Self-Evolving Multi-Agent General Intelligence Systems for Autonomous Scientific Discovery and Complex Problem Solving: A Comprehensive Review
Abstract
Artificial General Intelligence (AGI) research is rapidly transitioning from static, pre-trained large language models (LLMs) toward dynamic, self evolving multi-agent systems. These autonomous architectures exhibit unprecedented capabilities in continuous reasoning, adaptive execution, tool refinement, and complex hypothesis generation. This review paper synthesizes the paradigm shift toward self-evolving multi-agent general intelligence systems designed explicitly for autonomous scientific discovery and intricate computational problem-solving. We examine the core architectural abstractions, memory consolidation mechanisms, self-directed code mutation routines, dynamic multi-agent communication topologies, and reinforcement learning strategies that enable continuous adaptation without traditional weight fine-tuning. Furthermore, we analyze key domain applications spanning computational chemistry, automated mathematical modeling, drug discovery, and software engineering. Finally, we establish a standardized benchmarking synthesis, identify foundational system limitations—including recursive collapse, hallucinations, and safety alignment drift—and outline crucial future research trajectories required to realize robust, safe, and truly autonomous scientific intelligence.
KEYWORDS: Self-Evolving Systems, Multi-Agent Intelligence, Autonomous Scientific Discovery, Large Language Models, Dynamic Topology, Meta Prompting, Automated Reasoning.
Full Text:
PDF 97-108Refbacks
- There are currently no refbacks.