acceptodds
Under review as a conference paper at ICLR 2027

IntExpo: Exploring Diverse Target Intents for Cross-Domain Recommendation

Abstract

Cross-domain recommendation (CDR) leverages source-domain interactions to alleviate data sparsity in a target domain. Recent cross-domain sequence synthesis (CDSS) methods synthesize target-domain histories for source-only users to augment training data. However, a user's target-domain behavior depends on both that user's transferable preferences and target-specific intents. Source histories reveal transferable preferences that transfer across domains, but similar transferable preferences may lead to different target intents and behavior sequences. Existing CDSS methods often generate overly similar sequences for users with similar preferences, leaving other plausible intents underrepresented. We call this problem **Target-Intent Collapse**. To address it, we propose **IntExpo** to synthesize target-domain sequences from transferable preference profiles and diverse target-intent trajectories. Specifically, IntExpo comprises a _Transferable Preference Profiler_, an _Intent Generator_, and a _Behavior Generator_. Its Transferable Preference Profiler proposes discriminator-guided reinforcement learning (RL) to extract transferable profiles from source histories. The Intent Generator learns a conditional distribution over intent trajectories, using trajectories observed from users with similar profiles as additional supervision. Finally, the Behavior Generator produces multiple plausible target-domain sequences conditioned on a profile and diverse intent trajectories. Experiments across four cross-domain transfer settings show that IntExpo outperforms competitive baselines even under the same synthesis budget. Further analyses show that it generates diverse, plausible behaviors and alleviates the target-intent collapse problem. Our anonymous code is available at https://anonymous.4open.science/r/IntExpo-5672.

Primary area:
generative models
Keywords:
Cross-Domain Recommendation, Cross-Domain Sequence Synthesis, Target-Intent Collapse, Large Language Models

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_6wYEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on the "Target-Intent Collapse" problem in cross-domain recommendation, where existing methods tend to synthesize overly similar target-domain sequences for users with similar source-domain preferences. To address this, the authors propose IntExpo, a three-stage framework. First, a Transferable Preference Profiler uses reinforcement learning guided by a domain discriminator to extract domain-agnostic preference profiles from source histories. Second, an Intent Generator clusters historical sessions and learns a conditional intent distribution using a collaborative loss to capture multiple plausible intents. Finally, a Behavior Generator fine-tunes an LLM to generate specific target-domain sequences based on the user's profile and the sampled intent trajectories.

Critical strengths
  1. The paper identifies a practical issue in generating synthetic data for cross-domain recommendations, which involves mapping a single source history to multiple plausible target intents. Highlighting this one-to-many relationship is a valid and useful perspective for improving data augmentation.
  2. The authors conduct extensive experiments across multiple cross-domain scenarios, including cold-start settings. The detailed ablation studies also help validate the contribution of individual components.
Critical weaknesses
  1. The proposed framework is very heavy. It requires training a domain discriminator, running reinforcement learning, performing clustering, training an autoregressive transformer, and fine-tuning a large language model. The core motivation is simply to generate diverse target sequences. However, the paper does not compare IntExpo with a straightforward baseline, such as using a frozen large language model with an adjusted temperature or a prompt that explicitly asks the model to "generate 3 diverse intent sequences". Given the high training costs reported in Table 6, it is hard to justify this multi-stage pipeline without proving that simpler generative prompting strategies fail to achieve similar diversity.

  2. In Section 3.1, the method uses a domain discriminator as a reward signal to penalize domain-specific information. The main risk here is that the language model might learn to bypass the discriminator by outputting generic, meaningless phrases (e.g., "the user likes good products"). If the profile loses fine-grained personalization just to become domain-agnostic, it harms the subsequent generation steps. The paper lacks qualitative or quantitative evidence to show that the profiles remain informative after this RL training.

Questions for authors
  1. Could you provide a performance comparison against a simple baseline where a single frozen language model is explicitly prompted to generate diverse target sequences (e.g., via higher temperature or diversity-promoting prompts)? How do you justify the computational overhead of IntExpo compared to such simple approaches?

  2. Can you provide qualitative examples of the user preference profiles before and after the reinforcement learning training? How do you ensure the profiles do not collapse into generic, low-information text?

