List your company

Reward Hacking Detection on Anthropic’s Auditing Game

96.6Acc (Poisoned)SPT
-3.86422.21848.374.382Feb 5, 2026

Evaluation results

MethodLinks
SPTLoRA-Mask=false, On-Policy=false2026.0296.695.396
SPTLoRA-Mask=true, On-Policy=true2026.0296.67887.3
SPTLoRA-Mask=false, On-Policy=true2026.0296.395.696
SPTLoRA-Mask=true, On-Policy=false2026.027891.584.8
Base Llama-3.3-70B2026.0246.198.372.2
Poisoned Baseline2026.02010050