? Could the authors provide additional experiments to evaluate the impact of different noise types on editing quality and downstream recommendation performance?

6 comments
AuthorsOpenReview ↗

Response to Reviewer aXB8

Response to Reviewer aXB8

Thank you for your careful review and constructive feedback. We hope our responses below help resolve your concerns.

Q1:Comparing direct LLM editing with SeqEditor.


To address the advantages of SeqEditor over directly using LLMs for sequence editing, we respond from two aspects, including Mechanism Analysis and Experimental Validation.


Mechanism Analysis:


Directly using LLMs for editing relies only on semantic understanding, which causes two limitations: misclassifying rare genuine interests as noise, or inserting semantically related but behaviorally illogical items. SeqEditor specifically addresses these issues:


  • For denoising, the Quality Assessor selects sequences with high training value, which trains the editor to identify noise. Additionally, interest-aware RL penalizes the erroneous removal of genuine items.

  • For supplementation, rather than relying on the LLM's free generation, SeqEditor mines candidates from real interaction co-occurrence statistics (Eq. 13–14) and uses the Quality Editor for verification. This avoids behavioral logic deviations and hallucinations from pure semantic generation.

In summary, unlike direct LLM editing, SeqEditor replaces purely semantic-driven generation with precise filtering constrained by both data supervision and behavioral statistics. This ensures effective denoising while protecting long-tail genuine interests from misremoval.


Experimental Validation:


In Appendix C.9 we compared SeqEditor with three recent sequential denoising methods: ECL-SR(RecSys 2023) [1], SSDRec(ICDE 2024) [2] , and LLM4DSR (TOIS 2025) [3]. The results show that SeqEditor clearly outperforms these baselines.


  [1] Equivariant contrastive learning for sequential recommendation, RecSys 2023.
  [2] SSDRec: Self-Augmented Sequence Denoising for Sequential Recommendation, ICDE 2024.
  [3] LLM4DSR: Leveraging Large Language Model for Denoising Sequential Recommendation, TOIS 2025.

We further introduce a simpler prompt-based sequence-editing baseline (directly using LLMs for sequence editing):


  • PromptEdit: This method directly uses Qwen3-4B without editing training. The prompt instructs the model to assess sequence quality by considering item semantics, functional relationships, usage scenarios, interaction order, and interest shifts. It edits sequences identified as low quality.

The results for the methods in Appendix C.9 and PromptEdit are reported in Table 1.


Table 1. The performance of different editing-based methods.

MethodBeauty R@10Beauty N@10Sports R@10Sports N@10Clothing R@10Clothing N@10
ECL-SR0.06210.03230.03630.01940.02780.0132
SSDRec0.06250.03260.03780.01990.02840.0134
LLM4DSR0.06300.03290.03820.02010.02870.0135
PromptEdit0.06190.03210.03610.01920.02760.0130
SeqEditor0.06580.03410.04140.02140.03080.0142


The results show the effectiveness of SeqEditor. We have update the results and descrbution of PromptEdit in Appendix C.9 of our paper.

Q2: Investigating Incorrect Removals on Diverse and Long-tail User Preferences.


We thank the reviewer for raising this valid concern. We respond from both Mechanism Analysis and Experimental Validation.


Mechanism Analysis:


Genuine but infrequent interactions indeed pose a risk of misclassification. To address this, SeqEditor explicitly accounts for the cost of erroneous removal during training. The Interest-Aware Reward not only rewards correct noise identification but also penalizes the removal of genuine interactions, with the penalty scaling up as the user's interest breadth increases. This forces the editor to be more conservative when handling multi-interest behaviors, thereby reducing the risk of treating secondary interests as noise.


In addtion, while denoising cannot guarantee zero false positives, the subsequent supplementation operation can further restore and enrich the diverse or long-tail preferences based on behavioral co-occurrence statistics.


Experimental Validation:


We conducted experiments (using SASRec) to examine editing accuracy and the preservation of long-tail preferences.


First, Figure 9 in Appendix C.6 compares the standard F1 reward with our Interest-Aware Reward. The latter achieves higher denoising Precision across user groups with varying numbers of interests, validating its effectiveness in reducing false removals.


