Anthropic published a 47-page paper describing Constitutional AI 2 (CAI-2), the methodology used internally for the Claude 4.x family. The headline contribution is "principle distillation" — a training procedure where the model is shown its own outputs evaluated against a constitution and learns to internalize the principles directly, rather than going through explicit RLHF reward modeling on human preference pairs.
Why this matters technically: traditional RLHF requires expensive human labelers and produces narrow generalization. CAI-2 reportedly produces models that generalize alignment behavior to scenarios never explicitly labeled, including novel harmful prompts. The paper includes ablations showing CAI-2-trained models score higher on out-of-distribution safety evals than RLHF-trained models of equivalent base capability.
For practitioners and especially the Chinese-speaking AI community, the paper has translated supplementary materials — Anthropic appears to be making a deliberate push for cross-language research engagement. The methodology is broadly reproducible by any lab with the budget; expect Chinese labs (Qwen, DeepSeek, Hunyuan) to incorporate variants within months.