acceptodds
Under review as a conference paper at ICLR 2027

LIFT: A Lifecycle-aware Interaction Factorization Transformer for Unified Retrieval and Ranking

Abstract

Cascaded recommender systems use the same user history for retrieval and ranking, while the two stages operate with different information available at different points of an interaction. We propose the Lifecycle-aware Interaction Factorization Transformer (LIFT), which decomposes each interaction into ordered Request, Item, Context, and Action states and models them as a causal sequence. Retrieval reads the Request state, while ranking reads the Context state, allowing both tasks to share history modeling while preserving stage-specific information. LIFT instantiates this representation with Role-Conditioned Attention and a lightweight Pre-LN Bias. On ML-20M and Taobao, LIFT achieves the highest Joint Score among the evaluated joint models, improving over the strongest baselines by 4.9% and 3.6%, respectively. Loss-weight sweeps show favorable retrieval–ranking trade-offs, while ablations and scaling analyses further examine lifecycle sequence construction, model components, and capacity settings.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Unified Retrieval and Ranking, Sequence Modeling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rPGnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes LIFT, a lifecycle-aware interaction factorization framework that unifies retrieval and ranking in recommender systems. Each interaction is decomposed into ordered Request, Item, Context, and Action tokens, serialized as a causal sequence so that retrieval reads a Request state (pre-candidate) and ranking reads a Context state (post-candidate, pre-action). A shared decoder with Role-Conditioned Attention (shared Q/K, role-specific V/O) and a lightweight Pre-LN role bias enables shared historical modeling while preserving stage-specific information. On ML-20M and Taobao, LIFT achieves the best Joint Score among joint baselines, shows favorable retrieval–ranking Pareto trade-offs, and is supported by ablations on lifecycle design, model components, and scaling analyses.

Critical strengths
  1. The lifecycle factorization (R→I→C→A) is a clear and principled way to align what information is available at different stages, and to define stage-aligned readouts for retrieval (Request) and ranking (Context).
  2. Role-Conditioned Attention (shared Q/K; role-specific V/O) is a neat parameter-sharing scheme that matches the intent: common dependency matching across roles while allowing role-specific transformations.
  3. The Pre-LN role bias is a simple, low-overhead conditioning mechanism that integrates well with a Pre-LN Transformer.
Critical weaknesses
  1. The framework increases sequence length by 4× per interaction; while training/VRAM costs are analyzed, serving-time latency and throughput under realistic candidate counts are not reported in detail.
Questions for authors
  1. How are ranking negatives built during training, and are the evaluation ranking candidates exactly the top-K retrieved by LIFT's retrieval head, or a fixed pool per query?
2 comments
AuthorsOpenReview ↗

A-rPGn-1: Serving efficiency and cache overhead (W1).

We thank the reviewer for recognizing the merits of lifecycle factorization, Role-Conditioned Attention, and Pre-LN Bias. We appreciate the constructive questions regarding serving efficiency and the ranking training and evaluation protocols. We provide further analysis and clarifications below.

A-rPGn-1: Serving efficiency and cache overhead (W1).

We agree that LIFT's finer-grained lifecycle modeling introduces additional serving-time computation and cache overhead, making the quality–efficiency trade-off an important consideration.

Under matched model widths, depths, and cache precision, LIFT requires 4×4\times the historical KV-cache capacity of UniPinRec, while its ranking FFN computation and candidate-history attention computation increase to 2×2\times and 8×8\times, respectively.

However, comparisons at the same model width do not fully characterize the quality–efficiency trade-off. Figure 4(b) shows that LIFT with a hidden dimension of only 64 achieves a Joint Score of .2648, outperforming UniPinRec with a hidden dimension of 256 (.2582).

Under these configurations, the theoretical serving costs are summarized below:

MetricUniPinRec (d=256d=256)LIFT (d=64d=64)
Joint Score.2582.2648
Parameters18.23M13.88M
Historical KV cache1×1\times1×1\times
Ranking FFN FLOPs1×1\times0.125×0.125\times
Candidate-history attention FLOPs1×1\times2×2\times

LIFT requires the same historical KV-cache capacity as UniPinRec, while its ranking FFN computation is reduced to 12.5% and its candidate-history attention computation is 2×2\times that of UniPinRec.

These results suggest that LIFT's more effective lifecycle representation enables better recommendation quality with a smaller model width, thereby mitigating the serving overhead introduced by finer-grained sequence modeling.

