#machine learning
Dec 2025
AdvJudge-Zero: Binary Decision Flips in LLM-as-a-Judge via Adversarial Control Tokens
Under GRPO training, the hardened judge eliminates the reward-collapse failures (false-positive spikes and length collapse) the authors observe in the unhardened baseline on both MATH and GSM8K at ten seeds per condition.
Tung-Ling Li, Yu-Hao Wu, Hong-Liang Liu
· arXiv.org · 4 citations
· ⚡1