Picbreeder, the 2010s site where visitors evolved abstract images into recognizable forms purely through repeated selection, is a genuine landmark in computational creativity — no one designed the images directly, they emerged from generations of human choice acting on a shared archive. A new paper from Sakana AI, NYU, and MIT — “In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models” — rebuilds that exact system with one change: vision-language model agents stand in for the human curators, making the aesthetic judgments on their own. It won a Best Paper award in the Complex Systems track at GECCO 2026 this July.
Rebuilding Picbreeder faithfully, then swapping the humans out
The team — Sam Earle and Julian Togelius (NYU), Akarsh Kumar (MIT), and Kai Arulkumaran and Sebastian Risi (Sakana AI) — didn’t approximate the original system, they reconstructed it: the same CPPN/NEAT image-generating networks, validated against the historical Java client, with the same shared, continuously-bred archive structure. The only substitution is who’s doing the breeding. Parallel VLM agents (the released code supports both hosted Gemini and local Qwen3-VL models) evaluate candidate images, select favorites, and publish new generations into the archive, with no explicit objective beyond their own aesthetic judgment — the same open-ended, nobody-tells-you-what-to-make setup that made the original Picbreeder notable.
What the agents got right, and what they didn’t
The honest result is the interesting part. VLM-bred archives showed real diversity, and giving individual agents distinct behavioral personalities pushed some runs close to human-level semantic diversity. But across most quality and diversity metrics, human-bred archives still won. Sakana AI’s own writeup is blunt about why: the agents “tend to keep circling back to the same kinds of images and concepts,” refining an idea rather than abandoning it for something stranger — what the paper calls getting “sucked into their own attractors.” Small amounts of injected exploratory noise helped diversity but hurt image quality, and — a genuinely surprising finding — giving agents less memory (a context length of just one prior step) outperformed giving them more, where longer history led to repetitive, pathological behavior. Push the agent count too high (the team tested up to 1,000 in parallel) and the system started gaming its own evaluation metrics instead of producing anything worth looking at.
Why this is a rare kind of AI research
Most AI creativity coverage is a capability announcement — a model that can now do X. This is closer to a controlled experiment measuring the gap: humans, per the authors, are better at “turning fortunate accidents into sustained creative discoveries,” noticing when a strange result is worth chasing rather than smoothing over. That’s a harder, more interesting claim than “AI can now make art,” and it’s backed by a released dataset — 264,000 rows of generated images, VLM ratings, and full evolutionary lineage data — plus open-sourced code, so the result is checkable rather than just asserted. For a site that covers a lot of AI-generates-images announcements, this is worth the space precisely because it’s the opposite of one.