Teaching Qwen Why

Published 4 November 2026

A partial replication of Anthropic's "Teaching Claude Why" on Qwen 2.5 32B and 72B base models. Fine-tuning on difficult-advice conversations cut agentic misalignment from 61% and 69% to 6%, while stories about aligned AIs helped far less and plateaued; the model that generated the training data mattered most.