LLMs Can Judge Ideas But Can't Generate Them

🔥 What happened Researchers at Johns Hopkins turned 200 ICML 2026 papers into a Twenty-Questions-style game called Re-Search: six recent LLMs had to rediscover each paper's core idea using only yes/no questions. Nearly all progress happens in the first 30 of 100 questions — after that, models plateau hard. 💡 Why it matters The bottleneck is generation, not judgment: models spot a good idea or question instantly when handed one, but rarely produce it themselves. For autoresearch systems, that means injecting externally sourced ideas beats upgrading any single model's generation ability. ⚡ Our take The industry keeps optimizing the wrong skill. A model that can only judge ideas is a good editor, not a researcher. The winners will be systems that deliberately flood the search space with outside ideas.