Semantic-ID-based (SID) generative recommendation (GR) represents each item as a sequence of discrete tokens and predicts the next item through autoregressive SID generation. However, existing methods face two challenges. First, insufficient shared context modeling for user sequences: full-history and recent behaviors provide overlapping yet complementary evidence, but modeling them within a single sequence may dilute current cues, whereas processing them through separate streams may fail to coordinate them. Second, insufficient personalized adaptation in progressive SID decisions: although decoder states evolve with generated prefixes, existing models typically use a shared output-side transformation across recommendation instances, without explicitly selecting a transformation suited to the current user need. To address these challenges, we propose Multi-View Thinking with Adaptive Generation (MV-Thinker), a framework that Thinks Globally and Generates Personally. First, Multi-View Temporal Thinking uses view-specific thinking queries and readers over overlapping full-history and recent-history views, coordinating both sources into complementary user-conditioned contexts before SID generation. Second, Hierarchical Residual Adaptation organizes lightweight adapters into nested residual paths before a shared vocabulary projection, adaptively transforming context-bearing decoder states for SID decisions. Extensive experiments on four public datasets demonstrate that MV-Thinker consistently outperforms strong discriminative and generative recommendation baselines across multiple ranking metrics. The implementation code and datasets are available on: https://anonymous.4open.science/r/MV-Thinker
Figures & tables
Figure 1. Illustration of two interconnected challenges in SID-based generative recommendation. (a) Insufficient coordination between the full history and recent interactions may either dilute current cues or remove the historical background needed to interpret them. (b) Different recommendation needs may require different uses of contextual cues, whereas a fixed transformation follows the same rule across contexts and lacks need-guided adaptation.
Figure 2. Overview of the proposed MV-Thinker framework.
Dataset
Users
Items
Interactions
Avg. length
Instruments
24,772
9,922
206,153
8.32
Beauty
22,363
12,101
198,502
8.88
Games
50,546
16,859
452,989
8.96
Yelp
30,431
20,033
316,354
10.40
Table 1. Statistics of the processed datasets.
Dataset
Method
H@5
H@10
H@20
N@5
N@10
N@20
Dataset
Method
H@5
H@10
H@20
N@5
N@10
N@20
Instruments
Caser
0.0720
0.0856
0.1099
0.0640
0.0683
0.0744
Yelp
Caser
0.0111
0.0205
0.0317
0.0079
0.0109
0.0137
GRU4Rec
0.0782
0.0925
0.1122
0.0688
0.0734
0.0783
GRU4Rec
0.0176
0.0258
0.0384
0.0127
0.0133
0.0166
BERT4Rec
0.0816
0.1040
0.1227
0.0674
0.0745
0.0795
BERT4Rec
0.0182
0.0285
0.0451
0.0125
0.0158
0.0199
SASRec
0.0828
0.1019
0.1283
0.0710
0.0772
0.0838
SASRec
0.0202
0.0313
0.0469
0.0135
0.0171
0.0210
TIGER
0.0853
0.1047
0.1293
0.0727
0.0790
0.0852
TIGER
0.0216
0.0354
0.0531
0.0138
0.0182
0.0227
+CARE
0.0859
0.1052
0.1307
0.0731
0.0792
0.0855
+CARE
0.0247
0.0359
0.0532
0.0159
0.0201
0.0249
Table 2. Overall performance with TIGER and LETTER backbones on four datasets. “+” denotes a method applied to the preceding backbone. H@ K and N@ K denote Hit Ratio and NDCG at K , respectively. The best reported results within each backbone group are boldfaced, and shaded rows denote our method. ∗ indicates significant improvements over the strongest baseline in the same backbone group under a two-tailed t -test.
Model
K=1
K=3
K=5
K=10
K=20
TIGER
4.31
7.15
9.27
13.05
19.56
MV-Thinker
5.51
9.75
12.13
17.59
25.50
Table 3. SID catalog coverage (%) on Instruments under different recommendation budgets K .
Figure 3. Visualization of Top-20 SID coverage on Instruments. Each point represents a unique SID on a shared t-SNE map. Gray points are uncovered, while blue and red markers indicate coverage by TIGER and MV-Thinker, respectively.
