简体中文 · English

TAGR: Temporally adaptive generative recommendation for livestream ads¶

Fidelity: core-mechanism reproduction. Runs refreshable LSIDs with a stable vocabulary, multiscale intent, and behavior/value branches. Production online training and ad systems are not reproduced.

Paper information¶

Field	Details
Paper	arXiv 2608.24034
Institution	Kuaishou Technology
First public release	2026-08-25 (arXiv v1)
Author code	No author-released code found (checked 2026-08-26).
Local adapter or method key	`tagr`
Local reproduction code	`src/auto_research/reproductions/tagr/`

Original paper summary¶

Background and main changes¶

Livestream content, products, and feedback change quickly. Static semantic IDs, single-timescale histories, and off-policy preference optimization can become stale. TAGR adapts at three levels: LSID periodically refreshes semantic and collaborative assignments without changing the token vocabulary; IAG blends short-, medium-, and long-term intent; IOPO runs on-policy GRPO intermittently and continues NTP between bursts, preventing online RL from overwriting the behavior distribution.

flowchart LR
  A[Live content and collaborative behavior] --> L[LSID dynamic assignment / stable vocabulary]
  H[Multiscale user history] --> I[IAG intent encoder]
  L --> G[Autoregressive candidate generation]
  I --> G
  G --> B[BA-GRPO behavior alignment]
  G --> V[VA-GRPO value alignment]
  B --> O[IOPO burst + NTP maintenance]
  V --> O

Key figure from the paper¶

Figure 2: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ z_u=\operatorname{Fuse}(h_u^{short},h_u^{medium},h_u^{long}),\qquad p_\theta(c\mid u)=\prod_t p_\theta(c_t\mid c_{<t},z_u). \]

\[ \mathcal L_{IOPO}=\mathcal L_{NTP}+\lambda_b\mathcal L_{BA-GRPO}+\lambda_v\mathcal L_{VA-GRPO}. \]

Offline and online results¶

The system serves more than 400 million daily active users. Against production DLRM, online livestream entry rate rose +8.5%, cart click rate +7.4%, and revenue +16.1%. This is the concrete evidence for including it in the industrial implementation queue.

Local reproduction¶

Local comparison: A blend of first-order transitions and popularity is the baseline. The experimental arm adds two-level LSID, three-timescale intent, and bounded behavior/value gates. On MovieLens-1M, NDCG@10 falls from 0.1123 to 0.0821 (−26.86%), while head share falls from 0.7275 to 0.1303.

The negative result shows that forcing emphasis on freshness or long-tail items can hurt accuracy in the stable movie domain. It cannot be extrapolated to time-varying livestream-ad gains. See metrics/movielens-1m-seed42.json.

auto-research reproduce --paper tagr --dataset-dir data --seed 42

Scope¶

The local run does not use private Kuaishou logs, an online reward model, real GRPO bursts, ad value labels, or real-time serving. It verifies the three adaptation mechanisms and their failure boundary on non-time-varying public data.