🧪 Test?View on arXiv
SimHack: Predicting the onset of detectable reward hacking before RL training
Not provided
reinforcement learningreward hackingsimulationcurriculum learning
2610.28707
Builder Relevance
2h ago80%
Abstract
The paper presents SimHack, a simulator that predicts the onset of detectable reward hacking in reinforcement learning training, allowing for the selection of lower-risk training recipes.
Reality Card
Core Claim
SimHack predicts the median onset of reward hacking with a relative error of 11.3% and achieves a rank correlation of 0.85 across different training recipes.
Method / Result
SimHack can guide training decisions by screening candidate RL curricula, resulting in training for 96.6% longer without hacking.
Limitations
The initial hacking rate does not determine onset, leading to variability across training recipes and runs, which may affect reproducibility.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.
No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.