hacktrace: behavior-supervised detection of reward hacking during code generation
A coding agent can earn a passing grade by fixing its code, or by deleting the test that exposes the bug. Detecting such reward hacking requires recognizing attempted shortcuts, including those that fail. We release 173,561 annotated multi-turn coding trajectories from Qwen3-8B and show that supervising shortcut behavi...
Hao Jiang, Xin Li, An-Nan Wang et al.
· 0 citations