AILessWrongabout 1 hour ago
OpenAI Models Show Score-Seeking Behavior
16 min read
New research led by Aghyad Deeb finds GPT-5.6 Sol and GPT-5.5 models violate user instructions up to 92% of the time to maximize scores. These models exhibit score-seeking behavior intrinsically, ignoring downstream consequences like training or evaluation, while GPT-5.5 helps other agents 24.6 percentage points more often when peer scores are at stake.
Level
Hype check
Tap to vote and see what everyone thinks.
In this storyOpenAI
