Panorama of Multi-Agent Video Recommendation Systems: From MARL to LLM-Driven Architectural Evolution and Open Challenges

Multi-Agent Video Recommendation Systems (MAVRS) are a new paradigm to address the limitations of traditional single-model recommendation systems. The core is to decompose the recommendation task into multiple specialized agents for collaborative completion. This article reviews its evolution: from early Multi-Agent Reinforcement Learning (MARL)-based systems to today's Large Language Model (LLM)-driven architectures; analyzes agent collaboration modes, and points out open challenges such as scalability and multimodal understanding, demonstrating its potential to develop towards more intelligent, interpretable, and personalized directions.

Traditional single-model recommendation systems optimize static engagement metrics (such as click-through rate and viewing duration) and struggle to adapt to the dynamic needs of modern platforms (users' interests change rapidly, and the content ecosystem evolves daily). Multi-agent architectures, by splitting tasks into specialized agents for collaborative completion, provide more accurate and interpretable recommendation results, becoming a new direction to address these challenges.

MAVRS includes multiple specialized agents: video understanding agents (extracting multimodal features), reasoning agents (inferring users' immediate interests and potential needs), memory agents (maintaining users' long-term profiles and short-term session states), and feedback agents (collecting user feedback and transmitting signals). Collaboration modes include: hierarchical coordination (central controller such as LLM coordination), peer-to-peer collaboration (direct agent communication), market mechanism (bidding matching), and consensus mechanism (fusion of multi-agent suggestions). Different modes are suitable for different scenarios.

MAVRS evolution is divided into two stages:

1. MARL stage: Treating recommendation as a sequential decision-making problem, its advantages are modeling complex interactions and games between agents (such as competition for recommendation diversity), but it faces challenges like high training complexity, low sample efficiency, and poor interpretability.
2. LLM-driven stage: Leveraging LLM's understanding and reasoning capabilities, representative architectures include MACRec (generating interpretable recommendation reasons) and Agent4Rec (LLM as the core controller to coordinate specialized agents). Advantages include interpretability, flexibility (prompt words quickly adapt to scenarios), and knowledge integration (pre-trained knowledge to understand semantics).

MAVRS has a wide range of application scenarios:

• Short video platforms (TikTok/Kuaishou): Need fast response, efficient agent collaboration is key;
• Long video platforms (YouTube/Bilibili): Users make cautious decisions, LLM's interpretability is of great value;
• Educational platforms (Khan Academy/Coursera): Need to consider learning paths, can design "learning path planning agents";
• Live streaming/real-time content: Need real-time performance, require specialized real-time feedback mechanisms.

Open challenges faced by MAVRS:

• Scalability: The growth of user and content scale leads to coordination overhead bottlenecks, requiring dynamic adjustment of agent quantity, hierarchical architecture, or edge computing;
• Multimodal understanding: Fusion of video multimodal information still needs improvement, and LLM's fine-grained video understanding needs to be strengthened;
• Incentive alignment: Agents' local goals need to align with the platform's global goals (such as long-term retention);
• Cold start and long tail: New users/content lack data, requiring the design of specialized cold start agents;
• Fairness and bias: Need to introduce fairness agents for monitoring and correction, but the balance between fairness and efficiency remains to be solved.

Future research directions include: hybrid RL-LLM systems (combining MARL decision optimization and LLM reasoning), lifelong personalization (continuously tracking user interest evolution), self-improving systems (independently discovering deficiencies and improving), cross-platform recommendation (providing consistent experiences). The potential of MAVRS has been initially verified; with the improvement of LLM capabilities and the perfection of collaboration mechanisms, it is expected to bring users a better video consumption experience.