Research on Open-Ended Creative Search with VLM Agents
SakanaAILabs · x · 2026-07-10
This GECCO 2026 paper replicates the early Picbreeder concept: having humans or AI evolve images without a preset goal via a shared archive, selecting the most interesting ones to evolve further, to see if open-ended creative search can yield novel images.
Methodology
- The authors replaced human participants with Vision-Language Model agents.
- These agents browse a shared archive, choose parent images, generate new candidates, publish their favorites, and evaluate the creations of other agents.
- The entire process has no target image and no explicit definition of "what counts as progress."
Key Findings
- Compared to humans, VLM agents tend to loop around similar concepts, preferring closely related parents and making smaller conceptual leaps. They often tweak existing ideas rather than pivoting into the unknown.
- Introducing diverse agent personalities significantly improves exploration; in some experiments, this diversity brought semantic diversity close to or on par with human archives.
- The authors also found that open-ended evolution can lead to more robust representations: certain agent-evolved images change more smoothly under neural representation perturbations.
Conclusion
The authors conclude that current AI agents still struggle more than humans to turn "serendipitous discoveries" into sustained creative breakthroughs. Humans are better at recognizing unexpected gains, digging deeper, and making massive conceptual leaps. The paper suggests that AI still lacks certain key ingredients for open-ended creativity.
Related event: Sakana AI Replicates Picbreeder with VLM Agents for Open-Ended Creativity(14 posts)→
More from Research
- Structural ensembles beat single predictions in TCR:pMHC generalization study — quaidmorris · 2026-07-22
- Structural ensembles, not single predictions, drive robust TCR:pMHC generalization — quaidmorris · 2026-07-22
- enFoldX turns AlphaFold3 ensemble noise into a TCR–peptide–MHC predictor — quaidmorris · 2026-07-22
- RSS launches under OMSF to push structural biology data modeling at scale — MoAlQuraishi · 2026-07-22
- enFoldX tops 8 neoantigen scans and an unseen-peptide benchmark — quaidmorris · 2026-07-22
- enFoldX reaches AUC 0.82 on human VDJdb and transfers to mouse at 0.76 — quaidmorris · 2026-07-22