One Token to Fool LLM-as-a-Judge

Yulai Zhao, Haolin Liu, Dian Yu +4

Large language models (LLMs) are increasingly trusted as automated judges, assisting evaluation and providing reward signals for training other models, particularly in reference-ba…