Lin Zhu - Researcher

KuaFu: Compressing Long User Behavior into Understanding at Billion Scale

Jiahao Hui^*, Lin Zhu^*†, Yishen Hu, Jingdong Shu, Zetai Jiang, Xining Ran, Ben Tan, Yeshou Cai, Gong Chen, Haijie Gu, Jie Jiang

Preprint, 2026 · arXiv:2609.31045 · cs.IR, cs.CL, cs.LG

arXiv/PDF/HTML

KuaFu overview: each behavior item is encoded once by the item compressor, reduced by a two-axis adapter into a handful of memory tokens, then read by a frozen LLM decoder. — **KuaFu in production.** Each behavior item – a product page, a piece of content, or a long ad context split into trunks – is encoded once by the item compressor and reduced by the two-axis adapter (right) to a handful of memory tokens. These tokens are interleaved with action and intensity tokens and task instructions, and a frozen LLM decoder reads the full strip to emit interest and feature slots. Because the compressed result depends only on the item, it is cached offline and reused across users, tasks and requests.

A unified behavior-compression layer whose minimal unit is one behavior item: each item is compressed into 2–4 tokens of width 128–256, matching or exceeding uncompressed single-task production models on all five headline metrics while raising per-GPU throughput by 37%–350% and saving 190 GPUs. Deployed on the Tencent advertising and recommendation platform for ten months, lifting overall GMV by 1.37%.