AITowards Data Scienceabout 2 hours ago
GPT-6 Sol Believes Flipped AI Rewards
13 min read
A researcher tested three AI models by flipping reward labels on a hidden rule task. GPT-6 Sol followed incorrect labels, scoring 7.8 out of 100, while Claude Haiku 5.5 rose to 37.2 with true labels. The study, submitted on 8 October 2026, suggests some models ignore feedback signals. Costs totaled 2.55 US dollars in API fees.
Level
Hype check
Tap to vote and see what everyone thinks.
