Agent-Oriented Planning in Multi-Agent Systems
Ao Li, Yuexiang Xie, Songze Li, Fugee Tsung, Bolin Ding, Yaliang Li
TL;DR
This paper introduces AOP, a novel framework for agent-oriented planning in multi-agent systems that improves task decomposition, allocation, and robustness, leading to better real-world problem-solving.
Contribution
It proposes a new framework, AOP, with principles and feedback mechanisms, enhancing multi-agent planning efficiency and effectiveness over existing strategies.
Findings
AOP outperforms existing planning strategies in real-world tasks.
The framework effectively decomposes and allocates tasks rapidly.
Feedback loops improve robustness and adaptability.
Abstract
Through the collaboration of multiple LLM-empowered agents possessing diverse expertise and tools, multi-agent systems achieve impressive progress in solving real-world problems. Given the user queries, the meta-agents, serving as the brain within multi-agent systems, are required to decompose the queries into multiple sub-tasks that can be allocated to suitable agents capable of solving them, so-called agent-oriented planning. In this study, we identify three critical design principles of agent-oriented planning, including solvability, completeness, and non-redundancy, to ensure that each sub-task can be effectively resolved, resulting in satisfactory responses to user queries. These principles further inspire us to propose AOP, a novel framework for agent-oriented planning in multi-agent systems, leveraging a fast task decomposition and allocation process followed by an effective and…
Peer Reviews
Decision·ICLR 2025 Poster
The work shows three components which improve the ability of multi-agent LLM systems to improve the results of general LLM queries as compared to current methods. All three components are shown to meaningfully add to the overall model, and directly tackle the presented critical principles for design.
The work is difficult to follow at times, I believe it can be made much clearer, especially with regards to Figure 2. It is difficult to understand this diagram without enough direct context. There is no clear direction that it takes. The detector does not seem to lead to a replan as described. The ordering of the sections, particularly the related work, and how it reads in context at the end, does not make sense, was this just compiled in the wrong place accidentally? The results only have on
* The problem of multi-agent planning is highly important with high potential impact. * How to decompose problems into sub-tasks based on the capabilities of the individual agents is often the key problem. * Learning the capabilities of other agents is a hard and important challenge. * The paper is easy to follow.
* The paper does not relate to the rich and long history of multi-agent planning. A starting point could be [1]. * The design principles seem a bit backwards. Normally, the set of problems that can be solved by a set of agents is defined as the union of the problems that each agent can solve, then you have to check whether the particular problem is within that set. Normally, it is a hard computational problem to determine what problems an agent can solve. If the set of sub-tasks contain redundan
The paper is well-organized, presenting a detailed and thorough explanation of the proposed framework. A particularly compelling aspect is the approach of breaking down tasks into smaller components and assigning them to specialized agents according to their areas of expertise. This decomposition allows each agent to handle tasks within its domain, enhancing overall efficiency and precision. Additionally, the integration of a feedback system significantly boosts the framework’s performance by re
There are several limitations of the work that the authors may consider to address: 1. Expert agents: the difference between each expert agent is their input prompts, but they are using the same underlying model (GPT-4o). It could be more interesting to replace each expert agent with the current state-of-the-art model in their domain. I believe there are plenty of works to fine-tune the language models for code generation, solving math problems, etc. 2. I don't see the significance of the comm
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMulti-Agent Systems and Negotiation