Moving Beyond Static Recommendation Feeds

Traditional recommendation systems often rely on static, batch-processed algorithms that struggle to incorporate immediate user feedback or nuanced preferences. The 'Shape Your Feed' framework introduces an agentic approach to this problem, leveraging Large Language Models (LLMs) to act as an intermediary between the user and the recommendation engine. By treating the feed as a conversational interface, the system allows users to refine their content streams through natural language, effectively turning the recommendation process into a dynamic, iterative dialogue.

The Agentic Architecture

The core of the system is an agentic loop that manages the interaction between user intent and content retrieval. Instead of a single-shot prediction, the agent performs three critical functions:

  1. Intent Interpretation: The LLM parses natural language input to extract specific constraints, preferences, or feedback on existing items.
  2. Dynamic Retrieval Strategy: Rather than relying solely on pre-computed embeddings, the agent translates user intent into actionable queries that can adjust the retrieval parameters or re-rank the current feed.
  3. Feedback Integration: The system maintains a stateful context of the conversation, allowing it to learn from negative feedback (e.g., 'I don't like this type of content') and positive reinforcement in real-time, significantly reducing the 'cold start' friction for new user preferences.

Practical Implications for Industrial Systems

Accepted for the RecSys 2026 Industrial Track, this approach addresses the scalability and latency challenges typically associated with LLM-integrated systems. By offloading complex reasoning to the agent while keeping the heavy-lifting retrieval tasks optimized, the system demonstrates that conversational recommendation can be performant enough for production environments. The primary trade-off is the increased complexity in managing state and context across long-running sessions, but the benefit is a significantly higher degree of user agency and satisfaction compared to traditional 'black-box' recommendation models.