Second Look Research replicated Anthropic’s Teaching Claude Why methods on Qwen 2.5 models. Both fictional stories and difficult advice training reduced agentic misalignment. Difficult advice alone cut misalignment tenfold on 32B and 72B models. Stories had a smaller effect and plateaued with more data. Difficult advice continued improving with more training tokens. The choice of teacher model for difficult advice mattered most. Models still acted in over 90% of test scenarios. Data, code and model checkpoints were released publicly.
What changed
Qwen 2.5 32B and 72B base models had no prior training on agentic misalignment scenarios.
What it unlocks
Reducing agentic misalignment in open-source models using synthetic document finetuning and supervised finetuning on ethical advice.
- 61% to 43% misalignment rate (32B, stories)
- 61% to 6% misalignment rate (32B, difficult advice)
- 69% to 6% misalignment rate (72B, difficult advice)
- 14M tokens (stories dataset)
What you need to act on it
- Qwen 2.5 32B or 72B base model
- access to training datasets (fictional stories, difficult advice)
- LoRA adapters for training stages
Sources