3 million sampled coding attempts, or rollouts.
A simulator predicts when detectable reward hacking will dominate training
Using reference coding runs, SimHack forecasts hacking onset for unseen reinforcement-learning recipes and helps select curricula that delay it.
Big Tech
Junwei Deng · Alex Bie · Natalia Ponomareva
University of Illinois Urbana-Champaign · Google Research
Research Digest··3 min read
Deng and colleagues investigate whether reward hacking can be forecast before launching a particular reinforcement-learning run.
§