Elith Inc. (Headquarters: Bunkyo-ku, Tokyo; Representative Director: Koki Inoue) is pleased to announce that its research paper on a multi-agent system integrating specialized domains has been accepted at the Workshop on Agentic AI of ICLR (International Conference on Learning Representations) 2025, one of the most prestigious international conferences in the field of machine learning. "ICLR" is, alongside "NeurIPS (Neural Information Processing Systems)" and "ICML (International Conference on Machine Learning)," one of the most prestigious academic conferences in the field of machine learning.
Details here
https://openreview.net/forum?id=5XNYu4rBe4
Paper overview
Background and objective
In recent years, attention has focused on improving the performance of a single large language model (LLM), but effectively integrating deep knowledge across specialized domains has remained a challenge. Elith has developed a new multi-agent system in which multiple AI agents, each specialized in a different field, dynamically collaborate to achieve advanced reasoning that leverages specialized knowledge.
Proposed technology
In this research, we proposed four methods of connecting agents, in which each agent dynamically retrieves and updates knowledge from a database specific to its area of expertise.
- Decentralized: All agents communicate directly with one another, maximizing knowledge sharing between specialized domains.
- Centralized: Information is aggregated through a managing agent to form unified reasoning.
- Hierarchical: High-quality reasoning is achieved by refining problem-solving step by step.
- Shared pool: All agents share conversation history, promoting the reuse of information.
Experiments and results
To verify the effectiveness of this method, we conducted experiments using the academic paper database arXiv. The results showed that the multi-agent system had significantly higher reasoning accuracy and stability than a single-agent model. In particular, even higher performance was achieved in "expert mode," where each agent specializes in its own domain. Dynamically updating knowledge also enabled robust reasoning that always reflects the latest and most relevant information.
Future outlook
These research results suggest that multi-agent systems are effective for integrating specialized knowledge and building consensus on complex interdisciplinary challenges. Elith will continue to scale up the system and develop adaptive capabilities for dynamic connection methods, aiming for practical application across a variety of industries.
[About ICLR (International Conference on Learning Representations)]
ICLR is one of the most prestigious international conferences in the field of machine learning, and along with "NeurIPS" and "ICML," it attracts cutting-edge research results from around the world. ICLR 2025 received approximately 11,500 submissions, with an acceptance rate of 32.08%.