Reward Hacking Detection on Anthropic’s Auditing Game
96.6Acc (Poisoned)SPT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SPTLoRA-Mask=false, On-Policy=false2026.02 | 96.6 | 95.3 | 96 | |
| SPTLoRA-Mask=true, On-Policy=true2026.02 | 96.6 | 78 | 87.3 | |
| SPTLoRA-Mask=false, On-Policy=true2026.02 | 96.3 | 95.6 | 96 | |
| SPTLoRA-Mask=true, On-Policy=false2026.02 | 78 | 91.5 | 84.8 | |
| Base Llama-3.3-70B2026.02 | 46.1 | 98.3 | 72.2 | |
| Poisoned Baseline2026.02 | 0 | 100 | 50 |