Recent advancements in
software engineering have seen the rise of Large Language Models (LLMs) that have demonstrated exceptional capabilities in various coding
tasks. While there has been a focus on autonomous software agents based on LLMs for specific Software Engineering (SE) tasks, researchers from FPT Software AI Center, Viet Nam, have introduced HyperAgent, a generalist multi-agent system
designed to address a wide spectrum of SE tasks across different programming languages.
HyperAgent consists of four specialized agents—Planner, Navigator, Code Editor, and Executor—that manage the full lifecycle of SE tasks. Through extensive evaluations, HyperAgent has shown competitive performance across diverse SE tasks:
GitHub issue resolution: With success rates of 25.01% on SWE-Bench-Lite and 31.40% on SWE-Bench-Verified.
Code generation at repository scale (RepoExec): Demonstrating 53.3% accuracy when navigating through codebases and retrieving correct context.
Fault localization and program repair (Defects4J): Achieving an
accuracy rate of 59.70% in fault localization and successful fixes for 29.8% of Defects4J bugs.
The design emphasizes three main advantages over
existing methods:
Generalizability: The framework is designed to easily adapt to a wide range of tasks with minimal configuration changes.
Efficiency: Each agent is optimized to manage processes with varying levels of complexity.
Scalability: The framework is built to scale effectively when deployed in real-world scenarios where the number of subtasks is significantly large.
HyperAgent mimics typical developer workflows by incorporating stages for analysis, planning, feature localization, code editing, and execution/verification. Its success highlights the potential of generalist approaches in software engineering.
Future work could explore integrating HyperAgent with existing development environments and version control systems and enhancing its explainability.
For more information about FPT Software AI Center’s research paper or GitHub page on HyperAgent's capabilities please take a look at their resources connected with this article."
Written by Elias Thalassos
https://futurex.solutions/fpt-software-ai-center-hyperagent-swe-bench-defects4j-performance/?feed_id=367&_unique_id=66e45c99f1651
Comments
Post a Comment