  3. How sensitive is the framework to the 120-day session threshold? Why did you choose such a long temporal heuristic to segment intents instead of relying on the semantic understanding capabilities of the language models?

3 comments
AuthorsOpenReview ↗

Responses to Reviewer 6wYE

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

Q1: Comparison with a Frozen LLM and Computational Overhead.


Thank you for this suggestion. We have compared IntExpo with a simple frozen-LLM generation approach and address your question through Experimental Validation and Computational Cost Analysis.


Experimental Validation:


For a fair comparison, we use Qwen3.5-2B-Base, the same backbone as IntExpo’s Behavior Generator, to build the frozen-LLM baseline. We adapt IntExpo’s behavior-generation prompt by retaining the target-domain description, candidate-item information, and output-format requirements, while replacing the preference profile and intent conditions with the user’s source-domain history.


The model is instructed to generate a plausible target-domain sequence from this history. We apply no task-specific fine-tuning and evaluate higher temperatures T∈[1.5,2.0,2.5]T \in [1.5, 2.0, 2.5 ]. All methods generate one target-domain sequence per source-only user, keeping the generation backbone and synthetic data volume consistent. Full results are shown in Table 1.


Table 1. The performance of IntExpo and frozen LLM at different temperatures.

ModelsMovie → Book (R@10 / N@10)Book → Movie (R@10 / N@10)Clothing → Sports (R@10 / N@10)Sports → Clothing (R@10 / N@10)
Frozen LLM (T=1.5)0.0756 / 0.04680.1505 / 0.08690.1029 / 0.06300.0605 / 0.0375
Frozen LLM (T=2.0)0.0809 / 0.04930.1488 / 0.08570.0975 / 0.05780.0628 / 0.0386
Frozen LLM (T=2.5)0.0812 / 0.04880.1490 / 0.08700.0984 / 0.05890.0570 / 0.0382
IntExpo0.0853 / 0.05340.1570 / 0.09150.1155 / 0.06850.0854 / 0.0476

The results show that the frozen model performs significantly worse than IntExpo at different temperatures. Higher temperatures increase randomness but do not ensure target intents that match user preferences. IntExpo learns a preference-conditioned intent distribution from real interactions and similar users’ intent trajectories, guiding the generation of more useful sequence for downstream recommendation.


Computational Cost Analysis:


Compared with the frozen LLM model, IntExpo adds one preference-profile extraction per user and intent generation with a lightweight model. These steps explicitly model target intents compatible with user preferences and provide personalized guidance for behavior generation. IntExpo runs offline to produce additional training data for the recommender. Online recommendation still uses the same SASRec model, without invoking IntExpo’s profile or intent generation modules, so it incurs no additional online generation cost.


Furthermore, the frozen LLM had an average generation failure rate of 15.28% in this experiment (output format errors). We repeated inference for failed cases until all required valid sequences were obtained before comparing performance. IntExpo showed no such failures in this experiment, avoiding the corresponding retry costs. Together with the recommendation gains above, these results show that IntExpo’s additional steps improve both the usefulness of synthetic data and output reliability. Compared with the frozen LLM, these reusable offline computations therefore provide clear practical benefits and a reasonable trade-off between cost and performance.

Q2: Profile Examples and Safeguards Against Collapse Before and After RL


Thank you for your questions. We respond in two parts, including Qualitative Examples and Safeguards Against Collapse before and after RL.


Qualitative Examples:


IntExpo aims to reduce domain-specific information while retaining users’ core interests and preferred attributes. The illustrative examples below show this transition: profiles before RL focus on source-domain items or content formats, while profiles after RL express interests and needs that can transfer across domains.


Table 2. Illustrative preference profiles before and after RL.

Transfer Scenario and Source HistoryBefore RLAfter RL
Clothing → Sports : A lightweight waterproof hiking jacket, a thermal base layer, and durable quick-drying hiking trousers.Core interests: Functional clothing, including hiking jackets, outdoor trousers, and thermal underwear. Preferred attributes: Waterproof, quick-drying fabrics, lightweight and durable materials, warmth, and wearing comfort.Core interests: Hiking and outdoor activities in different weather conditions. Preferred attributes: Lightweight and durable options, protection from rain and cold, and comfort during sustained activity.
Movie → Book : A film about an astronaut solving survival problems through science, a documentary on spaceflight technology, and a documentary explaining cosmic evolution.Core interests: Science-themed films and documentaries about spaceflight and the universe. Preferred attributes: Scientific detail, clear explanations, and logically structured problem-solving plots.Core interests: Space exploration, understanding the universe, and solving problems through science. Preferred attributes: Concrete scientific detail, clear explanations of underlying principles, and coherent analysis and reasoning.

