A preprint reports that AMBER, a recommendation system built around compact Event Tokens, produced the lowest reported normalized entropy in its controlled offline ranking comparison. Normalized entropy, or NE, is a lower-is-better, calibrated log-loss measure. AMBER was 0.40% better than a fair version of the incumbent system, 0.30% better than a pointwise model, 1.00% better than an HSTU-style input, and 1.60% better than a system combining Semantic IDs with content-understanding embeddings. The study asks whether preserving more of each interaction’s temporal snapshot can improve industrial ranking and retrieval without pushing compute costs too high. It uses petabyte-scale chronological impression logs, with the main results using the full dataset and the ablations using a 10% subsample. Scores were calculated on unseen future dates, so the reported gains are offline model comparisons, not demonstrated changes in user behavior or revenue.
One event, one compact representation
AMBER changes the unit being fed into the sequence model. A learned encoder compresses each heterogeneous event snapshot into a compact Event Token, which is asynchronously pre-computed and cached for serving. The representation is first aligned with a frozen User large language model, or LLM, then trained jointly end to end with the downstream sequence model. The study also uses recurrent retraining. In the sequence design, retrieval receives one token per event, while ranking receives two. That lets the system use the same event-centered representation for both tasks while giving the ranking model more room to describe each event.
The retrieval result was measured with Soft Recall, a value-weighted score intended to capture end-to-end utility. AMBER’s improvement ranged from 0.31% to 0.51% across checkpoint ages. It exceeded CU embeddings by 0.28 to 0.32 percentage points and remained 0.31% better at Day 8, when CU embeddings became neutral. A separate offline test blended AMBER with the incumbent system and found lower Ensemble NE under both simulated delay windows: a 0.10% reduction at 10 minutes and a 0.16% reduction at 1 minute.
The information going in mattered
The size of the ranking gap depended on which features were available. In feature ablations, removing item features left a 6.65% NE gap. A version with only Semantic IDs and outcome signals had a 2.40% gap; adding content-understanding and other item categorical features narrowed it to 1.52%. Adding user features was associated with a further 1.10% NE improvement. These were offline ablations, so they show how the tested configurations compared, not that any single feature would produce the same change in a live service.
The cost question is part of the claim
The efficiency analysis counted more than the cost of running the User LLM. Its total-cost model also included the Event Tokenizer and feature materialization, the work needed to prepare features for serving. Once serving compute was included, scaling on the event side was more compute-efficient. Tokens co-trained with a 16-layer User LLM improved more rapidly downstream than tokens co-trained with a 2-layer model. The comparison relies on estimated serving costs, so the exact trade-off depends on the infrastructure used.
Context length changed the preferred token budget. Shared tokenization improved NE by 0.02% under both tested schemas. Giving each event two tokens produced a 0.10% benefit with unlimited context, but a 0.16% degradation with a limited history window. The pattern favored preserving more history when context was tight. A separate initialization test reported a 0.83% NE improvement with the LLM frozen, a further 1.04% after unfreezing, and a stable 0.20% advantage for pre-trained initialization over random initialization.
The paper also examined representation drift, meaning changes in the learned token space between checkpoints. Over a 20-day checkpoint gap, its best reported alignment method, DANN, produced a mean cosine similarity of 0.956 and k-nearest-neighbor accuracy of 0.856, alongside a 0.02% lower NE than the version without regularization. In this analysis, higher cosine similarity and lower checkpoint separability indicated better alignment. The drift result was therefore paired with no reported loss in the tested prediction score.
A result from full traffic
The paper then reports a large-scale evaluation on a full-traffic Facebook surface. The Event Tokenizer used 3.84 million dense parameters and about 200 GB of sparse embedding tables while encoding billions of events per day. Cached Event Tokens were added as historical features to the existing Incumbent ranker, a non-LLM architecture. The evaluation reported a 0.06% NE reduction. The authors state that a 0.02% reduction is considered statistically significant on their internal datasets and sufficient for measurable revenue impact, but the supplied analysis gives no p-value, confidence interval, sample count, or direct online revenue estimate.
What the study leaves open
The findings are best read as evidence about model quality and compute trade-offs inside the evaluated recommendation system. The evidence comes from internal industrial logs and offline comparisons, plus the internal full-traffic evaluation. Main results used the full dataset, while ablations used 10%; serving and feature-materialization costs were estimated on an infrastructure-dependent scale. Asynchronous tokenization introduces feature delay, and end-to-end LLM ranking-serving bottlenecks were outside the stated scope. The paper reports no confidence intervals or p-values for its small NE differences, and it does not establish effects on clicks, retention, or revenue. The document is arXiv:2608.25546v1, dated 26 August 2026, with no journal publication identified in the supplied metadata.
Paper data and sources
Original title: An Event is Worth One Token: Event Tokenization for Industrial-scale LLM Recommendation
Authors: Fan Xia, Zhaoheng Zheng, Iman Setayesh et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text