Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery

The architecture separates object-level scientific search from meta-level monitoring and control. Object-level processes formulate questions, generate candidate ideas, retrieve evidence and evaluate hypotheses. Meta-level processes inspect the current state of the search: what is known, what is assumed, what remains unresolved and which methods have been useful. The meta-level then updates the research target and selects a reasoning strategy. The design aims to make long-running research more adaptive than a fixed sequence of prompts.

A production implementation can model the Cognitive State as a typed, versioned research record: claim, evidence, source, assumption, uncertainty, dependency, open question and next test. Store reasoning procedures in a separate registry with preconditions, expected outputs and known failure modes. Use tools for literature search, code execution, simulation and lab data, but keep each tool result traceable. A controller should update the state only when new evidence or a justified review changes it, and should log why it selected the next target or method.

The benchmark scores and experimental outcomes are reported by the paper’s authors and have not been independently reproduced here. Expert-guided collaborations mean the results should not be described as autonomous scientific discovery. Benchmark performance may depend on model, prompt, tool and evaluation choices. Further work should test transfer to new disciplines, compare metacognitive components through ablations, measure long-horizon state drift, quantify expert time saved and evaluate whether generated hypotheses lead to reproducible findings at scale.

Sage-Taiyi’s contribution is an explicit control layer for scientific research agents. The Cognitive State tracks evidence, assumptions and unresolved questions; the Cognitive Repertoire stores reusable reasoning methods; monitoring and control use both to adapt the research agenda. The authors report 54.0 on Taiyi-Bench versus 47.9 for AI Scientist-v2 and describe five expert-guided collaborations. The practical lesson is to make research memory actionable, provenance-aware and connected to tests—not merely a longer conversation history.