Survey Proposes LIFE Framework for LLM Multi-Agent Systems
A new survey from arXiv (2605.14892) examines the challenges and progression of LLM-based multi-agent systems. While individual agents excel at reasoning, planning, and tool use, they falter in tasks requiring sustained coordination. Multi-agent systems enable structured collaboration but introduce risks such as error propagation across agents and rounds, leading to failures that are hard to diagnose and rarely result in structural self-improvement. Existing surveys treat individual capabilities, collaboration, and self-evolution separately, missing causal links. The paper proposes the LIFE progression: Lay capability foundation, Integrate agents through collaboration, Find faults via attribution, and Evolve through autonomous self-improvement. This unified review aims to address the gap by connecting these stages causally.
Key facts
- arXiv paper 2605.14892 surveys LLM-based multi-agent systems
- Individual agents strong in reasoning, planning, tool use but limited in sustained coordination
- Multi-agent systems enable structured collaboration among specialized agents
- Tighter coordination amplifies error propagation across agents and interaction rounds
- Failures are difficult to diagnose and rarely lead to structural self-improvement
- Existing surveys cover individual capabilities, collaboration, or self-evolution separately
- Proposes LIFE progression: Lay, Integrate, Find, Evolve
- LIFE stands for Lay capability foundation, Integrate agents, Find faults, Evolve through autonomous self-improvement
Entities
Institutions
- arXiv