Beneficial RL Training on Health Data Improves Cross-Domain Alignment

Models trained on a small amount of beneficial trait data from the health domain show improved alignment scores across unrelated tasks, mirroring prior findings that harmful training data causes broad misalignment. The effect generalizes beyond the training domain, suggesting domain-specific beneficial RL may be a scalable alignment lever. No model sizes or specific benchmark numbers are disclosed in the post.