What it is
Builds a benchmark and co-training setup for search-augmented image generation that learns when search helps and when it adds noise. Instead of blind search on every prompt, it learns a generator-specific knowledge boundary: co-train a gated search reasoner and the generator so search is used only for knowledge the model cannot internalize.
Key results
- Open generators score just 21-28/100 on SearchGen-Bench
- Search-intensive prompts cause about a 40-point drop vs NoSearch
- Adaptive search improves monotonically over no search and blind search
Search helps visual generation only when it is calibrated to the generator's changing knowledge boundary; otherwise retrieval noise hurts.