Figure 4. User-dependent inference routing over HRA paths on Beauty. Rows represent sampled test users, and columns indicate residual depths. Scores are first-position log confidences, min–max normalized per user. Darker cells indicate higher relative confidence, and users are grouped by the path selected during inference.
Figure 5. Ablation study of MV-Thinker on the Instruments and Beauty datasets. “w/o MVT” and “w/o HRA” denote the variants without Multi-View Temporal Thinking and Hierarchical Residual Adaptation, respectively.
Figure 6. Component analysis of MV-Thinker on Instruments and Beauty. “Only Full” and “Only Recent” retain only the full-history and recent-history views, respectively.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7. SID coverage and repeated recommendation exposure on Instruments dataset at K=20 . (a) Cumulative SID coverage in the same fixed random user order. (b) Coverage of SIDs recommended to at least the specified number of users (logarithmic horizontal axis).
Figure 8. Thinking-token layouts compared in Table 4 . In a - b - c - d , each number denotes the thinking tokens inserted before the corresponding target SID token. Only MV-Thinker reads the full-history and recent-history views.
Layout
H@5
H@10
H@20
N@5
N@10
N@20
2-2-0-0
0.0872
0.1079
0.1314
0.0740
0.0807
0.0866
0-0-2-2
0.0862
0.1071
0.1321
0.0732
0.0799
0.0863
0-0-0-2
0.0865
0.1081
0.1314
0.0736
0.0805
0.0865
2-2-2-2
0.0858
0.1073
0.1314
0.0734
0.0803
0.0864
MV-Thinker
0.0882
0.1088
0.1325
0.0751
0.0817
0.0877
Appendix
Table 4. Performance comparison of different thinking layouts and MV-Thinker. H@ K and N@ K denote Hit Ratio and NDCG at K , respectively. The best results are boldfaced.
Case I: Recent intent shift
Case II: Fine-grained variant
Case III: Intent + variant
Target
Zildjian 5B nylon-tip drumsticks
Ernie Ball Not Even Slinky (.012–.056)
Dunlop Max-Grip nylon picks (.73 mm)
Earlier
Guitar strings Guitar straps Guitar tuners
Electric-guitar accessories Amplifiers
Broad guitar accessories
Recent
E-drum kit Drum throne Zildjian sticks × 3
Ernie Ball strings, multiple gauges
Guitar strings → Dunlop .73 mm nylon picks
Method
Top-1
Rank
Top-1
Rank
Top-1
Rank
TIGER
Guitar tuner
>20
Super Slinky (.009–.042)
6
Clip-on tuner
>20
w/o MVT
Guitar strings
>20
Regular Slinky (.010–.046)
3
Guitar-string set
>20
Appendix
Table 5. Illustrative case-study scenarios on Instruments. Histories are divided into earlier context ( gray ) and recent activity ( blue ). Top-1 recommendations are highlighted as intent-consistent but different from the target ( orange ), or inconsistent with the current activity ( red ). Rank denotes the position of the ground-truth item.
Method
H@5
H@10
H@20
N@5
N@10
N@20
TIGER
0.0853
0.1047
0.1293
0.0727
0.0790
0.0852
Single-Path 1
0.0867
0.1076
0.1316
0.0741
0.0807
0.0868
Single-Path 2
0.0870
0.1058
0.1285
0.0733
0.0793
0.0851
Single-Path 3
0.0861
0.1057
0.1307
0.0737
0.0800
0.0863
Parallel Paths
0.0876
0.1082
0.1320
0.0745
0.0812
0.0874
MV-Thinker
0.0882
0.1088
0.1325
0.0751
0.0817
0.0877
Appendix
Table 6. Recommendation performance of different path organization strategies on Instruments. Single-Path 1–3 are trained separately; Parallel Paths uses parallel paths with temporal modeling. H@ K and N@ K denote Hit Ratio and NDCG at K , respectively. The best results are boldfaced.
Generative recommendation models user histories and item catalogs to generate personalized items, lists, or queries rather than only scoring candidates.
