#counterfactual examples

Google DeepMind and collaborators introduce Crome, a causal framework that improves reward modeling robustness in LLM alignment by using counterfactual data augmentation to tackle reward hacking issues.