Organizations: School of Artificial Intelligence, Shenzhen Technology University, Shenzhen, China · City University of Hong Kong, Hong Kong, Hong Kong SAR · Shenzhen Technology University, Shenzhen, China · McGill University, Montreal, Canada · Shenzhen University, Shenzhen, China
The sequential recommendation (SR) task aims to predict the next item based on users' historical interaction sequences. Typically trained on historical data, SR models often struggle to adapt to real-time preference shifts during inference due to challenges posed by distributional divergence and parameterized constraints. Existing approaches to address this issue include test-time training, test-time augmentation, and retrieval-augmented fine-tuning. However, these methods either introduce significant computational overhead, rely on random augmentation strategies, or require a carefully designed two-stage training paradigm. In this paper, we argue that the key to effective test-time adaptation lies in achieving both effective augmentation and efficient adaptation. To this end, we propose Retrieve-then-Adapt (ReAd), a novel framework that dynamically adapts a deployed SR model to the test distribution through retrieved user preference signals. Specifically, given a trained SR model, ReAd first retrieves collaboratively similar items for a test user from a constructed collaborative memory database. A lightweight retrieval learning module then integrates these items into an informative augmentation embedding that captures both collaborative signals and prediction-refinement cues. Finally, the initial SR prediction is refined via a fusion mechanism that incorporates this embedding. Extensive experiments across five benchmark datasets demonstrate that ReAd consistently outperforms existing SR methods.
Figures & tables
Figure 1. Illustration of the training-test mismatch in sequential recommendation. The training set contains common patterns (e.g., sports or electronics sequences). However, the test input may contain items or combinations that are rare in the training data. Without retrieval (left), the frozen model produces unreliable predictions. With ReAd (right), the test sequence queries a collaborative memory, retrieves relevant patterns, and fuses them with the initial prediction to produce an adapted output. A two-part diagram contrasts frozen sequential recommendation without retrieval against ReAd. ReAd uses a rare mixed test sequence to query a collaborative memory, retrieves relevant historical patterns, and fuses them with the initial prediction.
Method
Test-time
Retrieval
Test-time
Requires
Plug-and-play
Computational
operation
enhanced
param update
retraining
overhead
TTT4Rec
✓
✗
✓
✗
✗
High
TTA
✓
✗
✗
✗
✓
Low
RaSeRec
✗
✓
✗
✓
✗
High
ReAd (Ours)
✓
✓
✗
✗
✓
Low
Table 1. Learning paradigms for sequential recommendation: test-time training, test-time augmentation, training-time retrieval, and ReAd.
Figure 2. Overview of ReAd. (a) Online inference retrieves collaborative memories and fuses them with the initial prediction. (b) Offline retrieval learning trains cross-attention with recommendation and KL-based alignment losses. Blue and red boxes denote online and offline components, respectively. A two-panel flowchart of ReAd. The online panel retrieves similar representations from collaborative memory and fuses them with the initial prediction. The offline panel trains a cross-attention retrieval module with recommendation and alignment losses.
Statistic
Office
Beauty
Sports
Home
ML-1M
#Users
4,906
22,364
35,599
66,520
6,040
#Items
2,421
12,102
18,358
28,238
3,706
#Interactions
53,258
198,502
296,337
551,682
1,000,209
Avg. Actions
10.86
8.88
8.32
8.29
165.60
Sparsity
99.55%
99.93%
99.95%
99.97%
95.53%
Table 2. Statistics of the datasets used in experiments.
