Sage-Taiyi research puts revisable scientific assumptions at the center of AI discovery
A new preprint introduces Sage-Taiyi, a research-agent architecture that revises its working assumptions as evidence accumulates and uses those revisions to choose what to investigate next. The paper was submitted on November 9 at 17:38 UTC. arXiv record
The authors report a score of 54.0 on Taiyi-Bench, compared with 47.9 for their strongest baseline, AI Scientist-v2. The benchmark contains 265 tasks drawn from 50 papers across ten disciplines. These are results from the authors’ evaluation, not an independent certification of scientific ability. Technical paper
Five expert-guided collaborations connect proposals to experiments or simulations. The paper describes people selecting candidate ideas and validating their implementation. That role matters: the work should not be described as an autonomous replacement for scientists. Technical paper
The public repository currently contains a README and documentation. Its presence alone is not evidence that a complete executable research system has been released. Project repository
The design question is useful beyond a benchmark leaderboard. Can an agent record which assumptions failed, preserve the evidence behind that judgment and change its next experiment accordingly? That is a more demanding test than producing plausible hypotheses in isolation. Independent replication, accessible implementation details and comparisons across real research budgets would help establish how far this approach generalizes.
Sources and citations
Revision history
Published .