Please refer to A-tLCs-5 for a detailed analysis of computational complexity, cache overhead, and the quality–efficiency trade-off. We emphasize that these results characterize theoretical model-level costs rather than empirically measured end-to-end latency or throughput in a production environment. We will include this analysis and clarify this limitation in the revised manuscript.

AuthorsOpenReview ↗

A-rPGn-2: Ranking negative construction and evaluation protocol (Q1).

A-rPGn-2: Ranking negative construction and evaluation protocol (Q1).

Thank you for this question. We clarify the general formulation of the ranking loss, the training supervision, and the ranking evaluation protocol.

Ranking negatives during training. The ranking loss in Eq. (3) adopts a general formulation:

= \ell_{\mathrm{rank}} \left( \hat{\mathbf{y}}_{u,\leq T}^{\mathrm{rank}}, \mathbf{y}_{u,\leq T}^{\mathrm{rank}}, \mathcal{N}^{\mathrm{rank}} \right),$$ where $\mathcal{N}^{\mathrm{rank}}$ denotes an optional set of additional negative samples or candidate-sampling information, allowing the framework to accommodate different ranking objectives and negative sampling strategies. In our experiments, we adopt the standard binary cross-entropy (BCE) loss for ranking, with supervision labels constructed directly from observed interactions and their ground-truth feedback: - **ML-20M:** Interactions with ratings greater than 4 are treated as positive samples, while those with ratings of 4 or below are treated as negative samples. - **Taobao:** Purchases are treated as positive samples, while other observed behavior types (page views, favorites, and add-to-cart events) are treated as negative samples. Since BCE directly computes the loss from the predicted score and binary label of each observed interaction, it does not require constructing additional negatives for contrastive or pairwise learning. Therefore, $$\mathcal{N}^{\mathrm{rank}}=\varnothing$$ in our experimental setting. Specifically, the ranking loss is instantiated as: $$\mathcal{L}_{\mathrm{rank}} = -\frac{1}{|\mathcal{S}|} \sum_{i\in\mathcal{S}} \left[ y_i\log p_i + (1-y_i)\log(1-p_i) \right],$$ where $\mathcal{S}$ denotes the set of observed interactions used for ranking supervision, $y_i\in\{0,1\}$ is the ground-truth feedback label, and $p_i$ is the predicted probability of positive feedback. This does not imply the absence of negative ranking examples; rather, negative supervision is already provided by the binary labels derived from observed feedback, without requiring additional negative sampling. In contrast, retrieval is trained using InfoNCE with in-batch negatives and LogQ correction. Thus, the two tasks share historical modeling while retaining task-specific supervision and objectives. **Independent ranking evaluation.** The ranking metrics (AUC and GAUC) in Table 2 are computed on held-out observed interactions and their ground-truth feedback. All models are evaluated on the same test interactions and labels, rather than on the top-$K$ candidates produced by their respective retrieval modules. This evaluation primarily measures the ranking model's ability to predict user feedback. **End-to-end cascade evaluation.** To further assess the effectiveness of the complete recommendation pipeline, we conduct additional end-to-end cascade experiments. Unlike independent ranking evaluation, the ranking candidates in these experiments are exactly the top-$K$ items retrieved by each model's own retrieval module. We evaluate all positive interactions in the ML-20M test set and report Recall@10 and NDCG@10 of the final recommendation lists across 10 candidate set sizes ($K=10$–$100$). LIFT achieves the best average performance across all evaluated candidate sizes. For example, at $K=100$, the results are: | Model | Cascade R@10 | Cascade N@10 | |---|---:|---:| | UniPinRec | .0625 | .0328 | | HSTU | .0989 | .0473 | | LIFT | **.1147** | **.0596** | LIFT improves final-list Recall@10 and NDCG@10 over the strongest baseline, HSTU, by 16.05% and 25.93%, respectively. Please refer to A-tLCs-1 for the complete cascade evaluation results. We will clarify the definition of the negative sample set in Eq. (3), as well as the distinctions between training supervision, independent ranking evaluation, and end-to-end cascade evaluation in the revised manuscript.
Reviewer_cGahOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LIFT serializes each interaction into four ordered tokens (Request, Item, Context, Action) and models them in a single causal Transformer. Retrieval reads the Request state, before the candidate is known. Ranking reads the Context state, after the candidate and its context are known but before the action. Role-Conditioned Attention shares the Q/K projections (with role embeddings added) but uses a source-role-specific V and a receiver-role-specific O. A role-specific bias is also added before the FFN LayerNorm. RoPE is indexed by interaction, so the four tokens of one interaction share a position. Training is joint, with InfoNCE + LogQ for retrieval and BCE for ranking. On ML-20M and Taobao, LIFT gets the best "Joint Score" √(NDCG@100·GAUC), +4.9% and +3.6% over the best of SharedBottom, MMoE, PLE, HSTU and UniPinRec. Across a λ_ret sweep on ML-20M its operating points form the empirical Pareto front. Ablations cover token layouts, the two components, and capacity and sequence scaling.

