Stanford Paper Shows LLMs Run at a Fraction of Their Creative Capacity — and a Single Prompt Fixes It
A new Stanford study demonstrates that RLHF-tuned models suffer from 'mode collapse' that suppresses creative output. A technique called Verbalized Sampling boosts measured creativity by 2.1x by forcing models to explore low-probability outputs.
The prevailing narrative around frontier models is that they're approaching their ceilings — that GPT-5, Claude 4, and Gemini Ultra represent something close to the limits of current architectures. A new paper from Stanford challenges that assumption from an unexpected direction. As @simplifyinAI summarized, the researchers found that RLHF — the reinforcement learning from human feedback process used to align all major models — systematically suppresses the creative range of LLMs by pushing them toward safe, high-probability outputs. The models are more capable than they appear. They've just been trained to hide it.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.