Generative recommendation aims to predict item identifiers from user interaction histories, but explicit control over the properties of recommended items remains underexplored. A key challenge is to connect human-readable concepts to internal features whose interventions effectively change the recommended items. To address this, we propose Steerable Generative Recommender (SteerGR), a framework that uses sparse autoencoders for concept-level control in game discovery. SteerGR proceeds through three stages: (i) Concept Definition and Grounding for defining play concepts with an LLM; (ii) Feature Discovery and Selection for discovering candidate features by their concept associations and selecting one by its steering effect; and (iii) Sparse Feature Intervention for amplifying or suppressing selected features to steer recommendations without retraining the recommender. Extensive experiments with two industrial generative recommendation models on a real-world game dataset show that the selected features successfully steer recommendations toward or away from target concepts. We believe this work takes a first step to extend SAE-based steering from LLMs to industrial generative recommendation.
Figures & tables
Figure 1: Motivation for controllable recommendation. (a) Without explicit control, a recommender may return the fixed results regardless of the user’s goals. (b) Ours can steer recommendations toward discovery goals, e.g. , teaming up with friends or trying a scary experience.
Figure 2: Overall architecture of SteerGR. We first train a GR model, freeze its parameters, and train an SAE on its hidden states. (a) Play concepts are defined and scored over games and users. (b) Features are discovered by concept association and selected by steering effect. (c) The selected feature is amplified or suppressed at inference.
Feature selection
Cohen’s d
Target concept share (%)
Amplify ( ↑ )
Suppress ( ↓ )
Atomic ID
Original
–
5.75
5.75
Random
−0.03
5.76
5.74
Top- d
0.55
6.59
5.02
SteerGR
0.38
7.95
4.27
Table 1: Effect of feature selection on concept association (Cohen’s d ) and target concept share. ‘Original’ denotes no intervention, and ‘Top- d ’ denotes an ablation without intervention-based selection.
Recent plays (oldest → newest)
Original top-5
Amplifying collaborate with others
Amplifying defend against threats
Game, concept
Game, concept
Game, concept
Before
Game, concept
Before
1
Base-building survival defend
Rhythm platformer tactical
Two-player co-op course collaborate
>10
Hidden-threat survival collect
2
2
Territory-capture game defend
Hidden-threat survival collect
Rhythm platformer tactical
1
Rhythm platformer tactical
1
3
Stealth escape solve
Battle royale brawler compete
Hidden-threat survival collect
2
Army battle game defend
6
4
Wave-based shooter defend
Morph-collect adventure collect
Morph-collect adventure collect
4
Battle royale brawler compete
3
5
Melee PvP arena tactical
Character-select brawler roleplay
Battle royale brawler compete
3
Zombie-wave shooter defend
>10
Table 2: Case study of amplifying two concepts for one user (features 2639 for collaborate and 1252 for defend ). ‘Before’ denotes the rank of each game in the original top-10, and yellow shading marks the ground truth game. Icons follow the concepts. Game titles and thumbnails are anonymized. Additional examples are in Appendix C.4 .
Figure 3: Change in target concept share under amplification and suppression ( α=8 ). Values are percentage points relative to the original model, and positive values mean more games of that concept in the top-10.
Figure 4: Accuracy for concept-matched users under intervention, averaged over all concepts.
Figure 5: Accuracy of concept-matched users and all users under amplification across α . We show the four concepts of the Atomic ID with the largest gains in target concept share on the validation split.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Atomic ID
Semantic ID
Catalog size (games)
40K
100K
SAE training states
79K
381K
Validation users
3.2K
3.5K
Discovery users
6.4K
7.0K
Evaluation users
6.3K
6.9K
Mean history length
616
213
Appendix
Table 3: Data statistics used for the SAE study.
Hyperparameter
Atomic ID
Semantic ID
GR model
Optimizer
AdamW
AdamW
Learning rate
10−4
10−4
Adam (β1,β2)
(0.95, 0.99)
(0.95, 0.99)
Weight decay
0.06
0.06
SAE
Appendix
Table 4: Final hyperparameters for SteerGR.
Short label
Label
Definition
build
build and customize
Players place, modify, or decorate objects and structures to create a personalized environment or unique character appearance.
compete
compete for dominance
Players engage in structured matches or challenges to outperform rivals and secure a higher standing on a leaderboard.
traverse
traverse challenging terrain
Players navigate complex environments by jumping, climbing, and dodging obstacles to reach a specific destination.
manage
manage resources and wealth
Players harvest, trade, or invest in-game assets to grow their economy and unlock better equipment or upgrades.
collaborate
collaborate with others
Players combine their efforts, skills, or roles to achieve a shared objective or complete group tasks.
solve
solve logic puzzles
Players decipher patterns, analyze clues, or rearrange objects to progress through interactive brain-teasing sequences.
