ADIAS: Automated Design of Interactive Agentic Systems
- Source
- arxiv.org
- Author
- Lekang Jiang, Bohan Tang, Stephan Goetz, Yiwen Guo
- Date
Context
ADIAS organizes automated improvement around unresolved problems. Instead of reconstructing what went wrong from a succession of candidate agents, it keeps a persistent record for each issue, including evidence and the outcomes of attempted fixes. That record guides which problem the next code revision should address and how to approach it. Research abstract
The authors report improvements over their strongest baseline across five interactive benchmarks, with gains across four underlying models. Removing the persistent issue state or replacing issue-focused revision with candidate-focused policies reduced performance. These experiments support carrying repair history forward explicitly when a system repeatedly rewrites agent software. Evaluation and component tests
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
Checking sign-in…
Loading comments…