Alibaba researchers describe DREAM, an agentic control architecture that adds a perception-aware policy layer above existing recommendation pipelines, comprising an Intent Engine that fuses user signals into structured representations and a Meta Engine that performs layered reasoning for strategy planning and parameter translation. In large-scale A/B tests on Taobao’s homepage feed, re-ranking control alone improved item page views by 2.06% and GMV by 0.88%, with further gains when extended to fine ranking reaching 2.71% IPV and 1.31% GMV improvement.