Critical strengths

LIFT presents a simple, practical way to unify retrieval and ranking by explicitly modeling their different information-availability stages within one causal sequence. The shared history encoder also offers a natural KV-cache reuse mechanism.

The experimental setup is reasonably thorough, with chronological splits, validation-based model selection, three seeds, and both token-layout ablations and retrieval–ranking trade-off analyses.

Critical weaknesses
  1. Baseline fairness and attribution remain unclear. The UniPinRec results in Table 2 exactly match the [RICA] ablation in Table 4, but the paper does not explain their relationship. Moreover, layouts with an explicit current-request token recover most of LIFT's retrieval improvement, suggesting that access to request-time features may explain a substantial part of the gain. Without feature-matched baselines, the contribution of the proposed architecture remains unclear.

  2. The evidence for real-world effectiveness is limited. ML-20M and Taobao do not provide the explicit request, exposure, and candidate-stage logs needed to evaluate a production retrieval–ranking cascade. The paper also lacks an end-to-end evaluation of retrieving candidates and then ranking them, and reports no variance despite relatively small improvements on some metrics. The novelty of the role-conditioned attention design is also somewhat incremental without comparisons to simpler alternatives.

Questions for authors
  1. Is UniPinRec in Table 2 the same model as [RICA] in Table 4? Which features does each baseline see?
  2. What do HSTU and UniPinRec get with the current-request features added to the retrieval query?
  3. What is the standard deviation over the three seeds for Tables 2–4?
  4. Can you report an end-to-end cascade evaluation following Eq. (4)?
  5. Does "Basic Transformer" in Table 3 include role/type embeddings? How does a role-specific FFN compare?
  6. Why is LIFT faster than UniPinRec with blank tokens at the same sequence length?
  7. ML-20M ratings often share a timestamp (bulk rating sessions). How is order within a timestamp decided? Does "time since last = 0" leak session membership into the Request token?
  8. How is the performance when serve online, is there an AB test?
12 comments
AuthorsOpenReview ↗

A-cGah-1: Clarification of the [RICA] representation (W1-1 & Q1-1). & A-cGah-2: Feature availability across baselines (Q1-2).

We thank the reviewer for the detailed and constructive comments. We appreciate the recognition of our lifecycle-aware unified modeling framework and its practical motivation. To address the concerns regarding baseline fairness, contribution attribution, real-world applicability, and experimental reliability, we have conducted additional controlled experiments and provide further clarifications below.

A-cGah-1: Clarification of the [RICA] representation (W1-1 & Q1-1).

Yes, [RICA] in Table 4 corresponds to UniPinRec in Table 2. Since [RICA] merges all information from an interaction into a single token, we adopt the UniPinRec-style prediction mechanism: retrieval reads the preceding complete interaction state, [RICA]t−1[\mathrm{RICA}]_{t-1}, to prevent target-item leakage, whereas ranking reads the current interaction state, [RICA]t[\mathrm{RICA}]_t, with its ground-truth Action masked to prevent target-feedback leakage.

We will explicitly clarify this implementation detail in the footnote of Table 4.

We will also clarify the distinction between [RIC][A] in Table 4 and HSTU in Table 2. While HSTU in Table 2 retains the attention architecture proposed in its original paper, [RIC][A] in Table 4 uses the same Transformer backbone as LIFT. This controlled design minimizes the influence of encoder architecture, allowing us to more directly isolate the contribution of lifecycle sequence organization.

A-cGah-2: Feature availability across baselines (Q1-2).

