Tag
This paper proposes an evolutionary framework inspired by parallel tempering that uses multi-temperature sampling and information exchange to improve the diversity and quality of scientific hypotheses generated by large language models, demonstrated across molecular, equation, and algorithm discovery.