Sage-Taiyi uses evidence to guide scientific search
Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery · arXiv:2611.12762
In plain words
Scientific research can branch into many ideas, and evidence or unresolved questions from one branch may not shape what gets tried next. Sage-Taiyi is an architecture that keeps track of findings and uncertainty, then uses that record to choose later research targets and methods.
On a test of recovering published research decisions and proposing new ideas, it scored above AI Scientist-v2 overall. A chemistry collaboration also connected a proposed reaction to a measured laboratory result, though that result applies to one model reaction.
The research
A scientific search can waste effort when earlier findings, their conditions or unresolved conflicts do not influence what gets investigated next. On Taiyi-Bench, which tests whether systems recover published research decisions and propose good new ideas, Sage-Taiyi scored higher overall than AI Scientist-v2.
Because the test compares complete systems, it cannot show that tracking understanding alone produced the difference.
Sage-Taiyi is a scientific-discovery architecture that tracks evidence and uncertainty to guide which research target and method to pursue next.
On Taiyi-Bench, systems proposed research routes while receiving scientific knowledge progressively. A literature-grounded verifier scored recovery of published decisions and open-ended idea quality; Sage-Taiyi scored 53.95 Overall, against 47.85 for AI Scientist-v2. The score averages four measures in each group, then averages the two groups.
Both systems used GPT-5.6-Sol and generated 10 ideas per task, so the comparison concerns their full discovery systems on this test.
Sage-Taiyi ranked first on all eight measures, supporting the authors’ view that monitoring and regulation can help scientific ideation and its progression toward validation.
The evidence covers three named ideation tests and five expert-guided research collaborations. The paper evaluates complete systems and does not isolate how much each architecture component contributes, so these results do not establish performance across scientific research generally.
What the evidence shows
- Search needs a memory
Sage-Taiyi is a metacognition-inspired architecture for scientific discovery. It addresses a problem with iterative search: findings can lose their evidence or conditions, conflicts can go unresolved, and earlier outcomes may not shape later choices.
The architecture separates developing ideas from monitoring the search and using that understanding to guide the next target and method.
- A test of routes and ideas
Taiyi-Bench’s combined score summarizes recovery of published decisions and the quality of new ideas. Sage-Taiyi’s higher score than AI Scientist-v2, with both systems using the same base model and generating 10 ideas per task, indicates stronger aggregate performance on these tasks.
- The comparison favors Sage-Taiyi
The comparison holds the base model, task set and number of ideas per task constant: both systems used GPT-5.6-Sol and generated 10 ideas per task. Sage-Taiyi ranked first across the benchmark’s eight measures, which cover recovery of published decisions and qualities of new ideas.
The result supports an advantage for the complete Sage-Taiyi system on these tasks; it does not identify which components account for that difference.
- A route reaches the lab
A chemistry collaboration shows how search connected to a measured experimental result. For one cobalt-catalyzed model reaction, the initial conditions produced trace amounts, while optimized conditions produced an 84% isolated yield.
The cobalt source, ligand and ethylzinc reagent were reported key changes; solvent, temperature and reaction time stayed the same. This links a proposed route to validation in that reaction, without showing the yield transfers to other reactions.
Deep
- Original paper authors
- Weida Wang, Benteng Chen, Yigeng Jiang, Shuchen Pu, Weiping Shen
- Submitted to arXiv
- License
- arXiv non-exclusive