Thank you for this question. The following table summarizes the features available to each joint recommendation model in Table 2 during retrieval and ranking. Each model retains its native feature input interface while following the same information-availability constraints. The specific feature definitions and their lifecycle assignments for each dataset are provided in Appendix A, Table 5.

ModelRetrieval QueryRanking
SharedBottomRt,I<t,A<tR_t,I_{<t},A_{<t}Rt,I≤t,A<t,CtR_t,I_{\leq t},A_{<t},C_t
MMoERt,I<t,A<tR_t,I_{<t},A_{<t}Rt,I≤t,A<t,CtR_t,I_{\leq t},A_{<t},C_t
PLERt,I<t,A<tR_t,I_{<t},A_{<t}Rt,I≤t,A<t,CtR_t,I_{\leq t},A_{<t},C_t
HSTU(R,I,C,A)<t(R,I,C,A)_{<t}(R,I,C,A)<t,Rt,It,Ct(R,I,C,A)_{<t},R_t,I_t,C_t
UniPinRec / [RICA](R,I,C,A)<t(R,I,C,A)_{<t}(R,I,C,A)<t,Rt,It,Ct(R,I,C,A)_{<t},R_t,I_t,C_t
LIFT(R,I,C,A)<t,Rt(R,I,C,A)_{<t},R_t(R,I,C,A)<t,Rt,It,Ct(R,I,C,A)_{<t},R_t,I_t,C_t

Here, the subscript <t<t denotes historical interactions, while tt denotes the current request or candidate item. For SharedBottom, MMoE, and PLE, historical information is provided through aggregated features rather than an explicit interaction sequence.

Notably, HSTU, UniPinRec and LIFT have access to the same categories of input information during ranking, whereas LIFT additionally incorporates current request information during retrieval through an explicit Request token.

All models respect causal information availability: retrieval queries do not access the current target item or its candidate-dependent context, and ranking does not access the current ground-truth Action.

We will include this table in the revised manuscript to clarify the feature inputs and information availability of each baseline.

AuthorsOpenReview ↗

A-cGah-3: Controlling for current Request features (W1-2 & Q2).

A-cGah-3: Controlling for current Request features (W1-2 & Q2).

Thank you for this observation. We agree that access to current Request features may explain part of LIFT's retrieval gains. To further disentangle the contributions of feature availability and architectural design, we incorporate the same current Request features used by LIFT into the retrieval queries of HSTU and UniPinRec, while preserving their original history modeling architectures.

We consider two fusion strategies:

  • Late Fusion: Concatenates the historical representation and the current Request representation, followed by an MLP to generate the retrieval query.
  • Residual Fusion: Incorporates the current Request representation into the historical representation through a zero-initialized learnable residual coefficient, allowing the model to start from its original retrieval representation and adaptively learn the contribution of Request information.

The following table reports the results on ML-20M averaged over three random seeds:

ModelR@100N@100AUCGAUCJoint
HSTU.3157.0910.8368.7410.2597
+ Late Fusion.3133.0877.8340.7370.2542
+ Residual Fusion.3262.0930.8391.7449.2633
UniPinRec.3152.0894.8442.7460.2582
+ Late Fusion.3105.0873.8429.7451.2551
+ Residual Fusion.3123.0901.8423.7457.2592
LIFT.3361.0981.8496.7565.2724

While Late Fusion does not improve either baseline, Residual Fusion improves the Joint Score, particularly for HSTU.

We agree with the reviewer's observation that explicitly incorporating the current Request accounts for a substantial portion of LIFT's retrieval improvement. This is, in fact, one of the key motivations behind our lifecycle factorization. Conventional interaction-level sequence modeling merges the information of an interaction into a single token, so retrieval typically reads the preceding complete interaction state and cannot naturally incorporate the information already available at the current request.

By explicitly introducing a Request state, LIFT allows current-request information to interact with historical interactions throughout the Transformer layers, while the lifecycle ordering prevents access to the current candidate item and its feedback.

Our controlled experiments further show that although current-request information can improve baseline performance, LIFT still achieves approximately 3.5% higher Joint Score than the strongest Request-augmented baseline.

Moreover, [R][IC][A] in Table 4 already incorporates an explicit current Request and recovers most of LIFT's retrieval gains. However, its GAUC and Joint Score are .7390 and .2652, respectively, compared with .7565 and .2724 for the complete lifecycle factorization. This suggests that an explicit Request is important for retrieval, while further separating Item and Context improves joint retrieval–ranking modeling.

