Meta · Multimodal LLMs
Generative Recommendation · Season 2 · 2026 Frontier
Generative Recommendation in Deep Water: The 2026 Frontier
From learnable Semantic IDs, cascaded reasoning, and continuous tokens to page generation, advertising deployment, and generalization audits
Once a model can generate items, how should IDs learn, how does bias propagate, how can the system ship, and how do we prove the capability is real?
阅读中文版 →The question of this season
The model can already generate. The difficult part starts now.
A seven-paper roadmap
Each paper solves one problem and hands a harder one to the next.
01–02
Representation and reasoning
Let recommendation loss reshape the item language, then confront cascading bias in autoregressive prefixes.
DIGER
The Tokenizer Can Finally Hear Recommendation Loss
Make Semantic IDs differentiable and trainable by downstream gradients while using exploration and uncertainty decay to prevent codebook collapse.CARE
The Deeper the Generation, the Larger the Bias Can Grow
Interpret token decoding as cascaded ranking and use progressive history plus query-anchored parallel reasoning to limit prefix errors.03–04
Continuous generation and item creation
Question whether discrete SIDs are necessary and extend outputs from catalog items to newly generated content.
ContRec
Semantic IDs Are Not the Only Answer
Remove discrete quantization, generate preference vectors with continuous tokens and diffusion, then map them back to real items.DualFashion
The Recommender Starts Generating the Product
Use dual diffusion to generate both fashion images and structured descriptions, moving recommendation from catalog selection toward personalized design.05–07
Industrial systems and capability audits
Move from page generation and advertising pipelines to separating memorization from genuine generalization.
GenRec
One Next Item Is Not Enough: Generate the Whole Page
Train at page level, compress long histories, and align complex preferences with page-level reinforcement learning.GR4AD
Generative Recommendation Enters the Advertising Mainline
Encode business value into advertising Semantic IDs, list rewards, and low-latency decoding, with deployment evidence at hundreds of millions of users.How Well Does Generative Recommendation Generalize?
Do Semantic IDs Really Generalize Better?
Split test cases into memorization and compositional-generalization regimes to expose what aggregate NDCG can hide.Season boundary