AI
Knowing When to Yield: Grounded Arbitration of User Corrections in Text-Based Embodied Agents
Publication: cs.AI updates on arXiv.org
Author: Yezhou Cheng, Runjia Du, Zeming Liu, Hang Lyu, Zehua Yang, Bojun Lin
Source published: 02/10/2026, 04:00:00 UTC
Briefing created: 02/10/2026, 04:14:50 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 04:03:25 UTC.
Briefings
A 2026 study by GAVA investigates grounded arbitration for user corrections in text-based embodied agents. The research defines a choice between accepting, rejecting, inspecting the world, or asking the speaker. In ALFWorld, 162 checkpoints yielded 972 paired true/false interventions. GAVA achieved 100% correction accuracy through complete local inspections, establishing a verification contract. An exploratory training-only object-location prior reduced interaction and joint costs by 0.490 and 0.420 on 340 unseen scenarios. After freezing the policy, these gains replicated on 77 non-overlapping seen checkpoints covering 308 scenarios with 95% confidence intervals excluding zero. Semantic GAVA showed 98.8% accuracy but did not establish a general advantage of environmental value of information over clarification.
Key points
- The study defines grounded correction arbitration as a choice among accepting, rejecting, inspecting the world, and asking the speaker.
- In ALFWorld, 162 checkpoints produced 972 paired true and false interventions with 100% correction accuracy via complete local inspections.
Context and limitations
The study uses normalized claims, complete symbolic observations, and controlled speakers; it evaluates neither human participants, visual input, nor physical robots.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/10/2026, 04:14:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiebf6zxfyqvegiw67sxdx7a3xerm5rd3bq4fa2q76v4mnt5e6bl3q