This new AI research introduces a novel agent-oriented planning framework for multi-agent systems (MAS), focusing on task decomposition, allocation, and execution in environments involving multiple LLM-empowered agents. The framework centers around a new SUPER-agent that decomposes user queries into sub-tasks, assigns them to appropriate agents, and manages the collaborative process of generating a complete solution.
The key innovation lies in the integration of three design principles-solvability, completeness, and non-redundancy-which guide the decomposition and allocation processes, ensuring that sub-tasks are not only solvable by individual agents but also collectively sufficient to address the full query. A reward model is introduced to efficiently predict the performance of agents in handling sub-tasks without requiring the actual execution of tasks, which significantly enhances the framework’s scalability and efficiency.
The study further incorporates mechanisms for dynamic refinement of the decomposition process through re-describe and plan-in-detail methods. These methods handle unsolvable sub-tasks by either refining their descriptions or breaking them down into smaller, more manageable components. Extensive experiments demonstrate the superiority of the proposed framework over baseline methods in both single-agent and multi-agent systems, with a notable improvement of over 10% in accuracy compared to single-agent systems.
The inclusion of a feedback loop ensures ongoing enhancements to the SUPER-agent’s decision-making process, promoting robustness and adaptability in real-world problem-solving. This approach opens new pathways for more flexible, adaptive, and efficient multi-agent collaboration in complex, dynamic environments.
All rights w/ authors:
AGENT-ORIENTED PLANNING IN MULTI-AGENT SYSTEMS
arxiv.org/pdf/2410.02189
Future You: A Conversation with an AI-Generated
Future Self Reduces Anxiety, Negative Emotions,
and Increases Future Self-Continuity (MIT, Harvard)
arxiv.org/pdf/2405.12514
00:00 Complex queries and the new SUPER-AGENTS
03:49 Hard-to-Decompose Queries
05:20 SUPER-AGENT allocates sub-tasks to agents
07:35 NEW Reward Model evaluates Agent-subtask fit
10:32 What do do with unsolvable tasks
14:48 Hidden dependencies of Agents and SUPER-AGENT
18:51 Reduce the complexity (if possible)
21:20 Agent-oriented Planning in multi-agent systems
22:40 AI and Psychology (MIT and Harvard)
25:00 My live session w/ MITFutureYou (personal data)
#airesearch
#chatgpt
#aiagents
6 окт 2024