AI research assistants are good at producing ideas and poor at remembering why earlier ideas failed. Sage-Taiyi, described in a new preprint, is built around fixing that, and its most persuasive evidence is not a benchmark score but two chemical reactions that worked much better after its suggestions.
The design borrows from the psychology of metacognition, or thinking about thinking. One layer does the scientific search: it breaks a question into targets, generates and refines ideas, and stores evidence from experiments, papers and human experts. A second layer keeps a running picture of what that evidence supports, which assumptions are shaky and which questions are still open, and uses it to choose what to investigate next and which reasoning method to apply. The aim is to stop an agent from chasing whatever the latest result suggests while forgetting a failure three rounds back.
On Taiyi-Bench, which the team built from 265 tasks drawn from 50 papers across ten disciplines, Sage-Taiyi scored 54.0 overall against 47.9 for AI Scientist-v2, the strongest of the comparison systems, and ranked first on all eight metrics. It ran on GPT-5.6-Sol, and baselines included the same model inside Codex and Claude Code as well as Claude Opus 5 and other dedicated research agents, so the gain is attributed to the architecture rather than a stronger model. Papers were chosen from after the model's training cutoff to limit contamination.
The lab results carry more weight. In five collaborations guided by domain experts across chemistry, materials science and engineering, the system's proposals were tested for real. Optimising conditions for a cobalt-catalysed carbonylation raised its yield from trace amounts to 84%, and a photocatalytic allylic amination rose from 13% to 69%. Those are the kinds of gains a chemistry group would notice.
There are good reasons for caution. The benchmark is the authors' own, and ideas are scored by an automated, literature-grounded verifier that the team says agrees well with human experts, but that is still a model grading a model. The five collaborations were expert-guided, so it is hard to separate the system's contribution from that of the scientists steering it, and two strong reaction results do not tell us how often its suggestions fail. As a preprint, none of this has been peer reviewed, and the authors give no figure for the computing cost of each run.