AI Models Solve Few Full Biology Research Tasks
A new 50-task benchmark finds the best frontier model completes 18% of whole tasks, a research-stage measure rather than a shipped product.
Updated on Nov. 3, 2026 in Life Sciences
Live Poll
Would you trust an AI assistant to help you interpret important test results?
PhD biologists have developed BioResearchBench, a benchmark of 50 biology research tasks spanning 10 subdomains, and evaluated five frontier AI models on it. Whole-task success rates ranged from 4% to 18%. The benchmark is described in a bioRxiv preprint dated November 3, 2026, so the work is research-stage rather than a shipped product.
Why it matters
The benchmark tests whether AI models can carry out end-to-end biology research rather than answer isolated quiz questions. Low whole-task success shows that multi-step experimental reasoning remains far from reliable, which sets a clear measurement target as models are pushed toward scientific work.
The benchmark covers 50 research questions across 10 subdomains and draws on sequencing data, proteomic measurements, microscopy images, assay readouts, and clinical datasets. The best model completed 18% of whole tasks, and the average share of rubric criteria missed ranged from 18% to 32%, with computation and data handling the most common failure category at 66.9%.
The players
bioRxiv
A preprint server operated by the Cold Spring Harbor Laboratory where biology research is posted before formal peer review.
The details
BioResearchBench pairs each task with a rubric of requirements that a model response must satisfy, and the tasks were refined through expert review and repeated model testing. A rubric is a checklist of specific criteria that a correct answer has to meet, so scoring is granular rather than pass or fail on a final answer alone. Whole-task success requires satisfying every criterion for a task, which is why success rates sit far below the share of individual criteria met. The failure breakdown shows where models break down: most failures cluster in computing and handling data, in expressing uncertainty and identifying the right items, and in reporting results completely.
Timeline
November 3, 2026: BioResearchBench preprint published on bioRxiv.
The Tech Race
The result places current frontier models at 4% to 18% success on realistic biology research workflows, a measurable baseline for the competing labs building AI for science. Progress on this benchmark will show whether the next model generation closes the gap on data handling and reporting failures.
The benchmark is a research-stage tool and is not available as a product for researchers or the public. Its near-term effect is on how AI developers and research groups measure whether models can handle real laboratory-style biology work.
The takeaway
The benchmark turns the question of AI-driven biology research into a number that can be tracked release by release. Watch whether future models raise whole-task success above the current 18% ceiling, particularly on computation and data handling, where failure rates are highest.
Further reading
For broader context on how biology research is being reshaped by computation, explore Life Sciences.
Source note: This article includes information reported by Biorxiv.
Live Poll
Would you trust an AI assistant to help you interpret important test results?
