Skip to content
QbitAI · WeChat

Claude Team Tests New Training Method on Qwen

Claude团队用Qwen测试全新训练方法

Anthropic proposed MSM training between pretraining and alignment fine-tuning. Tests on Qwen2.5-32B and Qwen3-32B cut misalignment from 68% and 54% to 5% and 7%. The key point is MSM complements AFT rather than replacing it.

Why it matters: HKR-H/K/R all pass: Anthropic offers a concrete MSM alignment method with Qwen2.5-32B and Qwen3-32B rate drops. It is strong safety research, not a model launch or major product update, so 82 fits.

Read the original ↗Export Markdown