TIGER
The paper that turned recommendation into a generation problem — items as Semantic IDs, recommendations as text generation.
About
TIGER (Transformer Index for GEnerative Recommenders) is the 2023 NeurIPS paper from Google DeepMind that launched the generative-recommendation wave. The insight: instead of matching a user vector against an item index, tokenize every item into a short Semantic ID — a tuple of codewords from an RQ-VAE trained over content embeddings — and then treat recommendation as sequence generation. A T5 encoder-decoder reads the user's interaction history as token IDs and autoregressively generates the next item's Semantic ID, with constrained beam search guaranteeing only real items come out.
Because items are identified by content-derived tokens rather than opaque IDs, the model handles cold-start items naturally and dispenses with the discrete ANN index entirely — the transformer's parameters become the index. It beat SASRec-style baselines on Amazon benchmarks in the paper, and a community from-scratch replication measured +30% NDCG@10 over SASRec on Amazon Beauty.
Google never released official code; the ecosystem grew through community reimplementations (EdoardoBotta/RQ-VAE-Recommender, Tiger-DPO-RecSys), and the semantic-ID idea now underpins a whole subfield of generative recommenders.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (NeurIPS 2023, arXiv 2305.05065): outperformed ID-based sequential baselines (SASRec, S3-Rec) on Amazon benchmarks; 500+ citations
- Independent from-scratch replication (rhine-yangrui/Generative-Sequential-Recommendation, Amazon Beauty 5-core, all-rank): Semantic-ID + T5 achieved R@10 0.0604 / N@10 0.0326 vs SASRec R@10 0.0573 / N@10 0.0250 — about +30% NDCG@10 over SASRec, close to the paper's reference numbers (R@10 0.0648 / N@10 0.0384)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: peer-reviewed NeurIPS 2023 paper (arXiv 2305.05065) from Google DeepMind with named authors; 500+ citations; key design (RQ-VAE Semantic IDs + T5 + constrained beam search) independently reimplemented and validated, with a community replication landing near the paper's numbers. Strengths: paradigm-defining paper, independently reproduced, strong adoption as a research baseline. Docked for: (1) no official Google code or weights — all implementations are third-party; (2) headline numbers are author-reported. License: research paper; community code under various open licenses.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · NeurIPS 2023 paper (arXiv 2305.05065) from Google DeepMind (Rajput et al.); 500+ citations; RQ-VAE Semantic IDs + T5 + constrained beam search.
- · Independently reimplemented and validated: community from-scratch replication on Amazon Beauty landed at R@10 0.0604 / N@10 0.0326 vs SASRec 0.0573 / 0.0250 — near the paper's reference numbers.
- · No official Google code or weights; headline numbers are author-reported.