Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Visual Generation

Summary

This paper addresses the world-knowledge bottleneck in visual generation models, which render convincingly but hallucinate unknown content when users request novel characters, trending entities, or post-cutoff events. The authors argue the bottleneck is structural: generators are trained on fixed corpora while the visual world is open-ended. They introduce SearchGen-20K and SearchGen-Bench, comprising 20,839 prompts spanning 12 failure categories and 22 domains. Frontier open-source generators score only 21-28 out of 100, roughly 40 points below existing benchmarks. Naive retrieval-augmented search indiscriminately fetches content, injecting noise even for prompts the generator already handles. The root cause is identified as the generator-specific, evolving knowledge boundary: the dividing line between what the model has internalized via training and what must remain in external context. The authors show this boundary can be discovered through a teach-search co-training framework, where even minimal versions yield monotonic improvements, laying groundwork for recursive self-improvement in visual generation. Paper ID: arXiv 2607.05382.