Therefore, LIFT's contribution is not merely introducing additional Request features, but organizing interaction information according to its lifecycle, enabling different recommendation stages to effectively exploit their available information within a unified model.

AuthorsOpenReview ↗

A-cGah-4: Effectiveness in real-world recommendation scenarios (W2-1 & Q8).

A-cGah-4: Effectiveness in real-world recommendation scenarios (W2-1 & Q8).

Thank you for highlighting the importance of evaluation in real-world recommendation scenarios. We agree that ML-20M and Taobao lack explicit request and candidate exposure logs, and that these datasets alone cannot fully establish LIFT's effectiveness in practical recommendation settings.

To address this concern, we conduct additional experiments on the Trivago hotel recommendation dataset from the ACM RecSys Challenge 2019, which contains real user requests, request contexts, candidate impression lists, and click feedback.

After extracting clickout events, the dataset contains approximately 1.58M recommendation requests and 36.36M candidate impressions, with an average of 22.99 candidates per request. To ensure sufficient historical interactions for sequential modeling, we further apply a minimum user interaction sequence length of five. All models are trained and evaluated on the same filtered dataset under a consistent experimental protocol.

In this setting, candidates associated with the same request share the Request state and historical representations while retaining individual Item and Context states. Retrieval reads the shared Request representation, whereas ranking reads each candidate's Context representation and uses its corresponding click feedback for supervision.

The following table reports the results averaged over three random seeds:

ModelR@100N@100AUCGAUCMRRJoint
HSTU.2493.0887.7159.7034.3729.2498
UniPinRec.2572.0933.7008.6965.3613.2550
LIFT.2778.1007.7292.7122.3892.2678

LIFT achieves the best performance across all metrics, improving Joint Score by 5.02% and MRR over the logged impression lists by 4.37% relative to the strongest baseline for each metric. These results further support the applicability of lifecycle modeling to real requests with multiple candidate exposures.

In addition, we conduct end-to-end retrieval–ranking cascade experiments (see A-tLCs-1), demonstrating that LIFT consistently improves the quality of the final recommendation lists across different candidate set sizes, rather than merely improving independently evaluated retrieval and ranking metrics.

Regarding online performance, we currently have no online A/B test results for LIFT to report. We acknowledge that offline evaluation on real-world logs cannot fully substitute for online validation in a production environment.

For serving efficiency, please refer to A-tLCs-5, where we analyze historical KV-cache overhead, online computational complexity, and the quality–efficiency trade-off across different model widths. These analyses characterize theoretical model-level computation and cache costs rather than empirically measured online end-to-end latency or throughput.

We will incorporate these additional experiments and efficiency analyses into the revised manuscript and clarify the scope and limitations of our evaluation.

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

To address the differences in information availability between retrieval and ranking in cascaded recommender systems, this paper proposes a unified framework called the Lifecycle-aware Interaction Factorization Transformer (LIFT). The method decomposes each interaction into ordered Request, Item, Context, and Action states, enabling retrieval and ranking to share history modeling while reading stage-specific representations, supported by role-conditioned attention and a lightweight Pre-LN bias. Experimental results on ML-20M and Taobao demonstrate that LIFT improves joint retrieval–ranking performance and achieves favorable trade-offs between the two tasks.

Critical strengths

1.The motivation is clear: retrieval and ranking operate with different available information. Reading representations from distinct lifecycle positions provides a clear and interpretable way to address this difference.

2.Experiments on lifecycle factorization, alternative sequence layouts, blank-token controls, and component ablations collectively provide targeted empirical support for the core design.

3.The method achieves consistent improvements in joint performance on two public datasets, while sensitivity analyses of the loss weights demonstrate favorable retrieval–ranking trade-offs.

Critical weaknesses

1.The evaluation primarily relies on separate retrieval and ranking metrics and their geometric mean, without assessing the final ranked lists produced by reranking the retrieved candidates. Consequently, it does not fully establish the benefits of the complete retrieval-ranking cascade.

2.The lifecycle states used in the experiments are primarily constructed from public behavioral logs and historical statistics. Their applicability to explicit requests, multiple candidate exposures, and complex request contexts remains insufficiently validated.

