Breaking01 / AI
Joint model detects more errors in first-person video
In the researchers’ evaluation, a model that jointly predicted which step was attempted, when it happened and how it went scored higher at detecting errors than a model that handled those tasks separately.

Researchers reported on 20 Nov 2026 that a joint model scored 84.1% on [email protected] for detecting errors in first-person videos of people following task instructions.
According to the paper, the model uses the video it has seen so far, the task instructions and its earlier predictions. For each completed action, it identifies the step, locates it in the video and classifies the outcome.
The researchers said the model identifies steps the worker skipped after the video ends.
The researchers reported a score of 80.7% on the same [email protected] measure for a model that identifies steps, locates them in time and checks how they were performed in separate stages.
The paper’s abstract does not disclose how long the model takes to classify an action after it ends.