Google's TokenMinds generates LLM-based user tokens for YouTube recommendations
Billions of YouTube users now get recommendations from discrete SID tokens and dense embeddings...
TokenMinds, developed by Google/YouTube researchers, addresses a key limitation in user modeling for large recommender systems: dense embeddings alone struggle with representational constraints. The system extends the PLUM framework from item retrieval to user modeling, using an encoder-decoder architecture adapted from pre-trained LLMs. It produces two complementary outputs: discrete Semantic ID (SID)-based user tokens and traditional dense embeddings. The shared SID vocabulary allows unifying user behaviors across long-form and short-form video into a single model, which Google claims substantially reduces training and serving costs.
Validated through extensive offline experiments and live launches on multiple YouTube surfaces, TokenMinds serves full user traffic (billions of users) via an asynchronous infrastructure that decouples representation generation from downstream scoring. Focusing on ranking as the primary use case, the team found that SID-based user tokens and dense embeddings provide complementary value across different production ranking systems. This duality means recommender systems can benefit from both the semantic grounding of discrete tokens and the flexibility of dense embeddings, while achieving significant operational efficiencies at industrial scale.
- TokenMinds generates both discrete SID-based user tokens and dense embeddings via an LLM-based encoder-decoder architecture.
- Validated on full YouTube traffic (billions of users) with an asynchronous infrastructure decoupling representation generation from scoring.
- Unifies long-form and short-form video behaviors into a single model, reducing training and serving costs substantially.
Why It Matters
TokenMinds proves that discrete user tokens complement dense embeddings at industrial scale, reducing costs while improving ranking.