Optimo

This is the artifact for ASE 2026 submission titled "Multi-level Code Optimization via Mixture of Prompts".

Dependency

This project relies on the COFFE evaluation framework, you can download it from Github.

Please also set up your API_KEY and BASE_URL environment variable as required to access the GPT series models.

Dataset

We include the two benchmarks used in the evaluation of the paper in folder datasets/:

• COFFE-Function: datasets/function

• COFFE-File: datasets/file

• Effibench: datasets/effibench

Note that these datasets are formatted as required by the COFFE evalaution framework. Please copy the entire datasets/ folder and replace the corresponding one in the COFFE root folder, and then install COFFE by pip install ..

Code

We include the source code in the folder src/:

• code/mine.py: this file includes the logic of optimization strategy mining. You can directly run it by using python code/mine.py.

• code/opt.py: this file includes the logic of optimizing code using Optimo. You can directly run it by using python code/opt.py.

• code/evaluate.py: this file includes the logic of all metrices introduced in the paper.

• code/utils.py: this file includes the logic of accessing remote models.

Data

We include the Codeforces dataset which Optimo mines optimization strategies from and the mined optimization strategies and efficient API map in folder data/:

• data/codeforces: this folder includes the Codeforces solutions and corresponding small and large test cases.

• data/strategies: this folder includes the optimiztion strategies and efficient API map mined by Optimo.

Eval

We include the evaluation results repored in the paper in folder eval/. You could see the evaluation results of each research question described in the paper.

Results

We include the raw results, i.e., the optimized code, of baselines and Optimo in folder results/. The gt/ folder include the results of human-written code optimization, while the gpt4o/ folder includes the results of LLM-generated code optimization.