reWordBench Reveals Reward Model Fragility and Boosts Robustness

reWordBench finds reward models lose accuracy on paraphrases, sometimes below random guessing; a consistency term lets robust models win up to 59% of head‑to‑head tests. Read more: getnews.me/rewordbench-reveals-rewa... #rewardmodels #rewordbench