3.The contribution of role-conditioned attention has not been fully disentangled from the additional model capacity. For example, Table 3 demonstrates that this component is useful under the current setup, but lacks comparisons with parameter-matched standard Transformers or generic adapters. Therefore, it remains unclear whether the gains stem specifically from the role-based structure.

4.The paper reports mean results over three random seeds but does not provide standard deviations or confidence intervals. This limits the assessment of the stability of the relatively small component-level gains in Table 3. Several important implementation details are also missing, including the filtering rules for repeated target items, the batch size, and the configuration used for efficiency measurements.

Questions for authors

1.Could the authors report end-to-end cascade results by first retrieving the top-KK candidates and then reranking them, and compare the quality of the final ranked lists across different values of KK?

2.Please compare role-conditioned attention with parameter-matched standard attention or adapter-based models and report the performance differences in the key ablation experiments.

3.Could the lifecycle design be evaluated on data containing real requests and multiple exposures per request? If this is not feasible, a controlled simulation experiment would be helpful.

4.How does the proposed method compare with the original UniPinRec in terms of end-to-end inference latency, throughput, and cache overhead?

5.Please provide additional implementation details, including the filtering rules for repeated target items, the batch size, and the configuration used for efficiency measurements, as well as standard deviations or confidence intervals for the results in Table 3.

8 comments
AuthorsOpenReview ↗

A-tLCs-1: End-to-end cascade evaluation (W1 & Q1).

We thank the reviewer for the constructive comments and for recognizing the merits of our lifecycle-aware interaction factorization. We have conducted additional experiments to address the concerns regarding end-to-end cascade performance, real requests with multiple exposures, parameter-matched comparisons, and experimental stability. We also provide further implementation details and an analysis of serving-time computation and cache overhead.

A-tLCs-1: End-to-end cascade evaluation (W1 & Q1).

Thank you for suggesting an evaluation of the final reranked lists. We agree that this provides a more comprehensive assessment of the complete retrieval–ranking pipeline.

We therefore conduct additional end-to-end cascade experiments on ML-20M following Eq. (4). We use all positive interactions in the full test set as evaluation queries. For each test request, we first retrieve the top-KK candidates from the item catalog using each model's retrieval module, rerank them using its corresponding ranking module, and evaluate the final top-10 recommendation list using Recall@10 and NDCG@10. All models are evaluated on the same test queries under an identical evaluation protocol.

The following table reports the results across 10 candidate set sizes, averaged over three random seeds (Recall@10 / NDCG@10).

KKUniPinRecHSTULIFT
10.1131 / .0600.1224 / .0691.1282 / .0751
20.1156 / .0572.1389 / .0712.1478 / .0793
30.1078 / .0539.1383 / .0698.1470 / .0769
40.0966 / .0487.1319 / .0657.1430 / .0741
50.0886 / .0446.1251 / .0617.1398 / .0730
60.0816 / .0415.1190 / .0578.1335 / .0695
70.0756 / .0389.1135 / .0548.1279 / .0666
80.0707 / .0367.1086 / .0522.1232 / .0640
90.0664 / .0347.1034 / .0495.1194 / .0619
100.0625 / .0328.0989 / .0473.1147 / .0596

LIFT achieves the best final-list Recall@10 and NDCG@10 across all 10 evaluated candidate set sizes (K=10K=10–100100). At K=20K=20, LIFT outperforms the strongest baseline, HSTU, by 6.40% and 11.27%, respectively. At K=100K=100, these improvements further increase to 16.05% and 25.93%.

Notably, increasing KK improves retrieval coverage but also makes reranking more challenging, as the ranking module must identify the target item among more candidates. LIFT exhibits greater robustness to this challenge.

For example, at K=100K=100, LIFT and HSTU achieve relatively similar retrieval Recall@100 (.4144 vs. .4062). However, conditional on successfully retrieving the target item, LIFT retains it in the final top-10 list with a probability of .2768, compared with .2435 for HSTU. This suggests that LIFT's cascade performance advantage arises not merely from improved retrieval coverage but also from more effective reranking within their respective retrieved candidate sets.

These results demonstrate that LIFT's improvements in independently evaluated retrieval and ranking metrics translate into higher-quality final recommendation lists in the complete cascade, further supporting the effectiveness of lifecycle-aware unified modeling.

We will include the full cascade evaluation results and the corresponding evaluation protocol in the revised manuscript.

AuthorsOpenReview ↗

A-tLCs-2: Evaluation on real requests with multiple exposures (W2 & Q3).

