BenchBench-Cloning is a benchmark for large language models focused on molecular cloning, comprising 65 tasks derived from sequence records, lab notebooks, and published studies. The tasks fall into two categories: 50 protocol-development tasks where the model must produce an executable protocol for a given construct, and 15 construct-design tasks where the model must infer the constructs required for a scientific objective. Protocol tasks are evaluated by running them in a cloning simulator, while construct tasks use deterministic checks or rubric‑based LLM-as-a-judge scoring. Across eight systems with repeated attempts, GPT-6 Astra achieved the highest success rate at 30.6%, with other models ranging from 8.6% to 22.2%; even best‑of‑ten sampling does not saturate the benchmark. Integrating individual cloning decisions into complete, valid plans remains difficult even for frontier models, making BenchBench-Cloning a valuable long‑term benchmark for tracking progress in experimental planning.
Review