Let a team of agents develop testable research ideas
Researchers built DiscoveryRSI, where agents grow and revise life-science hypotheses into experiment plans; they report their 27B model compiled 50% more runnable tasks than its base.
Done with Custom agentDiscoveryRSI-27B
What they did
A research team built a system where several agents propose research hypotheses in life science. Each idea carries evidence, a plan and rules for disproving it. The agents revise ideas over rounds and turn some into runnable computer experiments. They also trained a 27B model on this research experience and run a live public demo.
How it went
The authors report their model had the highest rate of usable references in their comparison, at 6.3%. They say it picked 16.7% more unique hypotheses and compiled 50% more tasks than its base model.
Worth knowing
The authors warn it can produce plausible but unsupported claims or infeasible experiments, so every hypothesis needs expert review and checking.
Try it yourself
Help me develop research hypotheses on [research question] in [field]. For each, write a clear claim backed by sources, a plan to test it, what result would prove it wrong, and alternative explanations. Then critique each one, add missing controls or measurements, and revise. Keep a record of how each idea changed and why. Use only data I provide: [dataset or sources]. Show me the ranked list with sources before going further, and flag anything you are unsure of.
Swap the highlighted [parts] for your own details.
Source: GitHub · Gen-Verse · Oct 15, 2026
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified. Spotted a mistake? Tell us.
