Megadose AI progress, ranked and analyzed.

Retrofitting Code Using LLMs to Support Exceptional Behavior

· ArXiv · AI/CL/LG ·
EXCODER uses tests plus program analysis to add missing exception-handling code to Java methods.

The paper defines the task as restoring exception-related code removed from existing codebases, guided by Exceptional Behavior Tests. Its benchmark covers 304 methods across 75 GitHub Java projects. Paired with Qwen 2.5 Coder 32b, EXCODER reaches 85.92% pass@1 on developer-written tests, about 12.56 points above the baseline. The authors say manual inspection still found limitations in the generated code. ArXiv · AI/CL/LG's note

score 4

Categories: Research