Dataset
Metric
GRU4Rec
SASRec
BERT4Rec
S³-Rec
CL4SRec
CoSeRec
ICLRec
DuoRec
MCLRec
ICSRec
RaSeRec
TTA
ReAd (+SASRec)
ReAd (+DuoRec)
Office
HR@5
0.0277
0.0544
0.0376
0.0443
0.0480
0.0577
0.0564
0.0644
0.0671
0.0651
0.0669
0.0548
0.0585
0.0698
HR@10
0.0532
0.0899
0.0666
0.0658
0.0665
0.0811
0.0815
0.1011
0.1071
0.1051
0.1042
0.0896
0.0907
0.1090
HR@20
0.0985
0.1388
0.1166
0.1221
0.1121
0.1346
0.1195
0.1552
0.1648
0.1576
0.1641
0.1389
0.1427
0.1668
ND@5
0.0169
0.0369
0.0233
0.0296
0.0308
0.0394
0.0345
0.0410
0.0407
0.0416
0.0429
0.0371
0.0384
0.0456
ND@10
0.0252
0.0483
0.0326
0.0383
0.0385
0.0500
0.0439
0.0527
0.0538
0.0525
0.0549
0.0492
0.0489
0.0582
ND@20
0.0365
0.0605
0.0452
0.0523
0.0502
0.0611
0.0534
0.0663
0.0691
0.0679
0.0699
0.0625
0.0620
0.0727
Table 3. Performance on five public datasets. Bold and underlined values denote the best and second-best results; ReAd’s gains are significant under paired t-tests ( p<0.01 ).
Dataset
Metric
GRU4Rec
BERT4Rec
Orig
ReAd
Imp.
Orig
ReAd
Imp.
Office
HR@5
0.0277
0.0320
15.52%
0.0376
0.0437
16.22%
HR@10
0.0532
0.0595
11.84%
0.0666
0.0742
11.41%
ND@5
0.0169
0.0199
17.75%
0.0233
0.0271
16.31%
ND@10
0.0252
0.0288
14.29%
0.0326
0.0369
13.19%
Beauty
HR@5
0.0206
0.0225
9.22%
0.0340
0.0372
9.41%
Table 4. ReAd improvements on different backbones (all p<0.01 ).
Figure 3. Effects of the number of retrieved items K and KL-alignment coefficient λ . Two line-chart panels report recommendation performance across datasets as the number of retrieved items K and the alignment-loss weight λ vary.
Figure 4. Effect of top-ratio ρ in τtop . A line chart compares recommendation performance across datasets for different top-ratio values used in the entropy calculation.
Method
HR@10
NDCG@10
Beauty
Home
Office
Sports
Beauty
Home
Office
Sports
w/o Att
0.0786
0.0268
0.0901
0.0436
0.0413
0.0127
0.0484
0.0207
w/o KL
0.0863
0.0293
0.1074
0.0489
0.0473
0.0159
0.0571
0.0269
w/o Rec
0.0643
0.0218
0.0752
0.0345
0.0332
0.0095
0.0381
0.0198
w/o α
0.0845
0.0279
0.1044
0.0479
0.0451
0.0149
0.0536
0.0243
ReAd
0.0874
0.0298
0.1090
0.0496
0.0481
0.0163
0.0582
0.0274
Table 5. Ablation results on four Amazon datasets with DuoRec.
Dataset
Metric
SASRec
Random
Popularity
ReAd
Beauty
NDCG@10
0.0308
0.0212
0.0226
0.0367
HR@10
0.0598
0.0421
0.0453
0.0687
Office
NDCG@10
0.0483
0.0360
0.0378
0.0489
HR@10
0.0899
0.0734
0.0744
0.0907
Sports
NDCG@10
0.0184
0.0128
0.0196
0.0197
HR@10
0.0326
0.0224
0.0321
0.0339
Table 6. Retrieval-strategy comparison on four Amazon datasets (best in bold).
Figure 5. Item popularity stratified results on four Amazon datasets (NDCG@10). Each subplot shows absolute NDCG@10 (bars: SASRec in blue, ReAd in orange) and relative gain (green/red dots with dashed line) across item popularity quartiles (Q1: long-tail, Q4: popular). ReAd achieves substantial improvements on long-tail items (Q1) and moderately popular items (Q2-Q3). A single row of four charts for Beauty, Home, Office, and Sports compares SASRec and ReAd NDCG at 10 across four popularity quartiles and overlays ReAd's relative gain.
