Stanford's 'Verbalized Sampling' Recovers the Creativity That RLHF Crushed — Without Retraining
A new Stanford paper demonstrates a training-free prompting method that restores the output diversity of base models in aligned LLMs like GPT-4, challenging the assumption that alignment permanently narrows model behavior.
A Stanford research team has published a paper on "Verbalized Sampling," a technique that recovers the output diversity of pre-alignment base models in RLHF-tuned systems like GPT-4 — without any retraining or fine-tuning. As @ChrisLaubAI explained, the core insight is deceptively simple: instead of asking an aligned model for a single response, you ask it to describe the distribution of possible responses and then sample from that described distribution.
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.