Kuaishou Proposes Streaming Semantic User Profiling Framework

_reachsumit · x · 2026-07-20

Kuaishou has proposed a recommendation system framework named RECAP, specifically designed for short video recommendation scenarios.

The core of this framework lies in utilizing Large Language Models (LLMs) to maintain bounded structured user profiles, supporting streaming updates. Its key highlight is a feedback-driven mechanism that uses feedback signals aligned with recommendation goals to optimize the LLM's profile updating process, thereby capturing dynamic user interests in short video scenarios more accurately.

Original post →

More from Research

Research channel →