Figure 6. MovieLens case study: red dashed lines denote the test sequence and target item; blue dashed lines denote the two retrieved sequences and corresponding items. A MovieLens case study shows one test sequence and target linked by red dashed lines and two retrieved historical sequences and items linked by blue dashed lines.
Sequential Recommendation (SR) predicts users next interactions by modeling the temporal order of their historical behaviors. Existing approaches, including traditional sequential models and generative recommenders, achieve strong performance but primarily rely on explicit interactions such as clicks or purchases while overlooking item exposures. This ignorance introduces selection bias, where exposed but unclicked items are misinterpreted as disinterest, and exposure bias, where unexposed items are treated as irrelevant. Effectively addressing these biases requires distinguishing between items that were "not exposed" and those that were "not of interest", which cannot be reliably inferred from correlations in historical data. Counterfactual reasoning provides a natural solution by estimating user preferences under hypothetical exposure, and Inverse Propensity Scoring (IPS) is a common tool for such estimation. However, conventional IPS methods are static and fail to capture the sequential dependencies and temporal dynamics of user behavior. To overcome these limitations, we propose Time aware Inverse Propensity Scoring (TIPS). Unlike traditional static IPS, TIPS effectively accounts for sequential dependencies and temporal dynamics, thereby capturing user preferences more accurately. Extensive experiments show that TIPS consistently enhances recommendation performance as a plug-in for various sequential recommenders. Our code will be publicly available upon acceptance.
Sirui Huang, Jing Long, Qian Li +2
Hong Kong Polytechinic University Hong Kong SAR, China · University of Technology Sydney Sydney, Australia · Curtin University Perth, Australia +1
Sequential recommendation models, while powerful, have limited flexibility in responding to immediate user requests, making it difficult to adapt their recommendations to the user's timely interests. Unfortunately, existing user request adaptation methods often incur high computational overhead due to either 1) retraining the entire backbone network or 2) leveraging the inference ability of large language models (a.k.a. prompt engineering), limiting their applicability in large-scale recommendation services. This paper presents Control via Request-Aware Masking for Editing Recommenders (CRAMER), a framework that takes users' natural-language requests to immediately change sequential recommendation models' behavior. Specifically, inspired by the model control theory, CRAMER treats user requests as control signals to modulate frozen backbone parameters through masking, achieving instant adaptation to diverse requests while avoiding costly retraining. Experiments on multiple large-scale benchmark datasets show that CRAMER outperforms four state-of-the-art request-aware baselines across multiple recommendation metrics while achieving minimal overhead. Moreover, the proposed framework exhibits enhanced controllability and cross-domain adaptability, establishing a new paradigm for request-aware sequential recommendation.
Zhiyuan Julian Su, Naihe Feng, Zhen Luther Qin +1
Gaoling School of Artificial Intelligence, Renmin University of China · Faculty of Computer Science, Dalhousie University
Sequential recommender systems typically infer user preferences through single-pass encoding of interaction histories without iterative refinement, relying on increasingly deep architectures to capture complex patterns. In this work, we revisit sequential recommendation from a recursive inference perspective: can user preferences be modeled as a persistent latent state that is recursively refined? We propose RecRec (Recursive Recommendation), a lightweight model that maintains a compact latent state and updates it through a shared recursive module conditioned on interaction evidence. Unlike prior recursive models, RecRec introduces an evidence-anchored correction mechanism that stabilizes refinement by grounding each update in the original interaction context, preventing semantic drift during deep recursive reasoning. Experiments on three benchmark datasets under standard evaluation protocols show that RecRec matches or outperforms state-of-the-art sequential, graph-based, and reasoning-enhanced recommenders while using only 3.9M to 14M parameters. Ablation studies demonstrate that both recursive refinement and the evidence-anchored correction gate contribute significantly to performance, highlighting the effectiveness of recursive latent inference as a scalable alternative to deeper or language-based architectures. Code is available at https://anonymous.4open.science/r/RecRec-6B67/README.md.