Synthetic Underspecification for Agents
LHAW generates controllable underspecified long-horizon tasks by removing info across goals, constraints, inputs, context. Validates via agent trials, classifying ambiguity impacts. Releases 285 variants from benchmarks.






