One Token per Creator: A Simpler Route to Generative Recommendations

When the identifier becomes the bottleneck

In Beyond Semantic IDs: Atomic Creator Tokens for Large-Scale Generative Recommendation, Zhuosheng Liu and colleagues at Meta investigate recommendations of creators whom users do not already follow. Their models predict the next creator a user will engage with from profile information and a short engagement history.

The paper examines a generative approach in which a language model produces identifiers rather than written recommendations. Its Semantic ID, or SID, baseline compresses a previously learned creator embedding into a sequence of codes. The model predicts that sequence, which the system then translates back into creators.

The authors identify two weaknesses in this pipeline. Multiple creators can share one SID, so predicting the correct code does not necessarily identify the correct person. Their baseline resolves these collisions by ranking creators within a code by follower count. Separately, the creator encoder and quantizer remain frozen: recommendation training cannot adjust those underlying representations.

These are findings about the evaluated design, not proof that every possible semantic-identifier system must perform worse.

A dedicated token for each creator

Atomic Creator Token, or ACT, removes the intermediate code. Every candidate creator receives a unique vocabulary token, so there is no ambiguity when translating a prediction back into a creator.

The implementation uses Llama-3.2-3B. User profiles appear as text, while engagement histories contain creator tokens. The model scores the candidate creator tokens directly and returns the highest-scoring creators.

Training first warms up the new creator embeddings with the backbone frozen, using creator descriptions and co-engagement sequences. A second stage jointly trains those embeddings and adapts the backbone through Low-Rank Adaptation. Unlike the frozen SID representation, ACT’s creator embeddings can change in response to prediction errors.

The authors also tested preference-optimization methods, but none improved on supervised training. Their served model therefore uses the simpler supervised approach.

Larger gains in the production comparison

At the middle production vocabulary size, ACT increased creator-level Recall@15 from 23.58% to 43.41%. That is a 19.83-percentage-point improvement, or about 1.84 times the baseline result. Recall@15 measures how often the actual next creator appears among the 15 returned candidates; it is not a measure of user satisfaction.

Both systems used the same backbone, with the token layer as the controlled change. The middle setting involved approximately 10 million training engagement events and one million test examples, split by day.

Giving SID more code capacity narrowed, but did not eliminate, ACT’s lead. At the broadest vocabulary, a code space 64 times the nominal base capacity achieved 30.15% Recall@15, compared with ACT’s 38.87%.

That broadest comparison has an important qualification: SID was evaluated only on creators with available graph-network embeddings, approximately 76% of that catalog, while ACT covered all candidates. The paper reports production vocabulary and codebook sizes in relative units rather than exact counts.

Public data exposes the training requirement

On KuaiRand-27K, the researchers trained both approaches on identical samples, without the initial warm-up stage. The candidate set contained the 50,000 creators with the most training clicks, covering 90.4% of training clicks.

ACT achieved 2.763% Recall@15, compared with 1.859% for SID and 2.351% for global popularity. SID slightly outperformed ACT at the single-recommendation cutoff, so ACT did not win every comparison.

Smaller KuaiRand-1K experiments were less favorable: ACT failed to beat the best simple heuristic, even after restricting the vocabulary to popular creators. Those settings provided roughly five or 34 training examples per creator on average, versus about 331 in the larger experiment. Because total training data also differed, the comparison does not establish a universal minimum number of examples per token.

Faster serving, cautious online evidence

Under the reported hardware and decoding settings, ACT also reduced inference cost. In the middle production setting, GPU time per user fell from 97.5 to 35.0 milliseconds. The authors attribute this largely to generating one creator token instead of a five-token SID block, reducing decoding work and cached state.

Live A/B results were positive in direction but small. Product engagement rose 0.29% against the SID-enhanced production system and 0.66% against the classic ranker; the latter approached statistical significance at approximately p = 0.05. These results are suggestive, not decisive evidence of broad online improvement.

The practical boundary is explicit: unique tokens require training examples, and scoring them becomes more expensive as the vocabulary grows. The authors propose ACT for well-observed creators, with conventional retrieval handling the sparse long tail. Their central contribution is evidence that, within that setting, simplifying creator identifiers can improve both prediction and serving efficiency.