DREAM Technical Report

2026-08-10

AI Summary

This work addresses critical limitations in industrial recommendation systems—such as information fragmentation, rigid rule-based designs, and insufficient real-time intent awareness—by introducing DREAM, a novel architecture that operates atop existing models without replacement. DREAM features a programmable and auditable policy control layer, pioneering an agent-based control paradigm grounded in three-tiered intent modeling (L0/L1/L2) and meta-model reasoning. It enables end-to-end autonomous optimization through协同 operation of an intent engine and a meta-engine, integrating on-device signals, edge-cloud trigger chains, and memory-driven hierarchical reasoning (M1–M3). A dual-loop reward mechanism dynamically schedules policies and hyperparameters. Large-scale A/B tests on Taobao’s homepage demonstrate that DREAM boosts IPV by 2.06% and core IPV by 2.39% in reranking alone, with GMV up 0.88%; extending to fine-ranking further improves gains to 3.06% (IPV) and 1.31% (GMV), while PV sustains growth exceeding 1%.

Abstract

Industrial recommender systems commonly use cascaded retrieval, ranking, and re-ranking pipelines. Although efficient, these pipelines fragment information and objectives across modules, rely on rigid rules, and have limited awareness of real-time intent, leaving session-level shifts among browsing, comparison, and purchase insufficiently addressed. We present DREAM (Developing Recommender Engine with Agentic Methods), an autonomous optimization control architecture that adds a perception-aware, orchestrable, and auditable policy layer atop existing pipelines without replacing them. DREAM has two core components. First, a three-tier Intent Engine fuses on-device signals into structured L0/L1/L2 intent representations; its edge-cloud trigger chain reduces reporting volume to approximately 8.7%. Second, a Meta Engine uses a MetaModel for layered M1-to-M2-to-M3 reasoning: intent summarization, strategy planning informed by Strategy Memory, and parameter translation. It dispatches the resulting parameters through a unified outlet with safety guardrails. A Reward Dual Loop continuously optimizes both components by combining offline simulation for strategy-space exploration with online feedback for outcome calibration, forming a cycle of generation, execution, evaluation, and experience accumulation. Large-scale A/B tests on Taobao's homepage feed show that re-ranking control alone improves IPV by 2.06%, Core IPV by 2.39%, and GMV by 0.88%. Extending control to fine ranking raises these gains to 2.71%, 3.06%, and 1.31%, respectively, while consistently improving PV by more than 1%. These gains require neither replacement of pipeline models nor compromise of serving stability, supporting agentic meta-control as a viable paradigm for industrial recommendation.

Problem

• industrial recommender systems
• cascaded pipelines
• real-time intent awareness
• session-level intent shifts
• recommendation optimization

Innovation

• agentic recommendation
• intent-aware control
• meta-engine
• reward dual loop
• industrial recommender systems