AgentVQA: Offline Benchmark with 1,000 Human-Verified Samples for Visual Agents

Researchers from Maryland and Princeton released AgentVQA, an offline benchmark with 1,000 human-verified samples from real agent scenarios, spanning web agents, robotics, egocentric video, gaming, and spatial planning to evaluate VLM agents without costly online runs.

2026-10-23 ~ 2026-10-24 · 2 related posts

• AgentVQA: a unified benchmark finds visual grounding is the top failure mode for VLM agents — zlab-princeton · 2026-10-23
• AgentVQA: 1,000 human-verified examples to evaluate visual agents offline across 5 domains — liuzhuang1234 · 2026-10-24