Shape Your Feed: LLM Agent System for Conversational Recommendations

Current recommendation systems primarily infer user preferences from implicit behaviors like clicks, often failing to capture explicit, nuanced user interests expressed in natural language. This leads to a disconnect where users struggle to actively steer their content feeds. To bridge this gap, researchers propose Shape Your Feed (SYF), an agentic recommendation framework powered by Large Language Models. SYF features a three-tier architecture. The Perception Flow captures detailed user intent from various inputs, including text, voice, and UI interactions. The Serving Flow then performs real-time, agentic re-ranking and pruning of content candidates, informed by a dynamic Semantic Profile of user preferences. Finally, the Self-Evolution Flow refines the system's behavior using Direct Preference Optimization and an LLM-as-a-Judge ensemble, aligning it with human judgments. Online A/B tests showed significant improvements in feed relevance and user sentiment.