DIN
Deep Interest Network for Click-Through Rate Prediction
Attention over past behaviours, weighted by the candidate item.
No deep dive on this stop yet; the arXiv paper is linked below.
Time Machine · Large recommendation models
29 stops, ordered by the month each paper first appeared; pick a stop for the key points, and read our deep dive inline where there is one.
Generative recommendationScalingLong-sequence modeling← → stops · Esc closes
DIEN
Deep Interest Evolution Network for Click-Through Rate Prediction
Models how interests evolve over time with a GRU.
No deep dive on this stop yet; the arXiv paper is linked below.
SIM
Search-based User Interest Modeling with Lifelong Sequential Behavior Data for Click-Through Rate Prediction
Two stages: retrieve relevant behaviours, then run fine attention; sequences reach tens of thousands.
No deep dive on this stop yet; the arXiv paper is linked below.
P5
Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)
Writes rating, sequential recommendation and explanation tasks as text prompts for one language model.
No deep dive on this stop yet; the arXiv paper is linked below.
TWIN
TWIN: TWo-stage Interest Network for Lifelong User Behavior Modeling in CTR Prediction at Kuaishou
Makes the retrieval and ranking stages use the same relevance measure.
No deep dive on this stop yet; the arXiv paper is linked below.
TIGER
Recommender Systems with Generative Retrieval
Gives each item a Semantic ID and recommends by generating that code token by token.
No deep dive on this stop yet; the arXiv paper is linked below.
Scaling SR
Scaling Law of Large Sequential Recommendation Models
An early systematic measurement of scaling laws for sequential recommenders.
No deep dive on this stop yet; the arXiv paper is linked below.
HSTU
Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations
Recasts retrieval and ranking as generative modelling of action sequences; the first clear scaling curve in recommendation.
No deep dive on this stop yet; the arXiv paper is linked below.
Wukong
Wukong: Towards a Scaling Law for Large-Scale Recommendation
Stacked factorization machines for feature interaction, scaling across two orders of magnitude.
No deep dive on this stop yet; the arXiv paper is linked below.
TWIN V2
TWIN V2: Scaling Ultra-Long User Behavior Sequence Modeling for Enhanced CTR Prediction at Kuaishou
Compresses behaviours with hierarchical clustering to reach life-cycle-long sequences.
No deep dive on this stop yet; the arXiv paper is linked below.
OneRec
OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment
Replaces the retrieve-then-rank cascade with one generative model, then aligns it to user preferences.
No deep dive on this stop yet; the arXiv paper is linked below.
LONGER
LONGER: Scaling Up Long Sequence Modeling in Industrial Recommenders
Global tokens plus token merging put long sequences end to end into a Transformer.
No deep dive on this stop yet; the arXiv paper is linked below.
MTGR
MTGR: Industrial-Scale Generative Recommendation Framework in Meituan
Switches ranking to the HSTU architecture while keeping all cross features; training cost stays flat.
Our deep dive covers the Meituan tech-blog version (2025-05), not the arXiv paper.
OneRec TR
OneRec Technical Report
End-to-end generative recommendation fully deployed in the main short-video feed, with its compute and gains published.
No deep dive on this stop yet; the arXiv paper is linked below.
基础+专家
Realizing Scaling Laws in Recommender Systems: A Foundation-Expert Paradigm for Hyperscale Model Deployment
A large foundation model plus per-surface expert models, so scaling gains can actually ship.
OneRec-Think
OneRec-Think: In-Text Reasoning for Generative Recommendation
The model writes out its reasoning first, then outputs the items to recommend.
OneTrans
OneTrans: Unified Feature Interaction and Sequence Modeling with One Transformer in Industrial Recommender
Behaviour sequences and every other feature become one token sequence for a single Transformer.
10K 序列
Make It Long, Keep It Fast: End-to-End 10K Long User Behavior Sequence Modeling for Billion-Scale Douyin Recommendation
10K behaviours end to end in ranking at roughly one eleventh of the compute.
LLaTTE
LLaTTE: Scaling Laws for Multi-Stage Sequence Modeling in Large-Scale Ads Recommendation
Measures power laws along every scaling dimension of multi-stage sequence models.
MixFormer
MixFormer: Co-Scaling Up Dense and Sequence in Industrial Recommenders
Sequence modelling and dense feature interaction share parameters and scale together.
UxSID
UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence
Groups candidates by Semantic ID; each group shares one offline-computed user interest, served with a single lookup.
行为基础模型
Scaling Laws for Behavioral Foundation Models over User Event Sequences
A Chinchilla-style compute allocation guide for behavioural foundation models.
TM20K
Teacher Retains Full Tokens, Student Merges Efficiently: TM20K for E-Commerce Sequence Modeling in Ad Recommendation
A teacher reads all 20K behaviours, then distils the long-sequence knowledge into a token-merging student.
Densing Law
Towards a Densing Law for User Representation Learning at Billion-Scale Capacity
Behaviour data repeats so much that scaling stalls; the paper turns to the density of user representations instead.
SequenceO1
SequenceO1: End-to-End Ultra-Long (100K) Sequence Modeling in Recommendation with Low-Rank Caching
Compresses up to 100K behaviours into one cached per-user summary shared by all candidates.
OneTrans-V2
OneTrans-V2: Unifying Retrieval, Pre-rank, and Fine-rank with One Transformer in Industrial Recommender
Retrieval, pre-ranking and ranking become three tasks of one Transformer; user history is encoded once.
Predictable Scaling
Predictable Scaling: Efficient Model Exploration via Scaling-Law Prediction
Uses small models to reliably predict whether a change helps at large scale.
OneLatent
OneLatent: Latent Reasoning for Efficient Foundation Recommendation Models
Compresses write-the-reasoning-then-the-SID into three fixed latent tokens; no reasoning text at serving time.
PinGER
PinGER: Token-Efficient Generative Retrieval with Semantic IDs
Sees more in training and reads less in serving, bringing SID retrieval input cost down to embedding-retrieval levels.