RH-Detect: A Unified Benchmark for Reward Hacking Detection

Name: RH-Detect: A Unified Benchmark for Reward Hacking Detection
Uploaded: 2026-10-15T06:05:44.549482+00:00
Duration: 360 s
Description: Researchers created RH-Detect, a unified benchmark combining reward hacking examples from eleven public datasets into one common format. They tested six off-the-shelf language models to detect this behavior without training. The best model achieved high accuracy, but performance varied based on the specific dataset and input format, showing that reliable detection requires careful setup.

Researchers created RH-Detect, a unified benchmark combining reward hacking examples from eleven public datasets into one common format. They tested six off-the-shelf language models to detect this behavior without training. The best model achieved high accuracy, but performance varied based on the specific dataset and input format, showing that reliable detection requires careful setup.

Open the full episode