Issue Localization via LLM-Driven Iterative Code Graph Searching

Environment Setup

conda create -n cosil python=3.11
conda activate cosil
pip install -r requirments.txt

Model calls are routed through LiteLLM. Configure credentials and API endpoints with environment variables before running the scripts. For an OpenAI-compatible endpoint, for example:

export OPENAI_API_KEY="<your-api-key>"
export OPENAI_API_BASE="https://<your-endpoint>/v1"

For other providers, use the environment variables supported by LiteLLM and pass the corresponding LiteLLM model name to --model.

How to run?

Preparation

Generate the repository structure by running:

python get_lite_structure.py      # For SWE-Bench Lite
python get_verified_structure.py  # For SWE-Bench Verified

To avoid regenerating repository structure files repeatedly, you can use the cache provided by the Agentless Team. Download here.

Then export the repository-structure location before running localization scripts:

export PROJECT_FILE_LOC="<path to your repo structures>"

You can set this directly in run_lite.sh, run_verified.sh, patch_gen.sh, ablation.sh, and sample.sh, or export it in your shell.

RQ1: Effectiveness

To reproduce RQ1 results for SWE-bench Lite/Verified:

bash run_lite.sh
bash run_verified.sh

Results are stored in the results folder.

RQ2: Ablation

bash ablation.sh

RQ3: Application

bash patch_gen.sh

Then use the official SWE-bench evaluation method to evaluate the generated patches.

RQ4: Generalizability

bash sample.sh

Evaluation

Evaluate localization results on SWE-bench Lite or SWE-bench Verified with:

cd evaluation
python FLEvalNew.py --dataset ["lite"/"verified"] --loc_file ["path to your localization results"]

Acknowledgement

This repository is partially based on OpenAutoCoder/Agentless.

• Agentless
• SWE-Bench
• OrcaLoca