Knowing When to Yield: Grounded Arbitration of User Corrections in Embodied Agents (GAVA)

Forum topic · 小凯 · 2026-10-05

Summary

This paper formulates grounded correction arbitration for embodied agents: when a human correction may be wrong, the agent must choose between accepting, rejecting, inspecting the world, or asking the speaker. The proposed GAVA framework implements this interface using observation-bounded evidence, legal probes, and a one-step expected-loss rule. In text-only ALFWorld, 162 checkpoints produce 972 paired true and false interventions. Complete local inspections give GAVA and an always-verify baseline 100% correction accuracy, establishing an evidence contract rather than a comparative advantage. In same-episode execution, GAVA reduces interaction cost versus always-verify but ties a cost threshold under a perfect speaker. A training-only exploratory object-location prior cuts interaction and declared joint costs by 0.490 and 0.420 on 340 unseen scenes versus uniform, and gains persist after freezing the policy on 77 disjoint seen checkpoints (0.595 and 0.517, 95% bootstrap intervals excluding zero). Semantic GAVA makes four factual errors per queue (98.8% and 98.7% accuracy). arXiv: 2610.00282.

Paper Overview

Field: ML Authors: Yezhou Cheng, Runjia Du, Zeming Liu, Hang Lyu, Zehua Yang, Bojun Lin Published: 2026-10-05 arXiv: 2610.00282

Abstract

How should an embodied agent respond when a person's correction may be wrong? The authors formulate grounded correction arbitration as a choice among accepting, rejecting, inspecting the world, and asking the speaker. GAVA implements this interface with observation-bounded evidence, legal probes, and a one-step expected-loss rule.

In text-only ALFWorld, 162 checkpoints produce 972 paired true and false interventions. Complete local inspections give GAVA and the always-verify baseline 100 percent correction accuracy — establishing the evidence contract rather than a comparative advantage. In same-episode execution, GAVA reduces interaction cost against always-verify but ties a cost threshold under a perfect speaker.

An exploratory training-only object-location prior lowers interaction and declared joint cost by 0.490 and 0.420 relative to uniform on 340 unseen scenes. After freezing the policy, gains replicate on 77 non-overlapping seen checkpoints (308 scenes): 0.595 and 0.517, with 95% checkpoint-bootstrap intervals excluding zero. Joint cost also beats a fixed policy with the same prior, while calibrated no-VOI comparisons remain uncertain. Semantic GAVA makes four factual errors per queue, corresponding to 98.8% and 98.7% accuracy; all methods complete all tasks.

Takeaways

Results support selective information gathering with semantic priors under declaration cost, but do not establish a universal advantage of environment information value over clarification questions.
The study uses normalized claims, full symbolic observations, and controlled speakers; human participants, visual input, and physical robots were not evaluated.

--- *Auto-collected on 2026-10-05*