Show HN: DoubtBench – does Jev know when humans disagree? (github.com)

DoubtBench, a new benchmark for evaluating AI decision models, focuses on human uncertainty rather than merely assessing correctness against a single right answer. It utilizes human-rated data from NVIDIA's HelpSteer2 to create approximately 7,500 questions that measure not only a model's accuracy but also how well its stated probabilities align with areas of human disagreement. This is significant because it challenges traditional benchmarks that might ignore the nuances of human decision-making, offering a more holistic view of a model's performance. The results highlight that currently favored models, like Jev and various Claude iterations, show minimal correlation between uncertainty and the disagreement found in human responses. Jev 1.13 emerges as the most efficient option, achieving higher accuracy and lower operational costs compared to the Claude models. With a unique scoring system that balances accuracy and human agreement, DoubtBench encourages the development of models that can effectively recognize when to defer to human judgment, a critical aspect for applications requiring high reliability and trust in AI decision-making.