Interactive LLM Agent Review for Software Engineering Artifacts

Интерактивная проверка артефактов программной инженерии агентами на основе больших языковых моделей
Dae-Kyoo Kim
2026-03-30

agent collaborationartifact validitygraphical modelsinteractive LLM agentssoftware engineering artifacts
Large language model (LLM)-based agents are increasingly used in software engineering, yet the quality of generated artifacts, particularly complex graphical models, remains inconsistent and often below professional standards. Most existing approaches are limited to code generation and testing, providing little support for broader engineering tasks. This paper proposes an interactive review collaboration approach in which task-specialized LLM agents iteratively exchange feedback to refine artifacts across the development lifecycle. We evaluate this approach against non-interactive pipelines using a validity metric across three case studies: Tour Online Reservation System, Smart Wallet System, and Food Order and Delivery System. Results show that interactive collaboration improves artifact validity by an average of 41.26%, with gains in domain modeling (31.03%), sequence diagrams (72.73%), class diagrams (65.63%), implementation (462.5%), and testing (76.67%). The results demonstrate robust, cross-domain benefits, highlighting interactive collaboration as an effective way to mitigate LLM limitations, though performance varies by task and domain.
1
Across three case studies, interactive collaboration improved artifact validity by an average of 41.26% compared with non-interactive pipelines.
2
Interactive collaboration demonstrated robust cross-domain benefits, but its performance varied depending on the task and application domain.
3
The approach addresses broader software engineering tasks than code generation and testing, including refinement of complex graphical models across the development lifecycle.
4
The paper introduces an interactive review collaboration approach in which task-specialized LLM agents iteratively exchange feedback to refine software engineering artifacts.
5
Validity gains varied by artifact type: 31.03% for domain models, 72.73% for sequence diagrams, 65.63% for class diagrams, 462.5% for implementation, and 76.67% for testing.

software engineering artifacts, particularly graphical models, generated by LLM-based agents across the development lifecycle

the effect of interactive collaboration and iterative feedback among task-specialized LLM agents on artifact validity across modeling, implementation, and testing tasks

Publication Details
Publication Date
2026-03-30
Journal
Publisher
ISSN
Cited by
0
Access Type
Author Information
Authors
Dae-Kyoo Kim
Explore further
Open the scid.ai AI chat with a ready-made request: it will find papers on a similar topic and help build a literature review.
Find similar papers in the chat →
Make a presentation
100%