ARIA: Runtime Inspection and Amendment for Multi-Agent Systems
Source
arxiv.org
Author
Bolin Shen, Yushun Dong
Date
Why it matters
Errors in multi-agent runs compound as other agents build on them. A dataset of verified repairs lets you train a model that intervenes mid-run instead of re-running the whole task.
Key takeaways · AI-distilled
ARIA's core idea is to move repair search into data construction and training, so the learned model can intervene at runtime without repeatedly re-executing and testing alternatives.
Each of the 3,953 verified repair pairs specifies what went wrong, where to intervene, and how execution should continue, and is used to train a compact model to decide when and how to correct a run.
The framework is meant to plug into any multi-agentUsing several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.Full definition → that exposes its execution trajectories and supports replacing the continuation of a run.
The authors report strong repair performance in both post-hoc and runtime settings, with generalization to unseen frameworks and host models, though the abstract gives no figures.
Terms in this piece · Glossary
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
articleAgentR A Stateful and Recovery-Aware Software Architecture for LLM-based Auditable WorkflowsRiya Samanta, Bidyut Saha, Soumya Kanti Ghosh, Rajkumar Buyya
articleAppEval: A Unified Benchmark for LLM-Based Mobile Application Repair in ArkTS, Swift, and KotlinBang Xie, Hao Liu, Zhenyu Shi, Yonghao Zhang, Senjian Zhang, Zhiyuan Peng, Xin Yin, Chenhao Ying, Yuan Luo, Wei Chen, Haiming Jin, Shaocong Long, Xu Liu, Zhe Peng
articleEvaluating Agentic Code Repair Capabilities in Distributed SystemsYibo Yan, Huijuan Wang, Junzhou He, Yizhuo Liang, Shaoyu Wang, Huanchen Sun, Seo Jin Park