A-tLCs-2: Evaluation on real requests with multiple exposures (W2 & Q3).

Thank you for this suggestion. To further validate the applicability of lifecycle modeling to real requests with multiple candidate exposures, we conduct additional experiments on the Trivago hotel recommendation dataset from the ACM RecSys Challenge 2019. This dataset provides real user search behaviors, request contexts, candidate impression lists, and click feedback, including rich contextual information such as city, device, platform, filters, and displayed candidate prices.

After extracting clickout events, the dataset contains approximately 1.58M recommendation requests and 36.36M candidate impressions, with an average of 22.99 candidates per request. To ensure sufficient historical interactions for sequential modeling, we further apply a minimum user interaction sequence length of five. All models are trained and evaluated on the same filtered dataset under a consistent experimental protocol.

LIFT naturally extends to this one-to-many setting through the attention mask design described in the appendix. Candidates associated with the same request share the Request state and historical representations, while retaining their individual Item and Context states. Retrieval reads the shared Request representation, whereas ranking reads each candidate's Context representation, using its click feedback as Action supervision. Thus, lifecycle factorization is not restricted to one-to-one request–item interactions.

The following table reports the results averaged over three random seeds:

ModelR@100N@100AUCGAUCMRRJoint
HSTU.2493.0887.7159.7034.3729.2498
UniPinRec.2572.0933.7008.6965.3613.2550
LIFT.2778.1007.7292.7122.3892.2678

LIFT achieves the best performance across all metrics, improving Joint Score by 5.02% and MRR over the logged impression lists by 4.37% relative to the strongest baseline for each metric.

These results demonstrate that lifecycle modeling is applicable not only to interaction sequences constructed from behavioral logs, but also to recommendation scenarios involving real requests, rich contextual information, and multiple candidate exposures. We will include these experiments and the corresponding request-level lifecycle modeling details in the revised manuscript.

AuthorsOpenReview ↗

A-tLCs-3: Parameter-matched comparisons (W3 & Q2).

A-tLCs-3: Parameter-matched comparisons (W3 & Q2).

Thank you for this suggestion. To disentangle the contribution of additional model capacity from that of role-conditioned attention, we conduct three additional parameter-matched comparisons on ML-20M:

  • LIFT-WideV: Uses standard shared attention and matches the parameter count by increasing the output dimension of the Value projection and the input dimension of the Output projection.
  • LIFT-4W: Retains four sets of Value and Output projections but applies all four to every token and sums their outputs, removing role-dependent parameter selection while preserving the intermediate representation dimension.
  • LIFT-Adapter: Uses standard shared attention with role-specific adapters, providing an alternative implementation of role-conditioned transformations.

All four models have exactly the same parameter count: 18,716,793 total parameters, including 6,278,393 dense parameters. They also use the same lifecycle sequence representation, training budget, and evaluation protocol. The following table reports the mean and standard deviation over three random seeds:

ModelDense ParamsN@100AUCGAUCJoint
LIFT-WideV6.278M.0962±.0008.8463±.0011.7542±.0021.2693
LIFT-4W6.278M.0972±.0014.8463±.0010.7544±.0024.2707
LIFT-Adapter6.278M.0983±.0016.8485±.0011.7554±.0014.2724
LIFT6.278M.0981±.0003.8496±.0019.7565±.0021.2724

Under matched parameter budgets, LIFT improves Joint Score over WideV and 4W by 1.15% and 0.63%, respectively. These results suggest that simply increasing shared attention capacity without distinguishing lifecycle roles does not fully reproduce the benefits of role-conditioned attention.

Notably, LIFT-Adapter achieves the same reported Joint Score as LIFT, further supporting our motivation that organizing parameters according to lifecycle roles is beneficial, rather than the gains being restricted to a particular implementation of role conditioning.

LIFT adopts role-specific Value and Output projections as a direct and computationally efficient implementation. Each token activates only its corresponding role-specific projection matrices. Consequently, compared with standard shared attention of the same representation width, LIFT increases parameter capacity without increasing per-token V/O projection FLOPs. In contrast, WideV, 4W, and Adapter require additional projection computations.

Overall, these results suggest that LIFT's improvements arise not merely from increased parameter capacity but also from lifecycle-aware parameter organization. We will include these parameter-matched comparisons and the corresponding analysis in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.