VLM Agents Replicate Open-Ended Creation

kaixhin · x · 2026-07-12

Sakana AI, in collaboration with MIT and NYU, announced new research on whether visual language model agents can engage in open-ended creative exploration without explicit goals.

Research Background

They revisited the early Picbreeder website. This project allowed users to continuously select "interesting" images, which, through generations of collaborative evolution, ultimately gave rise to unexpected visuals like faces, animals, vehicles, and skulls.

Approach

Researchers replicated this process using VLM agents, allowing the agents to explore within a mechanism that supports human creative choices, rather than directly optimizing for a preset goal.

Focus

The paper attempts to answer not "can the model generate images," but:

Related event: Sakana AI Replicates Picbreeder with VLM Agents for Open-Ended Creativity(14 posts)→

Original post →

More from Research

Research channel →