ExCoder

About

This repo hosts the code and data for the following ISSRE 2026 paper:

Retrofitting Code Using LLMs to Support Exceptional Behavior

Authors: Linghan Zhong, Jiyang Zhang, Jayanth Srinivasa, Junyi Jessy Li, Milos Gligoric

@inproceedings{ZhongETA26excoder,
  author = {Zhong, Linghan and Zhang, Jiyang and Srinivasa, Jayanth and Li, Junyi Jessy and Gligoric, Milos},
  title = {Retrofitting Code Using LLMs to Support Exceptional Behavior},
  booktitle = {IEEE International Symposium on Software Reliability Engineering},
  year = {2026},
}

Data

All the data and results generated for the paper are available at EngineeringSoftware/Excoder on Hugging Face. Download them into _work/data and _work/results to reproduce the paper without re-running the data or experiment pipelines.

_work/ is the scratch root for everything the pipelines generate. It is not version controlled and does not ship with the repository, so create it before the first run:

mkdir -p _work

Setup

The Docker image brings up Python, Java 8, Maven 3.8.3 and llama.cpp together. It is the recommended path for the data and experiment pipelines, which build third-party Maven projects that need exactly those versions.

docker compose run --rm excoder                              # with GPU
docker compose -f docker-compose.cpu.yml run --rm excoder    # without

You land in /workspace with the environment active. Export AZURE_END_POINT and AZURE_API_KEY beforehand and compose forwards them in.

To install locally instead, with uv and Python 3.11:

uv sync
source .venv/bin/activate

That covers the qualitative and paper pipelines. The other two also need:

• Java 8 and Maven 3.8.3, to build the subject projects.
• llama-server from llama.cpp on PATH, plus a GPU. The local models are served over HTTP.
• AZURE_END_POINT / AZURE_API_KEY for gpt-5-mini.
• GGUF_DIR pointing at a directory for the model weights.

Building the PDF needs latexmk, pdflatex and bibtex.

If you use direnv, .envrc puts the repository paths and the virtualenv in every shell inside the tree; the pipelines do not need it.

Running from scratch

Four scripts, each of which runs all of its stages when given no arguments. Pass --help to any of them for the individual stages and options.

Rebuild the dataset

scripts/run_data_pipeline.sh --mock    # print the plan, verify the inputs
scripts/run_data_pipeline.sh           # for real

Re-run the experiments

export GGUF_DIR=/path/to/gguf
scripts/run_experiment_pipeline.sh --mock              # print the plan
scripts/run_experiment_pipeline.sh --download-models   # for real

Generate paper tables

scripts/run_qualitative_analysis.sh
scripts/run_paper_pipeline.sh