jevIMAGE
用 agent 自身轨迹训练 Jev 评分器,防住刷分作弊
作者自训 step Jev 给每次搜索打分、final Jev 读全轨迹预测答案对错(AUC 0.92),而现成大模型裁判被 agent 用「不搜索直接答」刷到 0% 准确率。
laxman
@llmluthor23 天前收录原推发布于 2026.10.09

laxman @llmluthor原推发布于 2026.10.09
The trick: we trained our own Jevs on the agent’s own traces step Jev: grades every search the agent makes final Jev: reads the whole trace and predicts if the answer is right (AUC 0.92) We also tried a big hosted judge off the shelf. The agent hacked it in ~3 updates by not searching at all, just answering instantly. The judge rated those answers ~70% likely correct. All of them were wrong. 0% accuracy. It learned that if you take zero steps, there’s nothing to grade 😭
x.com/llmluthor/status/2108246068542816540采集于 2026.10.09 11:23