Generative recommendation reformulates item retrieval as sequence generation, allowing a unified model to directly generate the next item from a user's interaction history. Semantic IDs further make this paradigm effective and scalable by representing each item as discrete codes, enabling knowledge sharing among semantically related items. Recent studies introduce explicit reasoning before Semantic-ID generation, helping models summarize user interests and infer possible preference transitions. However, reasoning is not inherently beneficial: Inaccurate or uninformative reasoning may mislead subsequent item generation and ultimately degrade recommendation performance. This raises a central challenge: how can a recommender select and learn effective reasoning traces and progressively evolve toward better reasoning from its own generations? In this work, we propose Evo-Rec, a three-stage framework for learning better reasoning and further enhancing it through reinforcement learning. First, we align Semantic IDs with their textual and behavioral contexts, enabling the model to understand and generate item identifiers. Second, we sample multiple candidate reasoning traces and retain those that improve the prediction of the ground-truth item, providing a stronger reasoning initialization through supervised fine-tuning. Third, we further optimize the reasoning policy through reinforcement learning with catalog-constrained item generation and ranking-aware recommendation feedback. Experiments on three Amazon Review benchmarks show that Evo-Rec consistently outperforms discriminative, generative, and reasoning-enhanced recommenders across all evaluation metrics. These results demonstrate the effectiveness of our framework in learning better reasoning for SID-based generative recommendation.
Mengdan Zhu, Yufan Zhao, Sophie Di +5
Emory University Atlanta, GA, USA · Microsoft Redmond, WA, USA · Cornell University Ithaca, NY, USA
Semantic IDs (SIDs) are now a central component of generative recommendation. Current SID-based systems assign three roles to the same token sequence. Shared prefixes are intended to organize related items, the complete SID identifies an individual item, and each generated token narrows the items that can still be returned. We systematically investigate SIDs from item encoding and SID construction to autoregressive generation and final recommendation. We examine how SID construction changes item representations and how those changes affect generation. Across three Amazon domains and eight SID constructions, SID neighborhoods recover only 32.2% of the encoder's ten nearest neighbors on average. Alternative item descriptions still retrieve the corresponding item first in 99.57% of controlled cases, yet change 38.4% of exact SIDs. These results show that SIDs retain broad organization but lose much of the encoder's fine local structure, while their exact tokens are not determined by item meaning alone. This loss becomes consequential during generation. After the final semantic token, TIGER retains only 29.9% of held-out targets that were plausible recommendations before SID filtering. Motivated by these findings, we propose Item-Supported Decoding (ISD), a lightweight inference-time method that allows a user-specific item ranking to support corresponding SID prefixes before beam search discards them. The same ranking then orders the generated items. ISD requires no additional parameters or retraining of the SID constructor or decoder. We empirically show that ISD improves NDCG@10 over the corresponding SID backbone in every evaluated setting, with relative gains of up to 31.2%. Our results show that SIDs provide useful coarse item organization, but their fine boundaries should not alone determine which items remain available during generation.
Junting Wang, Xinrui He, Yunzhe Li +1
University of Illinois Urbana-Champaign Urbana, Illinois, USA
Personalized recommendation depends on learning semantic and collaborative relations among items from online content and user interactions. Generative recommendation (GR) offers a promising paradigm by autoregressively generating discrete semantic IDs (SIDs) of items. However, randomly initialized SID token embeddings and GR models trained from scratch do not directly inherit the semantic relations inherent in items, while inter-item collaborative relations are learned largely implicitly through token-level attention and next-token supervision. To address this limitation, we propose REGER, a unified framework that explicitly incorporates inter-item relations into semantic pretraining, history encoding, and SID generation. We pretrain the GR model through generative semantic retrieval, enabling it to inherit inter-item semantic similarity by learning to generate semantic neighbors' SIDs. We combine global attention with local attention aligned with SID boundaries to capture historical context and directed transitions between adjacent items. We formalize item-level preference distillation on the SID prefix tree, deriving an exact factorization of item preference distributions into prefix conditionals and a weighted decomposition of item-level KL divergence into conditional KL terms. Guided by this decomposition, we renormalize the generator's probabilities over legal next tokens and match the resulting distributions to the teacher's conditional preferences along the ground-truth SID path, jointly optimizing the distillation objective with cross-entropy. Extensive experiments on real-world datasets validate the effectiveness of the proposed preference distillation and demonstrate that REGER consistently outperforms existing GR baselines while retaining efficient inference.
Yuxi Huang, Silin Zhou, Chenhao Wang +6
University of Electronic Science and Technology of China Chengdu, China