Generative Recommendation
Recasting recommendation as sequence generation over user action histories with large transformer models that scale with training compute.
Tier Grades
Tier A: Frontier AI labs
feasible
Meta and Kuaishou train and deploy generative recommenders up to 1.5 trillion parameters (HSTU).
Tier B: Indonesian AI labs
stretch
Scaling gains need tens of billions of daily actions (HSTU), data held by platforms rather than labs; Indonesian lab access is not in the sources.
Tier C: Institution units such as ITB
stretch
Open OneRec-Foundation 1.7B and 8B checkpoints allow study, but the scaling result needs industrial data (HSTU).
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
stretch
Proofs of value need a client's interaction logs at scale; small clients will not reach the HSTU regime.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeHSTU-based Generative Recommenders reach 1.5 trillion parameters, and quality scales as a power law of training compute across three orders of magnitude up to GPT-3 or Llama-2 scale.
- DataIndustrial streams of tens of billions of user actions a day (HSTU); OneRec encodes user behaviour sequences and decodes videos with sparse MoE and DPO-based preference alignment.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 12 authors on HSTU; 8 on OneRec.
- Time horizonOne to two years.
- Shelf-life signalsDeployed replacements of retrieve-and-rank cascades (OneRec at Kuaishou, HSTU on a platform with billions of users), Open LLM-backbone recommenders (OpenOneRec 1.7B and 8B)
- Pipeline roleTier A platforms publish architectures and open checkpoints; Tier C studies them on public data; Tier D applies smaller versions with clients who own large interaction logs.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
OpenOneRec / OneRec-FoundationrecommendationexemplarOneRec-V2recommendationexemplarOneRec (V1)recommendationexemplar
HSTU (Generative Recommenders)recommendationexemplarOneRec-Foundation 1.7Brecommendationproduced byOneRec-Foundation 8Brecommendationproduced byOneRec-Thinkrecommendationproduced by
MTGRrecommendationproduced byLinks
R&D TopicGenerative Recommendation
ModelsOneRec (V1)Kuaishou, recommendationOneRec-V2Kuaishou, recommendationOpenOneRec / OneRec-FoundationKuaishou, recommendationHSTU (Generative Recommenders)Meta, recommendation+4 more
OneRec (V1)Kuaishou, recommendationOneRec-V2Kuaishou, recommendationOpenOneRec / OneRec-FoundationKuaishou, recommendationHSTU (Generative Recommenders)Meta, recommendation+4 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 8
OneRec (V1)Kuaishou, recommendationexemplarOneRec-Foundation 1.7BKuaishou, recommendationproduced byOneRec-Foundation 8BKuaishou, recommendationproduced byOpenOneRec / OneRec-FoundationKuaishou, recommendationexemplarHSTU (Generative Recommenders)Meta, recommendationexemplarMTGRMeituan, recommendationproduced byOneRec-ThinkKuaishou, recommendationproduced byOneRec-V2Kuaishou, recommendationexemplarPapers
- Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)Shijie Geng et al., RecSys 2022 (per arXiv comment), 2022 founding
- OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference AlignmentJiaxin Deng et al., arXiv preprint, 2025 state of the art
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative RecommendationsJiaqi Zhai et al., ICML'24 (per arXiv comment), 2024 state of the art
- Recommender Systems with Generative RetrievalShashank Rajput et al., NeurIPS 2023 (per arXiv comment), 2023 state of the art
- Large Language Models for Generative Recommendation: A Survey and Visionary DiscussionsLei Li et al., COLING 2024 (per arXiv comment), 2023 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations arxiv.org https://arxiv.org/abs/2402.17152
- OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment arxiv.org https://arxiv.org/abs/2502.18965
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Generative Recommendation", https://aurion.semai.id/topics/generative-recommendation/, records of October 2026.