BRIDGE: Predicting Human Task Completion Time From Model Performance

Researchers specializing in artificial intelligence evaluation introduced BRIDGE, a novel psychometric framework, in a new technical paper published on the arXiv preprint server in February 2025 to revolutionize how we measure AI capabilities through the lens of human task completion time. The development aims to address the growing disconnect between abstract AI benchmark scores and practical human-interpretable measures of difficulty, providing a more tangible understanding of how machine intelligence relates to human labor. By grounding model performance in the time it would take a person to complete a specific task, the framework offers a standardized alternative to traditional, labor-intensive annotation methods. Technically, BRIDGE operates by identifying a latent difficulty scale extracted directly from model responses and anchoring that scale to human completion metrics. Historically, calculating these time-based metrics required hiring human subjects to manually complete and log every task within a benchmark, a process that proved prohibitively expensive, prone to subjective noise, and nearly impossible to scale alongside the rapid release of massive new datasets. BRIDGE bypasses these bottlenecks by using the statistical distribution of AI errors and successes to predict the temporal investment a human would typically need. This unified psychometric approach holds significant implications for the industry as it allows developers to quantify the efficiency gains or limitations of large language models and other AI systems in real-world professional contexts. Instead of relying on percentage-based accuracy scores which lack context, stakeholders can now utilize the BRIDGE framework to estimate the economic value and complexity of tasks an AI is capable of handling, effectively translating silicon performance into human units of work.