These examples illustrate the abstraction of source-domain descriptions into transferable preferences. Preferences for waterproof fabrics and thermal clothing become outdoor protection needs that can guide sports-equipment generation. Preferences for science films and documentaries become interests in scientific exploration and clear reasoning that can guide book generation. These examples illustrate how IntExpo’s RL-based profiler can effectively extract transferable preferences while preserving the user-specific information needed for personalized generation.


Safeguards Against Collapse:


Thank you for highlighting this concern. Our prompt explicitly requires the profiler to retain each user’s core interests and preferred attributes. We also use KL regularization to constrain deviations from the reference policy during RL; its role in preventing degenerate solutions is discussed in Section 3.1 of the paper.


Our w/o DD ablation shown in Table 3 (Section 4.3 in our paper) provides supporting evidence. Removing discriminator-guided RL reduces recommendation performance across all transfer directions. This supports the usefulness of RL-optimized profiles for personalized transfer, beyond simply making their source domain harder to identify.


Table 3. Ablation of discriminator-guided RL.

ModelMovie → Book (R@10 / N@10)Book → Movie (R@10 / N@10)Clothing → Sports (R@10 / N@10)Sports → Clothing (R@10 / N@10)
w/o DD0.0784 / 0.04540.1548 / 0.08970.1139 / 0.06600.0836 / 0.0462
IntExpo0.0853 / 0.05340.1570 / 0.09150.1155 / 0.06850.0854 / 0.0476

Q3: Session Threshold Sensitivity and Why Not Use LLM Segmentation.


We thank the reviewer for this question. We respond in two parts: (i) How the session split threshold Δ\Delta is chosen and how sensitive the framework is to it; (ii) Why we do not use an LLM for segmentation.


(i) Choice of the session split threshold and sensitivity analysis


Statistical Estimation. For each target domain, we take the 85th percentile of time gaps between consecutive interactions in the training sequences, rounded up to whole days, as that domain's threshold. Only the longest ~15% of gaps count as possible intent switches, so the threshold follows each domain's own rhythm. Across all four settings, this value lands near 120 days, so we use Δ\Delta = 120 days as the shared default.


Sensitivity Analysis. We further evaluated Δ∈[30,60,90,120,180]\Delta \in [30,60,90,120,180] days, with results shown in Table 4. Results show that performance peaks or is near-optimal around Δ = 120, and drops noticeably only when Δ is too small (sessions are over-fragmented) or too large (different intents are merged). We have added these results and analysis to Section 4.7.


Table 4. Sensitivity to the session threshold.

DatasetMetric30 Days60 Days90 Days120 Days180 Days
Movie->BookRecall@100.07560.08290.08420.08530.0816
Movie->BookNDCG@100.04990.05260.05290.05340.0521
Book->MovieRecall@100.15350.15570.15620.15700.1548
Book->MovieNDCG@100.08670.08670.08760.09150.0875
Clothing->SportsRecall@100.11080.11150.11340.11550.1150
Clothing->SportsNDCG@100.06210.05860.05750.06850.0621
Sports->ClothingRecall@100.08370.08420.08490.08540.0850
Sports->ClothingNDCG@100.04260.04460.04690.04760.0472

(ii) Why we do not use an LLM for segmentation.


We use a time-interval heuristic rather than LLM-based segmentation for three main reasons:


  • Stable Supervision. Intent-trajectory learning uses overlapping users' real target-domain sessions as supervision (Eqs. (6)–(11)). Time intervals give a deterministic, reproducible, and user-consistent segmentation criterion. LLM-based segmentation would vary with prompts, model versions, and sampling randomness, making the training target itself unstable.

  • Lower Computational Cost. Sessionization must be run offline over all overlapping users' sequences. Timestamp-based segmentation is an O(N) lightweight operation, whereas calling an LLM to decide each boundary incurs inference cost that grows linearly with data size.

