Ask, Don't Guess: Learning When to Clarify in Situated Instruction Following

Language agents in situated environments must decide not only what an instruction might mean, but whether it is useful and safe to act on that interpretation. Existing clarification benchmarks often treat this as static ambiguity detection. We instead frame clarification as a situated value-of-information decision under uncertainty, interaction cost, and task consequences. We introduce Clarify-to-Act, a procedural benchmark in which an agent observes a structured scene and instruction, then either acts immediately or asks one clarifying question before acting. Rewards are deterministic and combine task success, clarification cost, and wrong-action cost. Across different models, direct ask-when-needed prompting improves over guessing but remains miscalibrated: it over-asks when ambiguity is harmless and under-asks when wrong actions are costly. An expected communicative utility (ECU) policy reaches 0.98 utility with no missed or unnecessary clarifications in these main runs. The central lesson is that pragmatic clarification should be evaluated as utility-calibrated interaction, not ambiguity detection alone.