Second, we compare the downstream recommendation performance of SASRec models trained on original sequences versus those trained on SeqEditor-edited sequences, reporting the following metrics:


  • Category Diversity@10 (CD@10): The average number of distinct categories in the Top-10 recommendations. A higher value indicates greater recommendation diversity..

  • Tail Recall@10 (TR@10): Recall computed only for test items appearing fewer than 10 times in the training set, It measures the model’s ability to preserve long-tail interests.


Table 2. Performance comparsion in CD@10 and TR@10.

DatasetOriginal Sequence (CD@10)Edited Sequence (CD@10)Original Sequence (TR@10)Edited Sequence (TR@10)
Beauty4.28104.29560.01780.0197
Sports5.04285.07560.00930.0109
Clothing5.43215.42070.00680.0078

Table 2 shows that SeqEditor consistently improves Tail Recall (TR@10) across all datasets while maintaining comparable Category Diversity (CD@10). This indicates that SeqEditor enhances long-tail prediction without compromising recommendation diversity, effectively preserving diverse and long-tail user preferences.

Q3: Testing SeqEditor on Different Noise Strategies.


We thank the reviewer for this valuable comment. Random noise is primarily used to broaden training perturbations and reduce reliance on specific noise patterns. We clarify this via Methodological Analysis and Experimental Validation.


Methodological Analysis:

  • Related Studies: Real accidental clicks may result from exposure to popular items, confusion between similar products, or unintended touches. In the absence of ground-truth noise labels, random perturbation provides a practical self-supervised approach for constructing diverse training signals for sequence denoising. This strategy has also been adopted in previous studies. For example, STEAM (WWW 2023)[1] trains its sequence corrector using randomly inserted and deleted items, while LLM4DSR (TOIS 2025)[2] constructs LLM denoising data through random item replacement.
  [1] A Self-Correcting Sequential Recommender, WWW 2023.
  [2] Llm4dsr: Leveraging large language model for denoising sequential recommendation, TOIS 2025.
  • Training Setup: SeqEditor generates diverse perturbations for each sequence by varying the injected items, their number, and positions, and resampling them from the full item space in each training epoch. This encourages the model to learn stable behavioral patterns rather than memorize fixed noise.

Experimental Validation:


To examine the effect of different noise injection strategies, we train the Quality Editor using the following three strategies and evaluate their recommendation performance with SASRec as the backbone:


  • Random Noise (RN): We use the random noise injection strategy described in the paper.
  • Popularity-based Noise (PN): Items are sampled according to their occurrence frequencies and then injected into the sequences, simulating accidental clicks caused by exposure to popular items.
  • Semantic Noise (SN): Items are sampled with probabilities proportional to their textual semantic similarity to items in the sequence and then injected, simulating accidental clicks caused by confusion between similar products.

Table 3 shows that random noise (RN) consistently outperforms both popularity-based (PN) and semantic noise (SN). This indicates SeqEditor learns a generalizable editing policy, demonstrating strong robustness to diverse and challenging noise types. We have added this experiments in Appendix C.11 of our paper.


Table 3: Recommendation performance with different noise injection strategies.

Noise InjectionBeauty R@10Beauty N@10Sports R@10Sports N@10Clothing R@10Clothing N@10
PN0.06400.03350.04000.02040.02990.0134
SN0.06280.03280.03900.02000.02930.0132
RN0.06580.03410.04140.02140.03080.0142

We sincerely thank you again for your time and thoughtful feedback, which has helped us improve our manuscript. We hope our responses address your concerns and are happy to provide any further clarification if needed.

Comments on the paper

2 comments
Area_Chair_zHLUOpenReview ↗

Reminder: Reviewer–Author Discussion

Thank you to the reviewers who have already engaged in the discussion and provided valuable feedback. As the discussion deadline is approaching, I would kindly encourage the remaining reviewers to review the authors' responses and share their follow-up comments, particularly on whether their concerns have been adequately addressed.

Authors are also encouraged to respond promptly to any further questions. Thank you all for your time and contributions to a constructive discussion.

Dear AC, SAC and PC,

Thank you for your continued follow-up during the discussion and for repeatedly encouraging the reviewers to engage. We have responded to every question raised by the reviewers, and have provided both analytical and empirical clarifications for each of them. We hope these materials can be helpful for your final assessment.

Sincerely,

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.