← All benchmarks

InteractiveBench (task pass)

Agents · 2026-10-25

InteractiveBench, from Meta and Duke University, is the benchmark of the InteractiveWorld environment: 120 frozen multi-turn tasks in six categories of 20 (Guidance, Planning, Accessibility, Decision, Prevention, Finding), set in Isaac Sim indoor scenes with an animated human. The model under test is an egocentric AI assistant: it sees only the Human's first-person frames and the dialogue, and at each turn stays silent, gives natural-language guidance or declares the task done, while a fixed simulated Human agent (Gemini 3.1 Pro in the paper) turns that guidance into executable actions. After a scripted warm-up the loop runs until the assistant says done or the interaction limit is reached, and the simulator then checks each task's required object states. The score is the percentage of the 120 tasks in which every required end state holds.

Top models (higher is better)