  • Intent Boundaries Are Not Necessarily Semantic-Similarity Boundaries. LLM-based segmentation tends to group semantically similar interactions into one session, but real intent switches can occur between semantically similar interactions. Time intervals serve as a coarse proxy for external context changes (season, scenario, need cycle) and provide a segmentation signal independent of item semantics, preventing segmentation from being driven entirely by semantic similarity.

We sincerely thank you for your thoughtful feedback, which has helped improve our manuscript, and welcome any further questions or requests for clarification.

Reviewer_rXE3OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a Cross-Domain Sequence Synthesis (CDSS) method for cross-domain recommendation, which aims to generate diverse, plausible target-domain sequences to enhance downstream recommendation. It raises a challenge termed "Target-Intent Collapse," where similar source histories yield highly homogeneous target behaviors. To solve this, the proposed IntExpo framework comprises three cascaded modules. First, a Transferable Preference Profiler uses a domain discriminator as a reward to fine-tune an LLM via GRPO, extracting domain-agnostic profiles. Second, an Intent Generator clusters sessions into discrete intents and introduces a Collaborative Intent Distribution (CID) loss, leveraging similar users' trajectories to sample diverse intent trajectories. Third, a Behavior Generator uses LoRA to fine-tune an LLM to generate concrete sequences conditioned on the preference profile, sampled intent, and candidate items. Experimental findings demonstrate that IntExpo achieves sota performance on all 8 metrics across 4 scenarios, and case studies show IntExpo reproduces the intra-group dispersion of real target data and illustrates its ability to diverge from a single source preference into diverse, plausible intents like "camping" and "fitness."

Critical strengths

This paper observes that while users with similar source-domain preferences exhibit highly dispersed real target-domain behaviors in the representation space, the synthetic sequences generated by existing methods are highly concentrated. To address this, a novel method architecture named IntExpo is proposed. It employs a domain discriminator combined with GRPO to ensure preference transferability, utilizes discrete intent trajectories alongside a CID loss to guarantee intent diversity, and applies LoRA-based behavior generation to ensure the plausibility of specific sequences, demonstrating notable innovation.

Critical weaknesses
  1. CDR already has extensive work in areas such as cross-domain transfer and multi-domain generalization. The paper discusses and compares sample augmentation methods in isolation, resulting in a lack of depth in the scientific problem. Moreover, most baseline models used in the experiments are relatively old, failing to demonstrate its sota performance.

  2. The controllable boundaries and hallucination issues inherent in generative models are difficult to avoid, and reinforcement learning training is typically unstable. The proposed model utilizes multiple modules; however, the Transferable Preference Profiler and the Behavior Generator lack both theoretical and experimental principled analysis.

Questions for authors
  1. Why have mainstream representation-transfer-based CDR methods not encountered the target-intent collapse problem?

  2. The paper defines intent collapse solely based on the "dispersion" of representations. If a user's real intents are not divergent, does IntExpo still hold?

  3. IntExpo uses multiple large-scale LLMs (Qwen3-4B, Qwen3.5-9B) for two-stage training (GRPO+LoRA), while the baselines are mostly small-scale traditional sequential models. Can a performance comparison under fair parameter settings be provided?

  4. The paper relies on LLM generation, and the generation quality highly depends on prompt design. Although Appendix B.3 displays the prompts, are the used prompts robust?

  5. What is the failure case?

3 comments
AuthorsOpenReview ↗

Response to Reviewer rXE3

Thanks for your time and valuable comments. We hope the following responses will address your concerns.

Q1: Target-Intent Collapse Problem in Representation-Transfer-Based CDR Methods.


Thank you for raising this valuable question. The target-Intent collapse problem may also occur in representation-transfer methods. The CDR method DisCo [1] that disscused in our paper has noted that users with similar source-domain preferences may have different target-domain interests.


To examine this, we conduct an additional analysis of PTUPCDR, a CDR method based on representation transfer. We select three overlapping users from the test set as seeds and retrieve 200 nearest neighbors for each using source-domain embeddings. Within each group, we compute the mean pairwise cosine similarity between different users’ target embeddings, separately for real target data and PTUPCDR predictions.


Tabel 1. Mean within-group pairwise cosine similarity.

