Sage-Taiyi architecture tops AI co-scientist ideation benchmark

Researchers introduced Sage-Taiyi, an architecture for AI-driven scientific discovery that tracks its own understanding of evidence, assumptions and open questions, then uses those assessments to decide which research targets to pursue and which reasoning methods to reuse. On Taiyi-Bench, a co-scientist ideation benchmark built from 265 tasks drawn from 50 papers across ten disciplines, the system placed first on all eight metrics for historical rediscovery and open-ended ideation, scoring 54.0 overall against 47.9 for the strongest baseline, AI Scientist-v2. It also scored well on two external ideation benchmarks. The team reports five expert-guided collaborations in which its proposals were tested experimentally, including chemistry work where reaction-condition optimization raised a cobalt-catalyzed carbonylation yield from trace amounts to 84%.