Knowing When to Yield: Grounded Arbitration of User Corrections in Text-Based Embodied Agents
When faced with user corrections, how should an embodied agent respond? A new study formulates correction arbitration as a choice between accepting, rejecting, inspecting the world, or asking the speaker for clarification. The GAVA system implements this approach using observation-bounded evidence and legal probes, following a one-step expected-loss rule. In text-only ALFWorld experiments, 162 checkpoints generated 972 paired true and false interventions. Complete local inspections confirmed 100% correction accuracy, establishing the evidence contract rather than relying on a comparative advantage. GAVA reduced interaction costs compared to always verifying but tied under perfect speaker conditions. An exploratory training-only object-location prior further lowered interaction and joint cost in unseen scenarios by 49%. For more details, see the original paper at [arXiv:2610.00282v1](https://arxiv.org/abs/2610.00282).
When faced with user corrections, how should an embodied agent respond? A new study formulates correction arbitration as a choice between accepting, rejecting, inspecting the world, or asking the speaker for clarification. The GAVA system implements this approach using observation-bounded evidence and legal probes, following a one-step expected-loss rule. In text-only ALFWorld experiments, 162 checkpoints generated 972 paired true and false interventions. Complete local inspections confirmed 100% correction accuracy, establishing the evidence contract rather than relying on a comparative advantage. GAVA reduced interaction costs compared to always verifying but tied under perfect speaker conditions. An exploratory training-only object-location prior further lowered interaction and joint cost in unseen scenarios by 49%. For more details, see the original paper at arXiv:2610.00282v1.
Sources
- arXiv cs.AI — Knowing When to Yield: Grounded Arbitration of User Corrections in Text-Based Embodied Agents
由 VictoriaPark 自主 AI 编辑团队撰写；每项事实主张均链接来源，观点与报道严格分开。