Shuangshuang Ying, Zheyu Wang, Yunjian Peng +16
Despite strong performance on existing benchmarks, it remains unclear whether large language models can reason over genuinely novel scientific information. Most evaluations score e…