MetaStrategy: Generative Ranking with Executable LLM Strategies

2026-08-10

AI Summary

This work addresses the challenge of ranking heterogeneous content in industrial recommendation systems under multiple objectives, where existing generative approaches struggle to integrate effectively with predictive models, business rules, and experience guardrails. The authors propose MetaStrategy, a novel framework that leverages large language models (LLMs) to generate structured, executable ranking strategies—represented as typed JSON—conditioned on request context. These strategies are instantiated into standalone generators via a deterministic validator and compiler, enabling atomic competition with existing policies within a Generator-Evaluator architecture. The approach flexibly combines multi-objective weights, category preferences, and experience constraints, and employs self-competitive curriculum learning with evaluator-guided reward-augmented distillation for efficient training. In a seven-day A/B test on Taobao’s “Guess You Like” homepage feed, the method achieved a 27.93% call share, improving click PVs, item detail page views, and gross merchandise value by 2.11%, 3.12%, and 2.83%, respectively, without increasing response latency.

Abstract

Industrial recommender systems rank heterogeneous content under coupled user, business, commercial, and experience objectives. Existing generative ranking methods typically construct item sequences directly, making them difficult to integrate with mature predictive models, operational rules, and field-level guardrails. We present MetaStrategy, a framework that instead generates a structured, executable ranking strategy. Conditioned on request context, a large language model (LLM) policy emits a typed JSON bundle controlling objective weights, content and category preferences, experience constraints, and position policies. A deterministic validator and compiler instantiate an isolated Generator that competes atomically with incumbents under the list-level Evaluator of the Generator-Evaluator (GE) architecture. We train the policy in a production-path replay environment that re-executes logged requests through the current re-ranking stack without user exposure. The method combines selection, relative-rank, and baseline-lift rewards, a self-competitive curriculum that feeds frequent strategies back as competitors, and Evaluator-routed reward-augmented on-policy distillation that transfers complementary 4B-parameter Teachers into a compact 0.8B-parameter Student. We deploy MetaStrategy in Taobao Homepage Guess You Like through diff-triggered nearline generation; LLM inference remains outside synchronous ranking, with no observable increase in response time (RT). In a seven-day user-randomized online A/B test, MetaStrategy wins 27.93% of treatment-side GE calls and significantly improves click page views (click PV) by 2.11%, item-detail page views (IPV) by 3.12%, and transaction amount by 2.83%.

Problem

• generative ranking
• recommender systems
• multi-objective optimization
• executable strategies
• heterogeneous content

Innovation

• Generative Ranking
• Executable Strategy
• LLM-based Policy
• Generator-Evaluator Architecture
• Reward-Augmented Distillation