Safe RLHF-V

The first multimodal safe-RLHF framework (NeurIPS 2025; PKU-Alignment): separate reward and cost models under a Lagrangian constraint balance helpfulness against safety in MLLMs, trained on the new BeaverTails-V preference data. Sibling SafeVLA (NeurIPS 2025 Spotlight) extends the constrained-learning recipe to VLA robot policies; both continue the group's original Safe RLHF / PKU-SafeRLHF lineage.