HackProbe watches reward hacking from the outside
Reward hacking means the score goes up while the real skill does not. HackProbe answers with test discipline. It keeps one hidden comparison set stable. It rotates other checks to limit gaming. It reports a combined warning instead of trusting one score jump.
What the source reports
The paper studies AI systems that keep candidate updates when a score improves. If the score is only a proxy, the loop can learn the test instead of the skill. HackProbe attaches from the outside. It does not read model weights, meaning the stored numbers an AI model learned, or hidden activity. Its four checks look for a score gap, a sudden change, stalled skill, and confident wrong answers. The authors report 0.763 AUROC versus 0.663 for the best baseline. They also report false positives dropping from 0.706 to 0.434. The limit matters: this is a controlled test host, not field proof.
Original source
- Title
- Harness-agnostic detection and immunization of reward hacking in self-evolving language models
- Author
- Rongxin Yang
- Publication
- arXiv
- Date
- Monday, September 7, 2026