The RecGPT-V3 technical report presents a stateful, hybrid-modal recommender designed for Taobao. Its Memory Hub continually distills long-term behavior into structured user-memory units, reducing user-modeling computation by 55.8%. A hybrid foundation model jointly reasons over natural-language tags and Semantic IDs (SIDs), connecting open-world intent understanding with concrete item grounding. Latent Intent Reasoning compresses verbose rationales into learnable latent tokens while retaining decodable explanations, reducing output-token cost by 200x. In large-scale online tests in Taobao’s “Guess What You Like” feed, the report states gains of 1.28% IPV, 1.00% CTR, 1.97% TC, and 3.97% GMV, alongside a 52.4% reduction in end-to-end serving resource consumption.
No heat snapshots are available in the last 24 hours.