模型精选02:28UCLA博士生发布Trace-and-Amplify,可规模化采集训练时奖励黑客UCLA博士生搞了个Trace-and-Amplify,不用诱导就能抓训练时的奖励黑客,准确率从60%提到90%,比老方法强多了。#Trace-and-Amplify#Qwen2.5-Coder#DeepSeek-Coder#奖励黑客lmarena.ai@lmarena_ai原文稍后读已读值得跟进有用关注 Trace-and-Amplify