Transfer settingRealPTUPCDR
Movie→Book0.5840.693
Book→Movie0.6080.722
Clothing→Sports0.5620.649
Sports→Clothing0.6230.695

As shown in Table 1, PTUPCDR has higher within-group similarity than real data, suggesting that representation-transfer methods may also encounter target-Intent collaps problem.


IntExpo focuses on this issue in CDSS, where similar source histories lead to overly homogeneous target sequences. IntExpo explicitly learns a distribution over target intents to generate diverse and plausible target sequences, addressing this issue through synthetic training data.


In addition, to strengthen the comparison with recent CDR methods, we further add S²CDR (WWW 2026) [2] under the same evaluation protocol. As shown in Table 2, IntExpo outperforms S²CDR on both metrics across all four settings. We have added these results in the revised manuscript (Tabel 1 (Section 4.2) and Appendix C.2).


Table 2. Performance comparison with S²CDR.

Transfer settingS²CDR R@10S²CDR N@10IntExpo R@10IntExpo N@10
Movie→Book0.08060.04920.08530.0534
Book→Movie0.15270.08840.15700.0915
Clothing→Sports0.10260.05890.11550.0685
Sports→Clothing0.06410.03870.08540.0476
[1] DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation. AAAI 2025.
[2] S²CDR: Smoothing-Sharpening Process Model for Cross-Domain Recommendation. WWW 2026.

Q2: Does IntExpo Work When Users' Intents Are Not Diverse?


Thank you for raising this question. We respond to it from two perspectives: Methodological Analysis and Experimental Validation.


Methodological Analysis:


IntExpo learns intent diversity from observed behavior. It combines supervision from each user's own trajectory with the CID loss to fit the following target intent distribution (Appendix B.1):

Qu=11+λδZut+λ1+λPucolQ_u = \frac{1}{1+\lambda}\delta_{Z_u^t} + \frac{\lambda}{1+\lambda}P_u^{\mathrm{col}}

The first term represents the user's own observed trajectory, and the second represents the distribution of observed trajectories from similar users.


When the observed trajectories of both the user and similar users concentrate on the same trajectory Z0Z_0, this gives Qu=δZ0Q_u=\delta_{Z_0}, so the training target remains concentrated. Thus, CID learns the intent distribution supported by observed behavior and does not force intent diversity. IntExpo can therefore accommodate cases where a user's real intents are not divergent.


Experimental Validation:


To examine IntExpo's generation behavior when real target interests are concentrated, we conduct a group comparison in the Movie→Book setting. We first use a pretrained Movie-domain SASRec to encode the source histories of test users with interactions in both domains. For each user, we retrieve the top 200 most similar users to form a candidate group.


For each candidate group, we compute the average pairwise cosine similarity between users' source-domain(Movie) representations. Separately, we use a frozen Target-domain(Book) SASRec to encode their real target histories and compute the average pairwise cosine similarity between users' real target-domain representations. We then select two groups with comparable within-group similarity in Movie but different levels of concentration in Book:

  • Group A: Lower within-group similarity in real Book representations, indicating more dispersed target behavior.

  • Group B: Higher within-group similarity in real Book representations, indicating more concentrated target behavior.

IntExpo, w/o CID, and LASSO each generate one target sequence per user using only their Movie history. We encode all generated sequences with the same frozen Book-domain SASRec and compute the within-group average pairwise cosine similarity (Avg Cos). Higher values indicate more concentrated representations.


Table 3. Average pairwise cosine similarity.

MethodGroup AGroup B
Real target sequences0.3770.913
IntExpo0.4090.884
w/o CID0.5720.873
LASSO0.6240.867

As shown in Table 3, IntExpo reduces excessive concentration in Group A and maintains concentration close to the real data in Group B. These results show that IntExpo can preserve appropriate concentration when users' real intents are not divergent.

Q3: Performance Comparison Under Fair Parameter Settings.


Thank you for raising this concern about the fairness of our comparison. In our experiments, IntExpo and the LLM-based CDSS baselines LASSO and LGCD all use Qwen3.5-2B-Base as their sequence generation backbone, with the same model parameter count. Other model sizes, such as Qwen3.5-9B, are used to evaluate training and inference costs at different scales. The main performance comparison uses the shared 2B configuration.


In the sequence augmentation experiments, all three methods use the same data splits, generate one target-domain sequence per source-domain user, and use SASRec for downstream recommendation. Thus, the comparison controls for generation model size, augmentation volume, and the downstream recommender.


