A temporally adaptive recommendation system for live-stream advertising reported a 16.1% relative lift in advertising revenue in a randomized production comparison. The same test reported an 8.5% lift in the rate at which users entered live rooms and a 7.4% lift in shopping-cart clicks, compared with DLRM, the production retrieval baseline. It used 10% of production traffic and covered more than 40 million users on Kuaishou's e-commerce live-stream advertising platform.
These are relative changes, not percentage-point increases or cash totals. The paper gives neither the underlying baseline rates nor the revenue amounts, so the absolute size of the change cannot be calculated from the reported figures. It also provides no confidence intervals, p-values or other uncertainty estimates, leaving the precision of the lifts unclear.
A recommender built for changing live rooms
TAGR is a generative retrieval system built around temporal change. Its three main pieces are LSID, which constructs dynamic live-ad tokens; IAG, which represents user intent at multiple time scales; and IOPO, which aligns recommendations intermittently with preferences from the current policy. The design is meant to adapt as live-ad scenes, products and user behavior change.
From logs to live traffic
To test retrieval offline, the researchers used a billion-scale dataset with five days of training logs, two days of validation and two days of testing. The data covered more than 400 million users and hundreds of thousands of live ads. The evaluation checked whether the next-action live ad appeared among the system's top K generated candidates, a measure known as Hit Rate at K.
For the full TAGR configuration, Hit Rate at 128 was 0.7723 for live-room entry (LRE) and 0.6965 for shopping-cart click (SCC). These numbers describe how often the target appeared in the generated shortlist, not the size of the business gain. The authors did not report an offline DLRM HR@K value, considering that comparison not directly comparable with generated Top-K decoding.
What changed inside the model
The component analysis followed the model's assembly line. The reported relative revenue lift was 9.9% with LSID, 11.2% after MSI-Encoding, 13.5% after MF-NTP, 14.7% after BA-GRPO and 16.1% after VA-GRPO. In plain terms, the sequence added dynamic ad tokens, multi-scale intent modeling and preference-alignment stages. These were sequential comparisons, not separate randomized production tests, so they do not isolate the individual production effect of each component.
A token-focused comparison of full LSID with hashing reported a compression ratio of 1.01, a collision rate of 0.02, stability of 90% and a relative revenue gain of 3.8% over static SID. The figures address the token system's compactness, distinctness and consistency, but no uncertainty estimates were reported for this comparison.
The intent ablation compared a multi-stride setup with entry strides of 1, 2 and 10 against full MF-NTP. The multi-stride version reported LRE and SCC HR@128 scores of 0.7276 and 0.6709; MF-NTP reported 0.7468 and 0.6836. These are offline retrieval comparisons, not standalone estimates of revenue.
The paper also compared three preference-alignment update schedules. On its training-loss, reward, stability and retrieval measures, intermittent GRPO reported an NTP loss of 1.53, reward of 0.913, high stability and LRE HR@128 of 0.7723. Continuous GRPO reported 1.68, 0.864, moderate stability and 0.7469, while off-policy GRPO reported 1.79, 0.675, low stability and 0.6982.
Built to answer live requests
The serving engine generated 256 live-ad identifiers through beam search. The reported throughput was more than 2,500 queries per second per L20 GPU, with end-to-end retrieval latency below 100 milliseconds. Those figures describe the paper's reported operating setup; the supplied analysis does not give variability or fuller conditions for the measurement.
The reported segment analysis found a relative revenue lift of 28.8% among low-value users and 18.4% for cold-start live streams. The paper does not give subgroup sizes or uncertainty estimates, so these figures remain comparative findings within the tested setting.
Promising, but still bounded evidence
That tested setting matters. The evidence comes from one industrial advertising environment and a multi-week experiment, so it does not establish how TAGR would perform across other platforms, advertisers, products or user segments, or whether the reported lifts would last longer. The production comparison evaluated the full TAGR package at once, leaving the contribution of any single component unresolved.
TAGR is described in an arXiv version 1 preprint dated 25 August 2026. The supplied document also includes a complementary-experiments section covering preference-optimization backbones and retrieval coverage. The central result remains a relative comparison within that production setting: a reported 16.1% revenue lift, without an absolute revenue amount or uncertainty estimate to show how large or durable the gain is.
Paper data and sources
Original title: TAGR: Temporally Adaptive Generative Recommendation for Industrial Live-Streaming Advertising
Authors: Wencai Ye, Guangyi Liu, Chaoyi Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text