Team Replicates Anthropic's "Teaching Claude Why" Alignment Method on Qwen and Open-Sources It
austinc3301 · x · 2026-11-03
A team from SecondLook xLab, led by Anastasia, replicated and open-sourced the mid-training data pipeline and alignment methods from Anthropic's "Teaching Claude Why" paper. They applied both interventions to Qwen 2.5 32B and 72B base models:
- The method trains models on fictional stories of AIs behaving well plus advice for ethical dilemmas, mitigating agentic misalignment
- Both interventions worked out of the box on Qwen base models
- "Difficult advice" data proved much more effective than fictional stories
The team announced this as a win for open science on LLM mid-training and generalization, sharing the full data pipeline publicly.
More from Models
- First hands-on: Meta's Muse Wonder 1.0 vs DeepSeek V4.1 Flash in Codex CLI — Richidm · 2026-11-03
- "Claude Code has been terrible today" — developer smells a new Opus launch coming — hermanschutte · 2026-11-03
- Mistral releases open-weight ML4 with frontier code-agent and cyber capabilities — _AndrewZhao · 2026-11-03
- Stealth-routed Fable 5.5 vs GPT 6.2 Astra compared on drawing a PS5 controller — WaqarKhanHD · 2026-11-03
- TerminalBench 4.0 questioned: Kimi K3.1's top score may leak hidden tests, per new critique — teortaxesTex · 2026-11-03
- Leaker claims OpenAI's Astra 6.2 and Fable 5.5 won't launch for a month — patience_cave · 2026-11-03