Appendix
Table 5: LLM-generated play concepts. ‘Short labels’ denotes the name used in the paper, and the definitions are embedded to compute concept scoring.
Model
R2
L0
N@10 with block output
Rec.
Original
SAE
Mean
Atomic
0.750
32
0.477
0.408
0.064
83.4%
Semantic
0.951
64
0.214
0.206
0.036
95.5%
Appendix
Table 6: SAE quality on final evaluation users. ‘Rec.’ denotes the recovered share, computed as (SAE−Mean)/(Original−Mean) .
Concept
AUC
Genres
GPT
Gemini
build and customize
0.75
0.96
10
compete for dominance
0.84
0.90
10
traverse challenging terrain
0.96
0.97
6
manage resources and wealth
0.90
0.96
9
collaborate with others
0.73
0.81
12
Appendix
Table 7: AUC of concept scores, with LLM judgments on 400 games. ‘Genres’ denotes the number of genres out of 15 among the games selected for each concept.
Figure 6: Target concept share across the number of selected features per concept, Nsel∈{1,2,5,10,20} ( α=4 ). We use Nsel=1 by default.
Figure 7: Target concept share versus off-target change for SteerGR across the number of features per concept Nsel∈{1,2,5,10,20} ( α=4 ). Numbers mark Nsel .
Figure 8: Effects of the assigned feature on the target concept share, compared with other features. Results are for held-out users under amplification ( α=4 ). Changes are in percentage points relative to the original model.
Figure 9: Concept association of the selected features, measured by Cohen’s d of pre-activations between high- and low-scoring user groups.
Recent plays (oldest → newest)
Original top-5
Amplifying collaborate with others
Amplifying defend against threats
Game, concept
Game, concept
Game, concept
Before
Game, concept
Before
1
Free-item obstacle course showcase
FPS duel game compete
FPS duel game compete
1
FPS duel game compete
1
2
Downhill sled runner explore
Downhill sled runner explore
Downhill sled runner explore
2
Downhill sled runner explore
2
3
Bomb-passing game defend
FPS deathmatch game compete
FPS deathmatch game compete
3
FPS deathmatch game compete
3
4
Team arena shooter defend
Racket sports game compete
Racket sports game compete
4
Bomb-passing game defend
6
5
Co-op road-trip survival collaborate
Open-world action RPG manage
Team block co-op course collaborate
9
Team arena shooter defend
8
Appendix
Table 8: Case study of amplifying two concepts ( α=4 ; features 2639 for collaborate and 1252 for defend ). ‘Before’ denotes the rank of each game in the original top-10, and shading marks games labeled with the amplified concept. Icons follow the concepts. Game titles and thumbnails are anonymized.
Recent plays (oldest → newest)
Original top-5
Amplifying collaborate with others
Amplifying manage resources and wealth
Game, concept
Game, concept
Game, concept
Before
Game, concept
Before
1
Co-op horror escape tactical
Forest survival game collect
Forest survival game collect
1
Forest survival game collect
1
2
Toss-and-carry course collaborate
Monster escape survival collect
Monster escape survival collect
2
Steal-and-defend tycoon manage
9
3
Piggyback climb course collaborate
Team duel shooter collect
Team duel shooter collect
3
Elimination party game compete
5
4
Shared-plank course collaborate
Clicker simulator tactical
Elimination party game compete
5
Farming tycoon manage
6
5
Horror morph roleplay roleplay
Elimination party game compete
Two-player co-op course collaborate
>10
Team duel shooter collect
3
Appendix
Table 9: Case study of amplifying two concepts ( α=4 ; features 2639 for collaborate and 2787 for manage ). ‘Before’ denotes the rank of each game in the original top-10, and shading marks games labeled with the amplified concept. Icons follow the concepts. Game titles and thumbnails are anonymized.
Recent plays (oldest → newest)
Original top-5
Amplifying traverse challenging terrain
Suppressing traverse challenging terrain
Game, concept
Game, concept
Game, concept
Before
Game, concept
Before
1
Driving game –
Vehicle stunt simulator collaborate
Wall-jump course traverse
2
Vehicle stunt simulator collaborate
1
2
Open-world heist game collect
Wall-jump course traverse
Wall-jump drills traverse
4
Knife duel game –
5
3
Open-world action RPG manage
Chance-based game –
Staged wall-jump course traverse
>10
Chance-based game –
3
4
Fishing simulator collect
Wall-jump drills traverse
Wall-jump tower traverse
>10
Co-op flight survival collaborate
7
5
Team volleyball game compete
Knife duel game –
Short wall-jump tower traverse
>10
Checkpoint roleplay defend
>10
Appendix
Table 10: Case study of amplifying and suppressing the same concept ( α=4 ; feature 927 for traverse ). ‘Before’ denotes the rank of each game in the original top-10, and shading marks games labeled with the concept. Icons follow the concepts. Game titles and thumbnails are anonymized.
