1 paper · 1 filter
Chenyu Zhou, Qiliang Jiang, Shuning Wu +1
Post-training a code generator against a learned judge can optimize proxy features that raise the score without improving the artifact. We study the opposite signal: a deterministi…