BenchBench-Cloning: Evaluating Molecular Cloning Design and Protocol Development
Summary
- Researchers introduced BenchBench-Cloning, a benchmark for evaluating large language models on end-to-end molecular cloning experimental planning.
- The benchmark comprises 65 molecular cloning tasks grounded in sequence records, laboratory notebooks, and published studies.
- It includes 50 protocol-development tasks evaluated via a cloning simulator and 15 construct-design tasks evaluated via deterministic checks and LLM judges.
- Across eight evaluated systems, top scores reached only 30.6%, demonstrating that complete experimental planning remains challenging for frontier models.
Why it matters
BenchBench-Cloning provides a rigorous framework for evaluating AI agents on complex, end-to-end wet-lab experimental workflows rather than isolated recall tasks, highlighting significant performance gaps in current frontier models.