A preprint reports that a method which compresses each temporal snapshot into a compact Event Token recorded the lowest reported Normalized Entropy, or NE, in its main ranking comparison. AMBER improved NE by 0.40% against an Incumbent (fair) ranker matched on features, FLOP budget and training target. The gains were 0.30% against a Pointwise model, 1.00% against an HSTU-style input and 1.60% against Semantic IDs plus content-understanding embeddings. NE is an offline proxy for online ranking quality, so the result is a comparison of model metrics rather than a direct measure of user behavior.
The study asks whether event-centric representations with higher snapshot resolution can serve as a useful scaling dimension for industrial recommendation. AMBER turns each temporal snapshot into a compact Event Token that becomes a new input modality for an LLM. The representation is learned end to end, then precomputed and cached, decoupling snapshot resolution from real-time serving computation. Main results used the full chronological industrial impression-log dataset, while ablation studies used a 10% subsample.
The reported gains extended across tasks
On retrieval, AMBER maintained a 0.31% to 0.51% Soft Recall improvement over the Incumbent across the reported checkpoint ages. It also outperformed CU embeddings by 0.28 to 0.32 percentage points. The primary controlled ranking comparison used an Incumbent (fair) baseline matched to AMBER on features, FLOP budget and training target.
The results also addressed a practical complication of cached features: delay. Blending AMBER with the Incumbent produced an Ensemble NE reduction of 0.10% under a 10-minute delay and 0.16% under a 1-minute window. The test connects the method's cached-token design with the timing conditions under which those representations would be used.
What the ablations reveal
The ablations point to the information the system depends on. Removing all item features was associated with a 6.65% NE degradation relative to full AMBER. A shared tokenizer improved NE by 0.02% under both the four-token and two-token schemas, which the paper says is consistent with positive transfer across entity types. These were model ablations, so they identify performance differences between configurations rather than establishing a causal effect.
The paper reports a favorable shift in the compute-quality Pareto frontier relative to alternative recommendation paradigms. In plain terms, AMBER reached a better reported balance between recommendation quality and computing cost. It also reports that Event Tokenizer capacity could be scaled independently, improving downstream performance with minimal impact on real-time serving compute. That separation is central to the proposal to use richer event snapshots as a scaling dimension.
A test of drift across time
The authors tested drift across time as well. In a 20-day checkpoint-gap test, DANN reported a mean cosine similarity of 0.956, k-nearest-neighbor accuracy of 0.856 and an NE change of minus 0.02% compared with no regularization. These are checkpoint-alignment and model-metric results, not direct measures of user outcomes.
The evidence has a narrower reach
In the large-scale evaluation, Event Tokens used with the Incumbent ranker were associated with an approximate 0.06% reduction in NE. The authors state that a 0.02% change is statistically significant and sufficient to yield a measurable conversion lift. The supplied analysis does not include a p-value, confidence interval or formal test procedure, and it does not report a directly measured conversion outcome. The conversion claim is therefore an interpretation attached to the metric result, not an observed user-level result.
That distinction sets the limits of the evidence. The ranking measure is an offline proxy, and the reported gains are comparisons among model configurations on chronological industrial impression logs. They do not by themselves establish higher conversion or other online user outcomes. The main results used the full dataset, but the ablations used a 10% subsample, so the reported evidence combines full-data headline results with smaller diagnostic tests.
AMBER's advantage is therefore best read as evidence about the tested offline metrics, not as proof of a real-world conversion benefit. The work is presented as an arXiv preprint identified as arXiv:2608.25546v3.
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