This paper investigates how a text‑based embodied agent should react when a user’s correction might be wrong. We formulate the problem as grounded correction arbitration, offering four actions: accept, reject, inspect the world, or ask the speaker. GAVA implements this interface using observation‑bounded evidence, legal probes, and a one‑step expected‑loss rule.
In the text‑only ALFWorld benchmark we created 162 checkpoints, yielding 972 paired true and false interventions. Full local inspections enable GAVA to verify correction accuracy at 100%, establishing an evidence contract rather than a comparative edge.
During same‑episode execution, GAVA reduces interaction cost compared with an always‑verify baseline, yet a cost threshold remains under a perfect speaker. Adding a training‑only object‑location prior lowers interaction and declared joint costs by 0.490 and 0.420 respectively on 340 unseen scenarios relative to uniform GAVA.
After freezing the policy, the gains replicate on 77 non‑overlapping seen checkpoints (308 scenarios): interaction cost drops by 0.595 and joint cost by 0.517, with 95% checkpoint‑bootstrap confidence intervals excluding zero. Joint cost also improves over an identical‑prior fixed policy, while the calibrated no‑VOI comparison stays inconclusive.
Semantic GAVA makes four factual errors per cohort, yielding accuracies of 98.8% and 98.7%; all methods complete every task. The results support selective information gathering with semantic priors under declared costs, but do not establish a general advantage of environmental value of information over clarification. The study uses normalized claims, complete symbolic observations, and controlled speakers, without human participants, visual input, or physical robots.
Review