LabBench: Can AI agents decide what experiment to run next? (gamowlabs.com)

Gamow Labs has introduced LabBench, a novel evaluation framework designed to assess AI agents' capabilities in deciding the next experiment in biological research. This framework is built upon 20 real-world tasks derived from wet-lab records in drug discovery and genomics. Each task presents agents with historical lab data while withholding final decisions, requiring them to commit to a next step based on their assessments. Notably, the performance of five leading AI agents, including GPT-6 Astra and Claude Opus 5.5, revealed that while these models excelled at interpreting data, they struggled with explicitly choosing and ranking experiments, passing only 21% of criteria related to decision-making. The significance of this research lies in its focus on advancing the autonomy of AI in biological experimentation, a crucial step toward enabling AI to independently design and execute experiments. Despite the initial shortcomings, the findings suggest that AI agents possess latent knowledge that could be harnessed with improved training and methodologies. As agents like Astra show potential in decision-making, the pursuit of enhancing their capabilities could lead to breakthroughs in autonomous experimental design, ultimately transforming the speed and efficiency of scientific discovery in the life sciences.