Table 4. Recommendation performance with the same 2B generation backbone.

MethodMovie→Book R@10Movie→Book N@10Book→Movie R@10Book→Movie N@10Clothing→Sports R@10Clothing→Sports N@10Sports→Clothing R@10Sports→Clothing N@10
LASSO0.08380.05150.15100.08700.10990.06490.07520.0440
LGCD0.08120.04860.14970.08440.10430.06160.07030.0398
IntExpo0.08530.05340.15700.09150.11550.06850.08540.0476

As shown in Table 4 (Tabel 1 in our paper), IntExpo outperforms LASSO and LGCD on all metrics across the four scenarios under these shared settings. This shows that its gains do not come from using a larger generation model.


In the revised manuscript, we have clarified the configurations used in the main experiments and the model-size in Appendix C.2.

Q4: Are The Prompts Robust?


Thank you for raising this question about prompt robustness. We use the same prompt templates across experimental settings. As shown in Appendix B.3, the prompt structure is shared across domains. We adjust only the domain names, input content, and item ID prefixes. The prompts specify the task, provide the necessary information, and define the output format. The key elements of the three modules are listed in Table 5.


Table 5. Key prompt elements of the three modules.

ModuleKey Prompt Elements
Transferable Preference ProfilerTransferable core interests and preferred attributes; omit item-specific details; JSON output with “core_interests” and “preferred_attributes”
Offline Intent ExtractionImmediate session intent; observed item evidence only; no inference of long-term preferences; JSON output with “intent_name” and “intent_description”
Behavior GeneratorTransferable preference profile; current intent; selection from the provided candidate items only; return item IDs

These requirements define what the model should do, while allowing different wording. In preliminary tests, we tried different phrasings with the same meaning. We observed relatively stable recommendation performance when the task goals, input information, and output constraints remained unchanged.

Q5: What Is The Potential Failure Case of IntExpo?


Thank you for the question. A potential limitation of IntExpo is that the candidate item set may not fully cover a user’s specific needs. Even with an appropriate target intent, the generated sequence may therefore fall short of addressing those needs. This limitation also applies to other LLM-based recommendation methods.


We identified a potential failure case in the Clothing→Sports setting. The user’s source-domain interactions included items such as a thermal cycling mask and cycling apparel, while their target-domain interactions included a bicycle headlight for night riding. IntExpo inferred “outdoor cycling” as the target intent. Although the generated sequence reflected the user’s cycling interest, the absence of bike lights in the candidate set limited its ability to address their lighting needs.


Candidate set construction has emerged as a research direction in LLM-based recommendation. For example, CORONA (SIGIR 2025)[1] filters candidates through LLM reasoning about preferences and intents, while CRAG (WWW 2025)[2] combines collaborative retrieval with context-based filtering. Our work focuses on preserving target-intent differences and generating behavior sequences; improving candidate construction is a complementary direction. We have added this discussion to Appendix C.6.


[1] Corona: A coarse-to-fine framework for graph-based recommendation with large language models, SIGIR 2025.
[2] Collaborative retrieval for large language model-based conversational recommender systems, WWW 2025.

We sincerely thank you again for your time and thoughtful feedback, which has helped us improve the quality of our manuscript. We hope our responses address your concerns and would be happy to provide any further clarification.

Reviewer_Hb8WOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cross-domain sequence synthesis, where an LLM generates target-domain interaction histories for users who are only observed in the source domain, so that a target-domain sequential recommender can be trained on more data. The authors argue that existing synthesis methods suffer from what they call target-intent collapse: users with similar source preferences receive overly similar synthetic target sequences, so other plausible intents are underrepresented. IntExpo addresses this in three stages: (i) a profiler LLM fine-tuned with discriminator-guided RL to produce transferable preference profiles, (ii) an intent generator that tokenizes target sessions into discrete intents and learns a distribution over intent trajectories, regularized toward the trajectories of profile-nearest neighbors through a CID loss, and (iii) a behavior generator that produces a session for each sampled intent from an intent-specific candidate set. Experiments on four Amazon transfer settings with a common SASRec recommender report gains over nine baselines under a matched synthesis budget, together with cold-start results, ablations, a synthesis-scale study, and t-SNE visualizations.

