Meituan Technical Team Showcases Cutting-Edge AI Agent Research with Top Conference Paper Selections from ASX Team

Key Takeaways

• Strategic Focus on Agentic Systems: Meituan's ASX (Agentic System X) team is dedicated to building a comprehensive technology system centered on Large Language Model (LLM)-based agents.
• High-Impact Academic Contributions: The team has published dozens of high-quality research papers in premier AI conferences, including ICLR, NeurIPS, CVPR, and AAAI.
• Core Research Pillars: The technical depth of the ASX team is concentrated in three primary areas: LLM post-training, Agentic Reinforcement Learning, and multi-modal understanding.
• Practical Application in Search and Recommendation: The research is specifically tailored to enhance the capabilities of search and recommendation systems through advanced agentic frameworks.
• Selected Paper Interpretations: Meituan has curated six specific papers from their recent publications to provide deeper insights into their technical breakthroughs and methodologies.

In-Depth Analysis

The Strategic Vision of Agentic System X (ASX)

Meituan's Business R&D Platform has established the ASX (Agentic System X) team to spearhead the transition from traditional algorithmic models to more autonomous, LLM-based agent systems. This shift represents a significant evolution in how search and recommendation engines operate. Unlike static models that provide outputs based on fixed patterns, an "Agentic System" implies a level of autonomy where the AI can reason, plan, and execute tasks to meet complex user needs. By focusing on an agent-based architecture, Meituan is aiming to create a more interactive and adaptive user experience. The ASX team’s focus on building this system from the ground up indicates a long-term commitment to integrating Large Language Models into the core of Meituan’s technical infrastructure.

The publication of dozens of papers at top-tier conferences like ICLR (International Conference on Learning Representations) and NeurIPS (Neural Information Processing Systems) serves as a testament to the academic rigor and innovation driving the ASX team. These venues are the primary stages for global AI breakthroughs, and Meituan’s consistent presence there highlights its role as a major player in the international AI research community. The specific focus on "Agentic" systems suggests that Meituan is looking beyond simple chat interfaces, moving toward systems that can act as intermediaries between users and the complex array of services Meituan provides.

Core Research Pillars: Post-Training and Reinforcement Learning

A critical component of the ASX team's research is LLM post-training. While base models provide a foundation of knowledge, post-training is essential for refining these models into functional agents. This process typically involves fine-tuning the models to follow specific instructions, adhere to safety guidelines, and operate within the constraints of a search and recommendation environment. By deep-diving into post-training, the ASX team ensures that their agents are not just knowledgeable, but also reliable and aligned with the specific operational requirements of Meituan’s business platform.

Parallel to post-training is the team's work in Agentic Reinforcement Learning (RL). Reinforcement learning is the cornerstone of autonomous decision-making. In the context of an agent, RL allows the system to learn from interactions with its environment—in this case, the user and the search/recommendation database. By optimizing for long-term rewards rather than just immediate clicks, Agentic RL enables Meituan’s systems to develop more sophisticated strategies for user engagement. This research direction is vital for creating agents that can handle multi-step tasks, such as planning a full day of activities or navigating complex service queries that require a sequence of informed decisions.

Advancing Multi-modal Understanding in Search Ecosystems

The third pillar of the ASX team’s research is multi-modal understanding. In a modern digital marketplace like Meituan, information is rarely limited to text. Users interact with images, videos, and structured data across various services. For an AI agent to be truly effective, it must be able to process and synthesize these different modes of information simultaneously. The ASX team’s research in this area, often showcased at conferences like CVPR (Conference on Computer Vision and Pattern Recognition), focuses on bridging the gap between visual and textual data.

By enhancing multi-modal understanding, Meituan’s agents can better interpret user intent when it is expressed through a combination of media. For example, understanding the context of a food photo in relation to a search query for a specific restaurant requires a sophisticated multi-modal approach. This capability is essential for creating a seamless search and recommendation experience that feels intuitive to the user. The integration of multi-modal research into the ASX framework ensures that Meituan’s agents are equipped to handle the rich, diverse data environment of a comprehensive service platform.

Industry Impact

The work being done by Meituan’s ASX team has broader implications for the AI industry, particularly in the realm of commercial AI applications. As the industry moves from "AI as a tool" to "AI as an agent," Meituan’s research provides a blueprint for how large-scale service platforms can implement these technologies. The focus on top-tier academic validation ensures that these advancements are built on a foundation of peer-reviewed excellence, which is crucial for the long-term stability and scalability of agentic systems.

Furthermore, Meituan’s emphasis on the intersection of LLMs, Reinforcement Learning, and Multi-modality reflects the current convergence of AI disciplines. By successfully integrating these areas, the ASX team is demonstrating how complex, real-world problems in search and recommendation can be addressed through a unified agentic framework. This approach is likely to influence how other tech giants approach the development of their own autonomous systems, potentially accelerating the adoption of Agentic AI across various sectors of the digital economy.

Frequently Asked Questions

What is the primary focus of Meituan's ASX team?

The ASX (Agentic System X) team focuses on building a technology system centered on Large Language Model (LLM)-based agents. Their research is primarily applied to the fields of search and recommendation, aiming to create more autonomous and intelligent systems.

Which academic conferences have featured Meituan's recent research?

Meituan's ASX team has published dozens of high-quality research papers in several top-tier AI conferences, including ICLR, NeurIPS, CVPR, and AAAI. These conferences represent the leading edge of global research in machine learning, computer vision, and artificial intelligence.

What are the core research directions of the ASX team?

The team's research is concentrated in three main areas: LLM post-training, Agentic Reinforcement Learning, and multi-modal understanding. These pillars are designed to improve the reasoning, decision-making, and data-processing capabilities of AI agents.