量子位官网T2·Models
78 pts
UniPat AI introduced PaperBenchX, a benchmark evaluating large models on 93 real-world paper reproduction tasks. Results show that even the top-performing GPT-6 Astra achieved only a 13.98% full reproduction rate, revealing a significant gap between solving specific math problems and reliable end-to-end scientific replication. The benchmark compiles scientific conclusions into executable, replayable, and verifiable workflows to establish new standards for measuring AI progress in science.