The Language Agent Biology Benchmark (LAB-Bench) is a dataset of over 2,400 multiple-choice questions designed to evaluate AI systems on practical biology research tasks, including literature recall, figure interpretation, database access, and DNA/protein sequence analysis. Unlike existing benchmarks that focus on textbook-style questions, LAB-Bench assesses real-world scientific tasks, such as literature search and molecular cloning. The benchmark includes tasks like LitQA2, SuppQA, FigQA, TableQA, DbQA, SeqQA, and Cloning Scenarios, covering a wide range of biological research capabilities. The dataset is divided into categories, with 41 Cloning Scenarios being particularly challenging, requiring multi-step reasoning and domain-specific knowledge. The study evaluates several frontier models, including commercial and open-source LLMs, against LAB-Bench, comparing their performance to human experts in biology. Results show that while models perform reasonably well on some tasks, they struggle with complex sequence manipulation and protocol troubleshooting. Human experts outperformed models in most categories, especially in tasks requiring detailed biological knowledge. The study highlights the importance of high-quality distractors in benchmarking, as models often guess answers rather than reason through them. It also notes the difficulty in obtaining reliable human baselines for tasks like Cloning Scenarios, which require specialized expertise and time. The authors suggest that future benchmarks may rely on "human proofs of possibility" as models improve. LAB-Bench is publicly available for use, and the authors plan to continue updating and expanding it. The benchmark aims to serve as a tool for developing automated research systems and to evaluate the capabilities of AI in scientific research. The study underscores the need for domain-specific evaluation strategies as AI systems become more integrated into scientific workflows.The Language Agent Biology Benchmark (LAB-Bench) is a dataset of over 2,400 multiple-choice questions designed to evaluate AI systems on practical biology research tasks, including literature recall, figure interpretation, database access, and DNA/protein sequence analysis. Unlike existing benchmarks that focus on textbook-style questions, LAB-Bench assesses real-world scientific tasks, such as literature search and molecular cloning. The benchmark includes tasks like LitQA2, SuppQA, FigQA, TableQA, DbQA, SeqQA, and Cloning Scenarios, covering a wide range of biological research capabilities. The dataset is divided into categories, with 41 Cloning Scenarios being particularly challenging, requiring multi-step reasoning and domain-specific knowledge. The study evaluates several frontier models, including commercial and open-source LLMs, against LAB-Bench, comparing their performance to human experts in biology. Results show that while models perform reasonably well on some tasks, they struggle with complex sequence manipulation and protocol troubleshooting. Human experts outperformed models in most categories, especially in tasks requiring detailed biological knowledge. The study highlights the importance of high-quality distractors in benchmarking, as models often guess answers rather than reason through them. It also notes the difficulty in obtaining reliable human baselines for tasks like Cloning Scenarios, which require specialized expertise and time. The authors suggest that future benchmarks may rely on "human proofs of possibility" as models improve. LAB-Bench is publicly available for use, and the authors plan to continue updating and expanding it. The benchmark aims to serve as a tool for developing automated research systems and to evaluate the capabilities of AI in scientific research. The study underscores the need for domain-specific evaluation strategies as AI systems become more integrated into scientific workflows.