Li, Qin, Zhou and colleagues built an automated pipeline that turns recent headlines into image generation prompts whose search results are genuinely confusing.
Search agents that locate, verify and crop references improve image generation for recent events.
A new benchmark of 500 time-sensitive prompts shows frontier-model harnesses retrieve correct visual references only about half the time, while the authors' Ranger-8B reaches a 67.2 search score.
Chinese Tech
Yifan Li · Bowen Qin · Kun Zhou · Haochen Zhao · Zikang Liu · Zheng Li · +5 more
Renmin University of China · Tencent · Peking University
Research Digest··3 min read
The authors introduce Savanna, a benchmark of 500 image generation prompts synthesized from recent news, and show that even harness frameworks built on frontier models retrieve a correct visual reference for only about half of them.
§