A research framework called BRIDGE can predict how long a human would take to complete a task — just by looking at how AI models perform on benchmarks.
The system applies a two-parameter logistic Item Response Theory model, a method borrowed from educational testing, to jointly estimate task difficulty and model capability from existing benchmark data. The key finding: latent task difficulty scales linearly with the logarithm of human completion time. That relationship lets researchers infer human time estimates for new benchmarks without running expensive, noisy human annotation studies. The authors also used BRIDGE to independently reproduce METR's result that the horizon of tasks a frontier model can half-solve doubles roughly every six months.
That reproduction matters. METR's exponential scaling result has become a frequently cited data point in AI capability forecasting, and an independent derivation via a completely different method adds weight to it — or at least rules out one class of methodological artifact. More practically, BRIDGE offers labs and evaluators a cheaper way to ground benchmark numbers in something a non-specialist can interpret: clock time, not abstract accuracy percentages.
The caveat is that "human completion time" is itself a rough proxy for difficulty — a task that takes an expert ten minutes might take a novice two hours. Whether the linear relationship holds across skill levels, domains, and task types outside the benchmarks tested here is the question the paper leaves open.