#reward model training

The Goblin Incident Reveals What Frontier AI Training Really Breaks: Why Reward Models Leak Into Every Layer

When a single personality feature poisoned generations of models—and why GPT-5.6 exists to...