Online repository for Paper "AgentFL: Scaling LLM-based Fault Localization to Project-level Context"

Summary

This is the online repository for the ICSE'25 under review paper "AgentFL: Scaling LLM-based Fault Localization to Project-level Context".

Environment

Defects4J-V1.4.0 (Note that the buggy items in V1.4.0 is identical with V1.2.0, we use V1.4.0 to avoid some problems in V1.2.0)

Defects4J-V2.0.0

Python version >= 3.8.5

Defects4J Mod

Before running AgentFL, please apply the files under the AgentFL/Defects4J_mod directory to modify your Defects4J V1.4.0/V2.0.0.

Run AgentFL

It's easy to run AgentFL for localizing a bug with the following command:

python3 run.py --config --version --project --bugID --model

For example:

python3 run.py --config Default --version 1.4.0 --project Closure --bugID 26 --model GPT_3_5_TURBO

More configs can be seen under the directory AgentFL/Config

Results

We release all of the results of AgentFL in the AgentFL/Results directory, including the evaluation results on Defects4J V1.4.0/V2.0.0 and the ablation study result.

In the result of each bug, we record all of the prompts, responses, and intermediate outputs.

Human Evaluation Results

The human evaluation results can be found in the file AgentFL/EvaluationResult/DebugResult_d4j140_GPT35_human.xlsx