AI gaining the power of self-design will have significant impacts. Researchers have proposed a new system called "Agent Design Automation System" (ADAS), allowing meta-agents to automatically construct powerful AI peers using search algorithms.
The ADAS system includes three key components:
-
Search space: defines all possible agent systems that ADAS can create
-
Search algorithm: the method ADAS uses to find excellent agent designs in the search space
-
Evaluation function: used to judge the quality or performance of created agents
Researchers proposed the "meta-agent search" algorithm to implement this idea. At its core, it uses foundation models as meta-agents to iterate new agents based on an ever-expanding database. In theory, meta-agents can program any possible building blocks and agent systems from scratch.
In experiments, meta-agent search performed excellently in multiple benchmarks:
- Outperformed SOTA manually designed agents in the ARC challenge
- Significantly surpassed manually designed agents in reading comprehension and mathematics
- Also outperformed baselines in multi-task and scientific domains
This research demonstrates the potential for automatically designing increasingly powerful agent systems, opening new directions for AGI development. The ability of AI to self-design and iterate may accelerate the progress of AI technology.