RecGPT-V3: A Stateful, Hybrid-Modal Recommender System for Large-Scale E-Commerce
Researchers have introduced RecGPT-V3, a stateful, hybrid-modal recommender system deployed on Taobao that combines natural language reasoning with Semantic IDs (SIDs) to address efficiency and representation bottlenecks. It features a Memory Hub to maintain evolving user memory and Latent Intent Reasoning to compress verbose rationales into compact latent tokens. This development addresses critical production challenges in scaling LLM-based recommender systems, such as high latency and computational overhead. By deploying this system on Taobao, the team demonstrated significant commercial gains (like a 3.97% increase in GMV) while cutting serving resource consumption by over 52%. RecGPT-V3 reduces user-modeling computation by 55.8% using its Memory Hub and lowers output token costs by 200x through Latent Intent Reasoning. The hybrid-modal model allows the LLM to reason jointly over text tags and SIDs, creating a high-bandwidth channel for precise item grounding.
## BACKGROUND
Traditional recommender systems often rely on random ID hashing to represent items, which lacks semantic meaning and can lead to representation bottlenecks. Large Language Model (LLM)-based recommenders improve intent reasoning but typically suffer from high latency due to stateless processing (re-evaluating full history for every request) and verbose chain-of-thought generation.