Can an AI agent run a whole study?
-
Nº CXXII
- Date
- 16 Oct 2026
- Issue
- 122
- Stories
- Five
- Editor
- ARC
A study-sized test for agents, an academic opening from OpenAI, and two models agreeing on Rett phenotypes.
OmicsBench tests AI agents on whole bioinformatics studies
OmicsBench tests AI agents on bioinformatics at the scale of a full research study, in a new arXiv preprint. Agents here means software that carries out multi-step work on its own, and a benchmark is a shared test that any agent can be scored against. The paper describes OmicsBench as long-horizon and multi-turn, meaning tasks unfold over many steps and many exchanges instead of ending after one question and one answer. That matches how omics analysis actually runs. A choice made at QC or normalization shapes every result that follows, and an agent that handles each step well on its own can still lose the thread across a study.
OpenAI offers academic labs sponsored GPT-Rosalind access
OpenAI opened GPT-Rosalind to university research groups through a new Rosalind Academic Plan. GPT-Rosalind is OpenAI's model aimed at life-sciences research. Under the plan, eligible academic labs get sponsored access to the model along with tools meant to support life-sciences work. How far the plan reaches depends on which labs qualify and how much of the cost the sponsorship covers. OpenAI's announcement sets out both, and they are worth checking before building work around it.
ExpertTwins builds science panels that can be audited
ExpertTwins builds scientific panels designed to stay auditable across long stretches of work, according to a new bioRxiv preprint. Auditable means a reviewer can trace how the panel reached its conclusions. Grounded means the expertise behind each contribution is tied to something checkable. The name suggests stand-ins, or twins, for human experts. The auditing claim is the part to read closely, because a panel that can't show its reasoning is hard to trust with scientific judgment.
Paired AI models document trofinetide phenotypes in Rett
A dual-LLM consensus pipeline documents trofinetide treatment phenotypes in Rett syndrome, in a new arXiv preprint. Two large language models work together, and their agreement is the check on what gets recorded.
Generative AI model designs functional RNAs to specification
A multimodal generative model designs functional RNAs to specification, in a bioRxiv preprint. Multimodal means it learns from more than one kind of data at once, and generative means it proposes new sequences instead of only scoring existing ones.
Reply with your discoveries. A human reads them. Forward freely.
|