ARISE: A Repository-level Graph Representation and Toolset for Agentic Fault Localization and Program Repair
In the evolving landscape of software engineering, the challenges of fault localization (FL) and automated program repair (APR) are becoming increasingly complex. The recent introduction of ARISE (Agentic Repository-level Issue Solving Engine) presents a significant advancement in this domain, leveraging a novel multi-granularity program graph to enhance the capabilities of existing large language model (LLM)-based agents.
The Challenge of Existing Systems
Traditional graph-based systems primarily focus on the structural aspects of code repositories. They illustrate relationships between files, classes, and functions but often fall short in capturing the nuances of variable flow within procedures. This limitation hinders agents from achieving the semantic precision necessary for effective function- and line-level localization. As a result, the efficiency of fault localization and program repair processes remains suboptimal.
Introducing ARISE
ARISE addresses these shortcomings by integrating a multi-granularity program graph that extends beyond basic structural relationships. Key features of ARISE include:
- Statement-Level Nodes: The program graph connects nodes at the statement level, allowing for a more detailed view of code execution.
- Intra-Procedural Definition-Use Edges: This feature highlights how variable values flow through procedures, facilitating more accurate tracing of variable definitions and usages.
- Three-Tier Tool API: ARISE exposes its graph through a comprehensive API that enhances data-flow slicing, making it a first-class and queryable primitive for agents.
Evaluation and Results
ARISE was evaluated using the SWE-bench Lite, which comprises 300 real GitHub issues across 11 Python repositories. The performance metrics reveal significant improvements when using ARISE compared to the standard SWE-agent baseline:
- Function Recall@1: Improved by 17.0 points.
- Line Recall@1: Enhanced by 15.0 points.
- Repair Success Rate: Achieved a Pass@1 rate of 22.0%, representing a 4.7 percentage-point increase over the SWE-agent.
These localization advancements directly correlate with enhanced repair success, illustrating the effectiveness of the ARISE approach in practical scenarios.
Controlled Ablations and Future Directions
Controlled ablation studies indicate that the notable performance improvements stem primarily from the data-flow graph rather than the tool schema itself. Additionally, it was found that large code models can directly utilize structured slice outputs, eliminating the need for a natural-language summarization layer. This finding emphasizes the potential for ARISE to streamline the integration of advanced models into the fault localization and program repair workflow.
A Framework-Agnostic Toolset
ARISE’s graph builder and slicing API are designed as framework-agnostic tools, positioning them as valuable assets for future research in automated program repair. This adaptability ensures that ARISE can be seamlessly integrated into various development environments, fostering innovation and efficiency in addressing software faults.
As the software engineering field continues to evolve, tools like ARISE represent a crucial step toward more intelligent and effective fault localization and repair processes, ultimately enhancing the reliability and quality of software applications.
Related AI Insights
- Parloa AI Agents Transform Customer Service Experience
- RouteHijack: Exploiting Routing Vulnerabilities in MoE LLMs
- Top Travel VPNs for 2026: Secure & Fast Connections
- ARIS: AI-Driven Autonomous Research with Multi-Agent Collaboration
- Moonshot AI Raises $2B at $20B Valuation Amid Open-Source AI Boom
- Pass-Rate Rewards in Reinforcement Learning for Code Generation
- AI Data Center and Power Grid Co-Design for Sustainability
- Reward Hacking Benchmark: Testing Exploits in LLM Agents
- Cascade Token Selection Boosts Transformer Attention Speed
- Efficient On-Device Bipolar Agitation Detection with MP-IB
