Language Models Resist Alignment

PKU-Alignment's compression-theory account of alignment "elasticity" — post-trained models rebound toward the pretraining distribution, so alignment is shallower than it looks — ACL 2025 Best Paper. The theoretical anchor of the group's safety line, alongside its moral-progress ProgressGym benchmark (NeurIPS 2024 D&B Spotlight) and SAE-V multimodal interpretability (ICML 2025).