vault

Tiger

6 min read

Tiger

TIGER (Transformer Index for GEnerative Retrieval) is a generative recommender introduced by Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy (Google) in their 2023 NeurIPS paper “Recommender Systems with Generative Retrieval”. The model represents a fundamental architectural shift from prior sequential recommenders: instead of scoring items from a fixed vocabulary (as SASRec and BERT4Rec do), TIGER generates the next-item identifier as a sequence of Semantic IDs — discrete codewords derived from the item’s content via a residual-quantized variational autoencoder (RQ-VAE) over pretrained text embeddings. The Semantic ID representation has two properties that distinguish it from raw integer item IDs: (1) it is content-derived, so semantically similar items have similar Semantic IDs (improving cold-start handling); (2) it is a short token sequence (typically 4 codewords per item) rather than a single ID, allowing the recommender to generate item IDs token-by-token like a sequence-to-sequence language model rather than scoring against the full vocabulary. TIGER is the most influential example of the generative recommender paradigm and represents one of the most active research directions in 2024–2026.

1. The Generative Retrieval Paradigm

Standard sequence-aware recommenders (SASRec, BERT4Rec) are discriminative: they take a user’s history and produce a probability distribution over the existing item vocabulary. The vocabulary size is fixed at training time; new items added to the catalog after training have no embedding and cannot be scored.

Generative retrieval reframes the problem as a sequence-to-sequence task: given the user’s history (encoded as a sequence of Semantic IDs), generate the Semantic ID of the next item as an output sequence of codewords. The architecture is structurally identical to a sequence-to-sequence transformer used in machine translation or text summarization. The key difference: items are not represented as opaque integer IDs but as compositions of meaningful semantic codewords.

This has several consequences:

Cold items are tractable. A new item gets a Semantic ID immediately from its content (no training required). The generative recommender can produce the new item’s Semantic ID as long as the codeword space the new item occupies has been seen during training in some other item.

The model can generate “novel” items. In principle, the model can produce Semantic IDs that no item in the training catalog has, allowing for synthetic recommendations or for recommending items added between training runs.

Scoring is replaced by generation. Top-K retrieval becomes beam search over the Semantic ID space, which has different computational properties from softmax over a flat vocabulary.

2. Semantic IDs — The Core Representation

A Semantic ID is a short tuple of discrete codewords (c_1, c_2, c_3, ..., c_L) derived from the item’s content. The construction:

Step 1: content embedding. For each item, compute a dense content embedding using a pretrained encoder (the paper used SentenceT5 — a sentence-transformer variant). This produces a d-dimensional vector per item.

Step 2: residual quantization. Apply a Residual-Quantized Variational Autoencoder (RQ-VAE) — a hierarchical quantization scheme — to the content embedding. The first codeword c_1 is the index of the nearest codebook entry (out of, say, 256 codewords) to the content embedding. The residual (embedding minus codeword) is then quantized to a second codebook entry c_2. The residual after that is quantized to c_3, and so on. After L layers, the item is represented as the tuple (c_1, c_2, ..., c_L).

The Semantic ID is therefore a coarse-to-fine semantic representation: c_1 captures the item’s broad category; c_2 refines within that category; subsequent codewords add finer details.

The paper used L = 4 codewords with codebook size 256 each, giving 256^4 ≈ 4.3 billion possible Semantic IDs — a much larger space than typical item catalogs, ensuring that semantically similar items can have distinct IDs while still sharing prefixes.

3. The Generative Model

The recommender is a sequence-to-sequence transformer (encoder-decoder, T5-style). The encoder takes the user’s interaction sequence as a sequence of Semantic IDs flattened into a single token sequence:

Encoder input: [c^1_1, c^1_2, c^1_3, c^1_4, c^2_1, c^2_2, c^2_3, c^2_4, ..., c^t_1, c^t_2, c^t_3, c^t_4]

where c^j_l is the l-th codeword of the j-th item in the history.

The decoder generates the Semantic ID of the next item, one codeword at a time:

Decoder output: c^{t+1}_1, c^{t+1}_2, c^{t+1}_3, c^{t+1}_4

Both encoder and decoder are standard transformer stacks.

Training: standard sequence-to-sequence cross-entropy loss on the next-item Semantic ID generation task.

Inference: beam search over the Semantic ID space, restricted to valid Semantic IDs that correspond to items in the catalog. The beam search expands codeword by codeword; at each step, only the codewords valid for the next position (given the prefix) are considered.

4. Cold-Start Handling

A new item added to the catalog after training:

  1. Its content is encoded with the pretrained sentence-transformer (no model retraining needed).
  2. The RQ-VAE produces its Semantic ID (the codebook is fixed; new items just use existing codewords).
  3. The new item is added to the inventory of valid Semantic IDs for beam search.
  4. The recommender can immediately surface the new item if its Semantic ID is generated during beam search.

The cold-start handling is structural — it works without retraining or fine-tuning, simply by virtue of the Semantic ID representation being content-derived.

5. Empirical Results

The paper reports significant improvements over SASRec on standard sequential-recommendation benchmarks (Amazon Beauty, Sports, Toys subsets), with particularly strong gains on cold-item evaluation. The improvements scale with the size of the model.

The paper also demonstrates that the Semantic IDs produced by the RQ-VAE are interpretable — items with similar c_1 codewords are in the same broad category, etc.

6. When to Use TIGER

TIGER is appropriate when:

  • Cold-item performance matters substantially.
  • You have rich item content (text, image) for Semantic ID derivation.
  • You can afford the additional infrastructure (sentence-transformer for content embeddings, RQ-VAE training, beam-search inference).

In 2026, TIGER and generative recommenders more broadly are an active research area but not yet a default production deployment. They are most appropriate for high-cold-start workloads (news, e-commerce with high catalog turnover) and for systems where the additional complexity is justified by the cold-start benefits.

7. Strengths

  • Native cold-item handling through content-derived Semantic IDs.
  • Generative framing — same architecture as language models; benefits from NLP advances.
  • Interpretable Semantic IDs — coarse-to-fine codewords have meaning.
  • Scalable — beam search over Semantic ID space avoids the full-vocabulary softmax.

8. Weaknesses

  • Engineering complexity. Sentence-transformer + RQ-VAE + sequence-to-sequence transformer + beam-search infrastructure. Substantially more complex than a standard SASRec.
  • Beam-search inference is more expensive than direct softmax in some setups.
  • Codebook design matters. Codeword count, number of layers, codebook training regime all affect quality.
  • Limited production deployment as of 2026 — most commercial systems still use discriminative methods.

9. See Also

Graph View

Backlinks