DualChem Benchmark (scale.com)

Scale has launched the DualChem benchmark, a groundbreaking evaluation tool designed to measure the dual-use risk awareness of large language models (LLMs) specifically in the field of chemistry. Unlike traditional benchmarks that focus on general capabilities, DualChem assesses whether these models can discern and provide responses to questions that could be used for both legitimate scientific research and harmful activities. This is particularly significant as government agencies increasingly seek to understand the potential misuses of advanced AI systems. The benchmark consists of 600 expert-validated questions that represent realistic, ambiguous dual-use scenarios encountered in chemistry, ensuring that the evaluation reflects true applications rather than synthetic prompts. Initial findings from testing over 30 state-of-the-art models reveal that while these LLMs perform relatively well with an average accuracy of around 50%, they often provide substantive answers even when they flag questions as potentially dual-use. This indicates a disconnect between risk identification and refusal behavior, highlighting the limitations of existing refusal-based safety measures. Moreover, model responses showed weak alignment with human expert judgments regarding the severity of dual-use applications, underlining the need for rigorous assessment frameworks. DualChem not only fills a critical gap in AI evaluation but also emphasizes the importance of partnerships between industry, academia, and government to develop better methodologies for monitoring and ensuring the safe deployment of powerful AI technologies.