REWORDBENCH

Summary from the TLDR AI newsletter

This work shows that many popular reward models are brittle to basic rewording of the prompts. It includes a benchmark and a potential strategy to make the reward models more robust.