RealFeedApp Average score across the held-out subset was 46.7%
The SolCodex agent recovered 46.7% of the physical workflow, including measuring cell activity and returning trays with lids. Other models achieved scores ranging from 20.9% to 5.5%.
- Field
- robotics
- What they did
- Researchers created a benchmark platform to evaluate how well language models can control robots in wet lab settings, specifically testing a robot that had interrupted an experiment with cell trays.
- Why it matters
- This helps determine how effectively AI can restore complex physical protocols, such as measuring cell activity and handling spoiled samples, which is critical for automating biological research.
#lab robotics#solcodex#wet lab#gpt-6.1#benchmarking