Critical strengths

A well-motivated decomposition with a principled diversity regularizer. Conditioning synthesis on a transferable profile and on a distribution over discrete intent trajectories (Eq. 1, section 3.2) directly targets the one-to-many structure the paper identifies. With the CID loss, the intent-generator objective has a clean interpretation as fitting a mixture of the user's own trajectory and the neighbors' trajectories, and the appendix gives a bound showing that a closely fitted distribution cannot collapse to a single trajectory when neighbors disagree (App. B.1). The ablation and the synthesis-scale study isolate the effect of this loss, and the main comparison is budget-matched, with one synthetic sequence per user and the same downstream recommender for every synthesis method.

Critical weaknesses

W1. Lack of quantitative evaluation for the headline "target-intent collapse" phenomenon. The first contribution is to identify and investigate target-intent collapse, and the abstract and section 4.5 state that IntExpo alleviates it. The evidence for both claims is qualitative: two embedding scatter plots, both on the Movie to Book setting with LASSO as the only synthesis baseline (Fig. 1, Fig. 3). Specifically, no numeric diversity or concentration statistics are reported comparing real versus synthetic sequences, IntExpo versus the two CDSS baselines, or IntExpo with versus without the CID loss. This is a little surprising, since App. B.1 already defines an intent collision probability that could be estimated from repeated samples of the intent generator. The bound derived there is never instantiated with empirical quantities. I also noticed that, at the paper's setting of λ = 1, it can at best certify that two sampled trajectories for the same user differ with probability about one half, and the fit gap it depends on is not reported. In addition, the bound concerns repeated samples for a single user, whereas the collapse described in section 1 and illustrated in Fig. 3 occurs across users, with one sequence generated per user. The downstream gains in Tab. 1 are consistent with reduced collapse but do not isolate it. I would find the paper considerably more convincing with a quantitative collapse measure on all four settings.

W2. Reporting gaps around the main comparison. Several details needed to judge Tab. 1 to 3 are missing. The tables report five-run means without standard deviations, and the caption marks paired t-tests without saying what is paired or which baseline is the reference. The LLM backbones, prompts, and tuning of LASSO and LGCD are not stated (section 4.1, App. C.2), while IntExpo uses three different Qwen models. The ablation discussion singles out removing the intent generator or the CID loss as causing "the significant decline" (section 4.3), but in Tab. 2 the weakest variant differs across settings and is often the one without the discriminator or without behavior-generator fine-tuning. The hyperparameter curves in Fig. 7 do not say whether they are on validation or test data. These points matter most on Movie to Book, where the Recall@10 gain over LASSO is under two percent and every ablated variant falls below LASSO on that metric. In the anonymous repository I found the synthesis pipeline, but I could not find the downstream training and evaluation code or the baseline implementations, even though the reproducibility statement lists evaluation scripts, so these details cannot be checked from the code either. I expect most of these points could be addressed by reporting, without new experiments.

Questions for authors

Q1. Following the protocol of Fig. 3, could the authors report a quantitative collapse or diversity measure on all four settings for real data, IntExpo, IntExpo without CID, LASSO, and LGCD? Examples would be the collision probability from App. B.1, estimated from repeated samples together with the fit gap its bound depends on, or pairwise distances between synthetic target embeddings within source neighborhoods.

Q2. For Tab. 1 to 3, could you provide standard deviations over the five runs and state what is paired in the t-test and which baseline is the reference? Which LLM backbones and hyperparameters were used for LASSO and LGCD, and how were the baselines tuned? Would it be possible to release the downstream training and evaluation scripts and the baseline configurations used for these tables?

Q3. Could you reconcile the ablation discussion in section 4.3 with Tab. 2? Please also confirm that the main results use the profiler listed in Tab. 4 (Qwen3-4B, as in the released default config), since Tab. 6 benchmarks Qwen3.5 profilers and section 3.1 cites the Qwen3.5 report.

Q4. Were λ and K (Fig. 7) chosen on validation or on test Recall@10? In the cold-start protocol (section 4.4), is IntExpo's first-item distribution restricted to the intent-specific candidate set, and over what item pool do the CDR and CDSS baselines rank?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Hb8W

Thanks for your time and constructive comments. We hope the following response will serve to address your concerns.