Recent plays (oldest → newest)
Original top-5
Suppressing compete for dominance
Suppressing collect rare items
Game, concept
Game, concept
Game, concept
Before
Game, concept
Before
1
Ability arena fighter tactical
Item-code hub collect
Item-code hub collect
1
Block-breaking game –
6
2
Item-code hub collect
Reflex PvP game tactical
Reflex PvP game tactical
2
Platform brawler compete
5
3
Platform brawler compete
Collector hangout collect
Collector hangout collect
3
Physics puzzle solve
>10
4
Block-breaking game –
Platform fighting game compete
Anime fighting sandbox tactical
8
PvP tower climb compete
>10
5
Social utility game roleplay
Platform brawler compete
Story escape room solve
10
Precision jump course traverse
>10
Appendix
Table 11: Case study of suppressing two concepts ( α=4 ; features 838 for compete and 2855 for collect ). ‘Before’ denotes the rank of each game in the original top-10, and shading marks games labeled with the suppressed concept. Icons follow the concepts. Game titles and thumbnails are anonymized.
Figure 10: Prompt for generating play concepts. The name of platform and details are omitted for anonymity.
Figure 11: Prompt for judging concepts with LLMs. The judges see no concept scores, assigned concepts, or feature information. The name of platform and details are omitted for anonymity.
Generative recommendation (GR) is moving industrial recommender systems (RecSys) from scoring candidate items to autoregressively generating item identifiers. This shift promises higher recommendation quality, but makes serving substantially harder: each request executes a short sequence of dependent decoding steps that explore a large item space and select promising candidates on the latency-critical path. GR not only creates structural mismatches with conventional RecSys pipelines, but also challenges the architectural and optimization assumptions of LLM serving stacks, making direct adoption of existing systems difficult. We present RecoGEM, our production GR serving system. RecoGEM uses a runtime--engine decoupling architecture to move GR-specific serving responsibilities, such as candidate-state transitions and live artifact updates, to a dedicated runtime. Its execution engine exploits GR's short token sequences, large candidate populations, and predictable step shapes to optimize self-attention, candidate selection, and execution planning. RecoGEM has been deployed for over nine months across our production lines, serving applications with hundreds of millions of users and supporting more than ten GR model families. On our OneModel production workloads, RecoGEM reaches over 1,000 QPS per serving instance with P99 latency below 35 ms and a 2.88× throughput improvement over representative baselines.
Xiangrui Yang, Yunfeng Zhao, Jiawei Yang +8
Kuaishou Technology, Beijing, China · Peking University, Beijing, China · Renmin University of China, Beijing, China
Generative recommendation (GR) predicts the next item through its semantic ID (SID). Existing methods commonly fit SID vocabularies to pretrained language model (PLM) representations of item descriptions, incurring additional training costs and making vocabulary coverage dependent on the fitting domains. We observe that pretrained sparse autoencoders expose PLM features with human-interpretable semantic annotations, allowing feature indices to serve as a reusable SID vocabulary. However, reusing this vocabulary requires preserving item distinctions under compact token budgets and learning to favor target items over candidates with shared dominant features. It also requires extending feature interpretability from item content to retrieval decisions. To address these challenges, we propose API, a framework that Adapts the Pretrained Interpretable vocabulary for effective and explainable GR without additional vocabulary learning. Within API, Pretrained Interpretable IDs (PIDs) retain selected feature indices and relative activation strengths for item discrimination. With PIDs as supervision, the Semantic Prediction Adapter (SPA) learns user preferences from interaction histories through joint feature-index and strength prediction. SPA combines both predictions to score candidate items, with item-level supervision encouraging the observed next item to outrank its alternatives. API further explains candidate score differences through feature contributions paired with their pretrained semantic annotations. Extensive experiments show that API achieves competitive performance across backbones and decoding schemes, while preserving item distinctions under compact token budgets and enabling explainable retrieval.
Ziluowen Luo, Ruochen Liu, Jun Yin +6
Central South University Changsha, Hunan, China · The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · Hong Kong Polytechnic University Hongkong, China +4
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
Yuxuan Hu, Yingyi Zhang, Wenlin Zhang +11
City University of Hong Kong Hong Kong, China · Platform and Content Group Tencent Shenzhen, China