Ex Ante Task Difficulty Prediction Across 17 Agentic Benchmarks
August 7, 2026
This research demonstrates that task difficulty can be predicted from task descriptions without executing costly simulations. By using token-level entropy as a predictive signal, the method enables the construction of progressive training curricula across coding, math, and web navigation domains.
HOW THIS AFFECTS YOU
●
builderYou can optimize training pipelines by using difficulty forecasts to select tasks for curriculum learning.
●
researcherYou can use entropy-based metrics to better evaluate the difficulty of new agentic benchmarks.