📝 Paper Summary
Software Engineering Agents
Automated Program Repair
iSWE is a specialized Java coding agent that decomposes issue resolution into localization and editing sub-agents, utilizing rule-based static analysis and containerized compilation to ensure correctness.
Core Problem
Most automated software engineering agents are optimized for Python and perform poorly on Java due to its strict type system, compilation requirements, and verbose object-oriented structure.
Why it matters:
- Enterprise software relies heavily on Java, yet existing leaderboards (like SWE-bench) and agents focus disproportionately on Python
- Java's compiled nature and multi-file dependencies make Python-centric approaches (which often rely on simple linters or bash scripts) ineffective or unsafe
- Generic agents often fail to localize bugs in large Java repositories where functionality is spread across many files
Concrete Example:
In an `apache/rocketmq` issue, a generic agent might fail to identify that a fix requires editing a second file not explicitly mentioned in the bug report. iSWE's localization agent uses call-graph tools to find this dependency.
Key Novelty
Language-Aware Two-Stage Agency (iSWE)
- Decomposes the task into two specialized ReAct (Reasoning + Acting) sub-agents: one dedicated solely to finding where to edit (Localization) and one to applying changes (Editing)
- Replaces generic text-processing tools with strict, rule-based Java static analysis tools (via CLDK) to provide the LLM with reliable symbol, inheritance, and call-graph information
- Validates edits using a containerized Java compiler and build environment rather than just a linter, ensuring type safety before submission
Architecture
The two-agent pipeline showing the flow from Issue Description to Final Code
Evaluation Highlights
- Achieves state-of-the-art issue resolution rates on the Java splits of both Multi-SWE-bench and SWE-PolyBench [Specific percentages not included in input text]
- Reduces LLM inference costs by 2x to 3x compared to leading agents by using efficient, language-aware tools that require fewer interaction turns
- Demonstrates high localization capability, successfully identifying necessary code locations not mentioned in issue descriptions (e.g., in `apache/rocketmq`)
Breakthrough Assessment
7/10
Strong engineering contribution applying language-specific rigor to coding agents. Addresses a neglected domain (Java) with appropriate tooling, though the core architecture (ReAct) is standard.