DREAM Improves Taobao Feed Metrics Without Replacing Rankers

Large recommender systems usually split retrieval, ranking, and re-ranking into separately optimized stages. That design is efficient, but it makes session-level changes in a shopper's goal difficult to carry across the pipeline, and policy changes often become brittle collections of rules. DREAM adds a control layer above those existing stages rather than proposing a replacement ranking model: it infers structured intent, converts it into bounded strategy parameters, and learns from offline and online feedback.

Core Contribution

The paper's central contribution is agentic meta-control for a production recommendation stack. Its distinction from an end-to-end recommender is operational: the underlying retrieval and ranking services remain in place, while DREAM observes user signals and selects parameterized interventions through a unified outlet with safety guardrails. This is a more deployable proposition than asking an established feed system to replace its models, provided the control interface is expressive enough to affect outcomes.

Figure 1 contrasts this optimization arrangement with prior recommendation paradigms: the proposed system places perception, planning, execution, and feedback around the existing pipeline rather than collapsing the pipeline into one agent.

Technical Approach

The Intent Engine represents a user at three levels. L0 holds relatively stable profile and long-horizon behavior; L1 captures the active demand, including category, scenario, audience, temporal context, and confidence; L2 records fine-grained preference, such as brand tendency, price preference, decision state, and real-time psychology. The Main Agent is a synchronous 0.8B language model. It updates the active intent list from the L0 profile, prior intents, and a compact current behavior pack, then decides whether the case merits asynchronous refinement.

That split is deliberate. A dual-layer router combines rules and a model confidence signal, sending only selected cases to two specialized 4B agents: a context subagent for uncertain category or intent structure, and an expert for unresolved item-level preferences. The report says roughly 6.3% of requests enter this asynchronous tier, while the edge-cloud trigger chain reduces reporting volume to about 8.7%. The resulting design reserves larger-model work for ambiguity rather than putting it on every serving request.

DREAM also runs a scheduled Dreaming pass over a full-day behavior trajectory. Unlike online updates, which can only insert or update an intent, Dreaming applies six traceable operations: KEEP, CORRECT, ENRICH, MERGE, ADD, and KILL. These explicitly address stale, fragmented, or missing intent records. A self-evolution loop then uses specialized-agent outputs to distill the Main Agent on its own generated intent trajectories, while routing itself is not supervised by that objective.

The Meta Engine translates intent into an M1 baseline and six ordered M2 modules. These can adjust weights for CTR, IPV, CVR, and GMV; alter card-type and category preferences; impose exposure or density constraints; and switch selected slots to pure-CTR ranking. The allowed controls are discrete and bounded, which makes the proposal easier to audit than unconstrained text-generated actions.

Results and Analysis

The strongest product evidence is the Taobao homepage-feed A/B test. With re-ranking control only, the authors report +2.06% IPV, +2.39% Core IPV, and +0.88% GMV. Extending control into fine ranking raises those gains to +2.71%, +3.06%, and +1.31%, respectively; PV improves by more than 1% in both settings. The incremental GMV gain is modest in percentage terms, but it matters more than a generic offline ranking score because the intervention is evaluated in a live commercial feed and does not require changing the serving models.

The paper also evaluates its intent machinery before downstream optimization. On LLM-as-a-Judge evaluation, routing raises the overall score from 71.32% to 78.20%, and routing plus self-evolution reaches 84.74%. Search-term recall rises from 50.57% to 56.05%; search-filter recall rises from 24.25% to 26.40%, though it falls from 26.79% under routing alone. The authors' own interpretation—that routing and self-evolution contribute complementary gains—is mostly supported, but the filter result shows that better aggregate intent structure does not improve every preference field monotonically.

A matched evaluation of 683 users compares final online intent with the consolidated Dreaming state against the same next-day behavior. The report says the largest gains occur in intent type, priority, and subcategory, while already reliable fields such as temporal node and need category change only marginally. That pattern fits the mechanism: full-trajectory consolidation should help category resolution and organization more than stable fields. The evidence is persuasive for this deployment, but it does not establish that the same policy layer will transfer to other marketplaces, traffic mixes, or objective functions.