Q1: Quantitative Evaluation of Target-Intent Collapse Across All Four Transfer Settings.


Thank you for the suggestion. We extend the analysis in Section 4.5 (Figure. 3) to all four transfer settings, quantitatively comparing real target data, IntExpo, IntExpo w/o CID, LASSO, and LGCD.


For each setting, we follow the grouping procedure in Figure. 3: we select three held-out overlapping users as seeds and retrieve 200 nearest neighbors for each using source-domain SASRec embeddings, following Figure. 3. All methods use the same three groups and generate one target sequence per user from source histories alone. Real target histories serve only as a reference.


We encode real and generated sequences with the same frozen target-domain SASRec. We then compute the mean pairwise cosine similarity between different users’ target embeddings within each group, then average across the three groups. We use the original embeddings before t-SNE projection. Higher similarity indicates more concentrated target embeddings. The results are shown in Table 1.


Table 1. Mean within-group cosine similarity.

Transfer settingRealIntExpoIntExpo w/o CIDLASSOLGCD
Movie→Book0.5560.6310.7330.8770.841
Book→Movie0.6280.6860.7650.8890.862
Clothing→Sports0.5310.5940.6990.8430.808
Sports→Clothing0.5930.6520.7480.8710.837

Across all four settings,IntExpo’s within-group similarity is closest to that of real data, providing quantitative support for its ability to alleviate target-intent collapse, consistent with Figure. 3. We have added the evaluation protocol, results, and analysis to Appendix C.5.

Q2: Statistical Reporting and Reproducibility.


We run each method five times using five different random seeds. All methods share the same seed set, data split, and test users within each setting. We have added standard deviations to Tables 1–3. For each setting and metric, we use two-sided paired tt-tests to compare IntExpo with each baseline in Tables 1 and 3 and each ablation variant in Table 2. Run-level metric values are paired by seed, giving five pairs per test. The captions now explain the comparisons, pairing procedure, and significance markers.


LASSO and LGCD use the same LLM backbone as IntExpo’s behavior generator, Qwen3.5-2B-Base. All other hyperparameters follow the best settings reported in their original papers. We have added the configurations and parameter selection details to Appendix C.2.


In addition to the training and evaluation code for IntExpo’s sequence generation pipeline already available in the anonymous repository, we have updated the downstream training and evaluation scripts and the baseline configurations used for these tables.

Q3: Clarification of the Ablation Results and Profiler LLM Backbone.


Thank you for pointing out the imprecise wording in Section 4.3. We have revised the discussion to describe the ablation results in Table 2 more accurately and clarify how each component’s impact varies across transfer settings and metrics. We have also removed general statements such as “significant drops” after removing IG or CID.


We confirm that the main experiments used Qwen3.5-4B as the profiler. The Qwen3-4B labels in Table 4 and the previously released default configuration were incorrect. After switching the model weights, the local directory retained its old name, which was carried over into the table and configuration. We have corrected the model label in Table 4 and renamed the directory to reflect the model actually used. We apologize for the confusion.

Q4: How λ and K Were Selected and How the Candidate Pool Was Constructed for the Cold-Start Experiment.


Thank you for the valueable suggestion. We selected λ=1\lambda=1 and K=16K=16 based on validation Recall@10 and used the same values across all four transfer settings. Test results were not used for hyperparameter selection. We have clarified this in Appendix C.3.


For a fair comparison using exactly the same candidate items, we construct a shared candidate set Cu=Izucand∪{iu∗}C_u=I_{z_u}^{\mathrm{cand}}\cup\{i_u^\ast\} for each test user uu, where iu∗i_u^\ast is the true first target item. If this item is absent from the intent-specific candidate set, we add it as a regular candidate without marking it as the ground truth. IntExpo’s first-item distribution is restricted to CuC_u, and all CDR and CDSS baselines select their Top-10 items from the same CuC_u. We have added this protocol to Section 4.4.

We sincerely thank the reviewer for the thoughtful feedback, which has helped improve the quality of our paper. Please let us know if you have any further questions or concerns; we will be glad to address them.

Thank you for the response. My concerns have been adequately addressed, so I will keep my initial score.

Dear Reviewer Hb8W,

Thank you for confirming that your concerns have been addressed. Thank you for your time and effort, and please feel free to contact us if you have any further questions.

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.