ARIA: Runtime Inspection and Amendment Framework for Multi-Agent Systems

01 ABSTRACT

Local errors in LLM-based multi-agent systems can propagate and cause task failure, while post-hoc diagnosis or repeated execution does not provide a runtime intervention policy. The authors build a dataset of 3,953 verified repair pairs and train a compact model to learn when and how to repair ongoing executions. Experiments show strong repair performance in both post-hoc and runtime settings, with generalization to unseen frameworks and host models.

02 KEY FINDINGS

1. Constructed the first large-scale dataset pairing failed multi-agent trajectories with execution-verified repairs, containing 3,953 repair pairs, each specifying what went wrong, where to intervene, and how execution should continue.
2. Proposed ARIA, an MAS-agnostic runtime inspection and amendment framework designed to integrate with any multi-agent system that exposes execution trajectories and supports continuation replacement.
3. The trained compact model demonstrates strong repair performance in both post-hoc and runtime settings, generalizing to unseen frameworks and host models.