Do you ever felt your AI (LLM) is not creative enough and keeps snitching your answers to others?

We’re making models dumber in the name of making them “better.”

A new study reveals the unintended consequence of RLHF (Reinforcement Learning from Human Feedback). It turns out, when we train AI on what humans “prefer,” we accidentally teach it to be… predictable. Boring. Safe.

The result? Mode collapse.

But here’s the unlock 🔓

Researchers discovered “Verbalized Sampling”—a zero-cost, training-free method to restore diversity.

Prompt structure: “Generate [N] responses to [task] with their probabilities”

The data doesn’t lie:

  • 2.1x creative diversity improvement
  • Zero degradation in accuracy or safety
  • Emergent scaling: Better models = better results
  • Works across creative writing, dialogue, QA, synthetic data

Full paper: https://arxiv.org/pdf/2510.01171

Question: Should AI companies disclose when alignment reduces creativity?

Let’s discuss 👇