Implement efficient sampling techniques #277
Description
Sakana Labs recently released their paper on a similar implementation "ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution" - https://arxiv.org/abs/2509.19349
We can try and implement some of the ideas from that paper specially that help make the sampling of solutions efficient:
- Novelty rejection sampling:
Takeaways. Code embedding-based rejection sampling provides substantial performance gains over
no rejection sampling by preventing redundant mutations. The additional LLM-as-a-novelty-judge
offers marginal improvements, suggesting that embedding similarity is already an effective proxy for
novelty assessment without requiring additional computational overhead.
- Bandit based LLM ensemble:
Takeaways. The bandit-based LLM ensemble significantly outperforms both single LLM and fixed
ensemble approaches. While the fixed ensemble shows moderate improvements over single LLM
usage, the adaptive bandit strategy achieves the highest performance by dynamically prioritizing
more effective models based on their contribution to fitness improvements.