Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors (Extended Abstract)Ido AmosJonathan Berantet al.2025IJCAI 2025