MetaMuse: Algorithm Generation via Creative Ideation

🎉 Accepted to ICLR 2026!

Purpose: drive algorithm invention rather than imitation by reducing availability bias with a structured creative workflow.

Impact on real system workloads (global cloud provider):

• Cache replacement: up to 35.76% fewer cache misses
• Online bin packing: up to 30.93% reduction in bin usage
• Solution diversity: up to ~1.8× more diverse than baseline LLM approaches

Core ideas:

• Performance-grounded diversity: measure diversity using real performance impact.
• External stimuli steering: use curated keywords to trigger creative reasoning.
• Waypoint reasoning: synthesize executable solutions through structured checkpoints.

conda create -n metamuse python=3.8
conda activate metamuse

pip install -r requirements.txt

Notes: openbox is installed from PyPI by default. For source install, see https://github.com/PKU-DAIR/open-box.

Create src/.env with:

AZURE_OPENAI_ENDPOINT=...
AZURE_OPENAI_API_VERSION=...
AZURE_OPENAI_API_KEY=...
AZURE_OPENAI_MODEL_GPT=...

Optional alternative providers:

AZURE_LLAMA33_ENDPOINT=...
AZURE_LLAMA33_API_VERSION=...
AZURE_LLAMA33_API_KEY=...
AZURE_LLAMA33_MODEL=...

AZURE_DEEPSEEKV3_ENDPOINT=...
AZURE_DEEPSEEKV3_API_VERSION=...
AZURE_DEEPSEEKV3_API_KEY=...
AZURE_DEEPSEEKV3_MODEL=...

Example (BPP Online):

python run.py \
  --algo rsdict \
  --problem bpp_online \
  --trace_folder /path/to/trace_folder \
  --code_folder /path/to/code_out \
  --tot_llm_call_num 50 \
  --hint_word_count 5

Example (Cache):

python run.py \
  --algo rsdict_sf \
  --problem cache \
  --trace_folder /path/to/trace_folder \
  --code_folder /path/to/code_out \
  --capacity 10000 \
  --tot_llm_call_num 50 \
  --hint_word_count 5 \
  --consider_obj_size

All log/code output directories are created automatically when the run starts.

If you notice bugs or run into any issues with the codebase, please feel free to open an issue or submit a PR on this repo.

@article{ma2025algorithm,
  title={Algorithm Generation via Creative Ideation},
  author={Ma, Ruiying and Liang, Chieh-Jan Mike and Gao, Yanjie and Yan, Francis Y.},
  journal={arXiv preprint arXiv:2510.03851},
  year={2025}
}