Breaking01 / AI
Meta researchers release 120-task test for wearable AI aides
The benchmark scores the final state of a virtual world rather than written answers, but the repository notes list no pass rates by model.

Meta's research organisation released InteractiveBench, a 120-task test for wearable AI assistants, for non-commercial use on 24 Oct 2026, according to explainx.ai.
The repository documentation fixes the tasks at 20 in each of six categories: Guidance, Planning, Accessibility, Decision, Prevention and Finding. The model under test plays the assistant and sees only the person's first-person view and the dialogue. A second, fixed model plays the person and carries out simulator actions on the assistant's guidance.
According to the documentation, the paper used Gemini 3.1 Pro in the human role for every assistant evaluation. A task passes when "every goal holds in the final world state". Success depends on the state of the virtual world at the end, not on the conversation.
The simulator, InteractiveWorld, uses NVIDIA Isaac Sim 5.1 and the scene dataset BEHAVIOR-1K 3.9.0, the documentation states. It offers 44 scenes of homes and venues and 36 families of human actions. Assistant models connect through a local open-source server or the OpenAI, Anthropic or Gemini APIs.
Running requirements. The repository documentation lists Linux with an NVIDIA GPU and about 100 GB of disk space, and one simulator worker uses about 4 GB of GPU memory. The main project is under CC BY-NC 4.0, a Creative Commons licence that requires credit and bars commercial use. The MANO body model and AMASS motion data require registration and are limited to non-commercial research.
The repository notes do not give pass rates for individual models.