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

This paper studies the world-knowledge bottleneck in visual generation, where models can render well but confidently fabricate unfamiliar or newly emerging entities. It introduces SearchGen-20K and SearchGen-Bench, covering 20,839 prompts across twelve failure categories and twenty-two domains, plus a pre-executed SearchGen-Corpus-1M for reproducible offline research. Frontier open generators score only 21–28/100 on the benchmark. The authors find that naive retrieval can hurt by adding irrelevant context. Their teach-then-search co-training framework discovers a generator-specific knowledge boundary and yields monotonic improvement, providing a replayable harness for agentic, search-grounded visual generation.