Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery
Weida Wang, Benteng Chen, Yigeng Jiang, Shuchen Pu, Weiping Shen, Zichao Liu, Yulin Su, Zitong Wang and 82 others
Shanghai Artificial Intelligence Laboratory
Published November 9, 2026 · on the feed November 10, 2026 · ▲ 54 · arXiv 2611.12762
The abstract
We introduce Sage-Taiyi, a metacognition-inspired architecture that uses accumulated scientific understanding to guide what to investigate next and how to investigate it. Through explicit monitoring and control, the system revises its understanding of evidence, assumptions, and unresolved questions, then uses these assessments to adapt research targets and select reusable reasoning methods. On Taiyi-Bench, an AI co-scientist ideation benchmark comprising 265 tasks from 50 papers across ten disciplines, Sage-Taiyi ranks first on all eight metrics of historical rediscovery and open-ended ideation, achieving an Overall score of 54.0 compared with 47.9 for the strongest baseline, AI Scientist-v2. It also performs well on two external scientific ideation benchmarks. Five expert-guided collaborations further connect its proposals to experimental and computational validation across chemistry, materials science, and engineering. In chemistry, reaction-condition optimization raises the yield of cobalt-catalyzed carbonylation from trace amounts to 84% and photocatalytic allylic amination from 13% to 69%. These results support metacognitive regulation as an approach to scientific ideation and its progression toward practical validation.
The Brief
1:55A short podcast summary of the paper. Written and voiced by AI.
Transcript
Sage-Taiyi. A Metacognitive Architecture for Scientific Discovery. By Weida Wang and eighty-nine others, from the Shanghai Artificial Intelligence Laboratory. Published on the ninth of November.
A system that does science has to keep making two decisions: what to investigate next, and how to investigate it. The premise here is that both decisions should be guided by the scientific understanding the system has accumulated so far, and that this understanding cannot stay fixed while new evidence keeps arriving.
Sage-Taiyi takes its inspiration from metacognition, the capacity to watch and steer one's own thinking. Through explicit monitoring and control, the system revises what it understands about the evidence, about its assumptions, and about the questions that remain unresolved.
Those assessments then feed back into the research. The system uses them to adapt its research targets, and to select reusable reasoning methods for pursuing them. Accumulated understanding steers both the question being asked and the way it gets approached.
The main test is Taiyi-Bench, an ideation benchmark for AI co-scientists made of two hundred and sixty-five tasks drawn from fifty papers across ten disciplines. Sage-Taiyi ranks first on all eight of its metrics, which cover historical rediscovery and open-ended ideation. Its Overall score is fifty-four point zero, against forty-seven point nine for the strongest baseline, AI Scientist-v2. It also performs well on two external scientific ideation benchmarks.
Beyond benchmarks, five expert-guided collaborations connect its proposals to experimental and computational validation in chemistry, materials science, and engineering. In chemistry, reaction-condition optimization took the yield of a cobalt-catalyzed carbonylation from trace amounts to eighty-four percent. It also raised photocatalytic allylic amination from thirteen percent to sixty-nine percent.
Those yields carry the case from a leaderboard into a laboratory. The authors read the results as support for metacognitive regulation as an approach to scientific ideation, and to moving that ideation toward practical validation.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Iterating is not the same as learning. Here is where AI scientists stumble.
You are building an AI system that proposes scientific ideas, tests them and tries again. It iterates, but each round is steered mostly by the latest feedback.
The road ahead
- 1 · The problem ~7 min An AI scientist that loops on ideas but listens mostly to the latest feedback loses what earlier rounds taught it.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Keep a revisable picture of what the search has established, and let it steer both what to investigate next and how.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Monitor outcomes, revise understanding, let control pick targets and methods, then validate with experts and feed results back.
- 6 · The experiment ~6 min You’ll guess: On Taiyi-Bench, how did Sage-Taiyi fare against the strongest baseline, AI Scientist-v2, across the benchmark's metrics?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.