We also introduce SciArena-Eval, the first meta-evaluation benchmark for scientific literature tasks built on collected human preference data. The goal is to understand – and improve – LLM-based evaluations in this area.
Both SciArena and SciArena-Eval are now available.
