batuhanaktas/

frontier-science-bench

	OpenEvolve FrontierScience Benchmark Report




	
	Executive Summary

This report presents a comprehensive benchmark comparing three approaches for solving FrontierScience problems using the nvidia/nemotron-3-nano-30b-a3b model:

Zero-shot Code Generation - Single LLM call to generate Python code Multi-turn Feedback - Up to 10 retry attempts with error feedback OpenEvolve - Evolutionary code optimization with 10 iterations

	Key Results




	

Method Solved Accuracy… See the full description on the dataset page: https://huggingface.co/datasets/batuhanaktas/frontier-science-bench.