Function vectors as a model diffing tool: 17 heads repair a bad fine-tune

·LessWrong··

I take two fine-tuned models trained to give bad medical advice, one on Qwen2.5-7B and one on Llama-3.1-8B, from the Model Organisms for Emergent Misalignment collection, and I found out that I could make one safe by simply copying 17 attention heads from the base model it was trained from (which I'm assuming is the good one). The interesting thing is that the reverse is not true. There is also a single direction you can pull out of the difference between the two models, and removing it partly c...

Read full article →

Related Articles

Unauthenticated remote code execution in Nginx
speedping · Hacker News · 6h ago
Highly classified US national security communications network goes down
monkeydust · Hacker News · 1d ago
Harvard researchers find AI coding agents shifted the bottleneck to code review
judahmeek · Hacker News · 1d ago
New Evidence Shows School Phone Bans Are Gradually Improving Test Scores
donohoe · Hacker News · 1d ago
Java4Shell: RCE in OpenJDK when processing Base64
werrett · Hacker News · 1d ago