Can VLMs Create Open-Endedly Like Humans
kaixhin · x · 2026-07-11
This paper discusses: when no one explicitly tells AI what to create, can it generate creative exploration like humans?
In collaboration between MIT and NYU, the research team published the GECCO 2026 paper "In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models". They replicated the open-ended evolution process of the early website PicBreeder: without setting predetermined goals, users simply select "interesting" images, allowing them to gradually evolve into unexpected forms like faces, animals, vehicles, and skulls through multiple rounds of selection and mutation. In the experiment, VLM agents explore, select, and evolve images in a shared archive, publish works, and evaluate others' works, but without any target images or definition of "progress."
Results show:
- AI can indeed do open-ended exploration, but with clear limitations: VLM agents are more easily attracted by existing appearances or semantics, tending to optimize ideas at hand rather than actively thinking outside the box to find truly unexpected new directions.
- Diverse agent personas significantly improve exploration: Introducing agents with different personalities improved exploration quality, bringing the semantic diversity of the generated archives closer to human-created ones.
- Humans still hold a unique advantage: The paper points out that VLMs have not yet reached the human ability to continuously transform accidental discoveries into long-term creations.
Related event: Sakana AI Replicates Picbreeder with VLM Agents for Open-Ended Creativity(14 posts)→
More from AGI Musings
- The Evolution of LLM Business Models: Selling Outcomes Over Tokens — yacineMTB · 2026-07-22
- Bindu Reddy says GPT-6 is coming soon, with Alibaba, DeepSeek and Kimi close behind — bindureddy · 2026-07-22
- Bindu Reddy says the industry still lacks a way to train 20T models and scale post-training RL — bindureddy · 2026-07-22
- Advanced AI Models Are Becoming Impossible to Plug and Play — emollick · 2026-07-22
- AI suggested a better composition, and that made one user uneasy — Sydde · 2026-07-22
- The Thimble and the Waterfall: AI's Data Bottleneck and Feedback Loops — dyamins · 2026-07-22