spark-code

Code and experiments for the Self-Improving AI Code Generation Through Co-Evolving Policy and Reward. This repository adapts the SPARK framework (Liu et al., 2025) from mathematical reasoning to code generation, where execution provides both deterministic verification and rich diagnostic feedback.

Quick start

git clone https://github.com/amarsaikhanb/spark-code.git
cd spark-code
pip install -e .

# Smoke test (no training)
python scripts/run_experiment.py --gpu 0 --condition C --output ./smoke \
    --iterations 1 --rollouts 4 --train-problems 20 --eval-samples 2 \
    --skip-frontier --no-wandb --reflection-eval-problems 10

To reproduce the headline runs:

# Condition A: exec-only GRPO baseline
python scripts/run_experiment.py --config configs/condition_A.json --gpu 0

# Condition C-regular: SPARK-style co-evolution (best run)
python scripts/run_experiment.py --config configs/condition_C_regular.json --gpu 1

What's in this repo

spark_code/        # importable Python package (data, sandbox, model, training, eval, utils)
scripts/           # CLI entrypoint and bash wrappers
configs/           # JSON configs for the three published runs
results/           # coming soon
tests/             # pytest suite (sandbox, extraction, metrics)

Citation

If you use this code or build on this work:

@misc{batjargal2026sparkcode,
  author = {Batjargal, Amarsaikhan},
  title  = {Self-Improving AI Code Generation Through Co-Evolving Policy and Reward},
  year   = {2026},
  url    = {https://github.com/amarsaikhanb/spark-code}
}

License

MIT. See LICENSE.