A new framework lets text-based AI agents decide whether to accept, reject, verify, or question a human's correction, instead of just obeying it.
The system, called GAVA, was tested in ALFWorld, a text-based simulated household environment, across 162 checkpoints producing 972 paired true and false corrections. When it could fully inspect its surroundings, GAVA matched a baseline that always double-checks: both hit 100 percent correction accuracy. The bigger result came from adding a trained prior on likely object locations, which cut interaction cost by about 49 percent and overall declared cost by about 42 percent on 340 unseen scenarios, with comparable gains of roughly 60 percent and 52 percent replicating on a separate set of 308 scenarios. A simplified, language-only version of the system made four factual errors per test group, for about 98.7 to 98.8 percent accuracy, and every method completed every task.
Plenty of AI products already let a person overrule the machine, from voice assistants to coding tools, but few have a formal way to weigh when checking the facts is worth the cost versus when it's cheaper to just ask or just comply. As agents take on more real-world tasks, that calculation matters: constant double-checking becomes its own burden on the person giving instructions, while blind deference lets bad corrections stand unchallenged.
Still, this is a bench result, not a deployed assistant: no humans, no robots, no visual input, only simulated speakers in a text-only world.