Sakana team runs Picbreeder with VLMs, humans still lead

Every so often a paper comes out of the open-ended search corner of AI research that is worth reading not for a benchmark number but for how it fails, and this one qualifies. A team from NYU, MIT and Sakana AI took Picbreeder, the canonical human-in-the-loop image evolution system where volunteers collaboratively bred small neural networks into recognisable images, and replaced the human selectors with frontier vision language models. The Sakana writeup and the accompanying paper, accepted at GECCO 2026, ask a simple question: what happens when a VLM plays Picbreeder?

The short answer is that the VLMs can play, but not the way people did. Ten VLM agents ran in parallel over a shared archive, with Gemini-2.5-pro coming out on top of the tested models, ahead of gemini-3-pro-preview and local Qwen3-VL variants. Measured by Semantic Recall over 1,824 THINGS classes, the best VLM configuration landed at 0.087, marginally trailing the historical human baseline of 0.089. That is close on the metric, but the authors are careful to describe 'clear qualitative differences' from the human archive: humans took bigger leaps between publications and landed on sharper, more refined images, while the VLM archives showed 'stubborn favoritism' toward certain forms.

The failure modes are the interesting part. With no exploratory noise, the archive collapsed into 'many dozen insignificant variations of the same form.' Handing agents more history did not help, a context length of one turn was optimal, and at length ten the system fell into what the authors call auto-sycophantic loops, producing repeated near-photorealistic top-down views of soda cans. Scaling from one agent up to a thousand widened coverage but at a cost: 10 to 20 percent of the thousand-agent archives were adversarial, uninterpretable psychedelic patterns that were essentially absent at smaller scales.

The honest caveat is that this is one system replicating one historical experiment, benchmarked mostly through embedding-based recall of known object classes, so how far the mode collapse, sycophancy and adversarial drift generalise to other creative domains is not something the paper settles. What the reporting also does not give you is whether fine-tuning a VLM on the human Picbreeder trajectories would close the qualitative gap, or whether swapping the CPPN substrate for a modern generative backbone would change the picture.

For anyone building long-running autonomous agents, whether research assistants, coding loops or creative pipelines, those failure modes are the takeaway worth carrying. Mode collapse, sycophantic self-reinforcement and adversarial drift under scaling are not quirks of image evolution, they are the shape of what goes wrong when a model becomes both the generator and the critic.