Agent Fidelity
A dimension for evaluating world models that measures how accurately a model reproduces human poses, motion trajectories and behavioral intent. High agent fidelity indicates that predicted or generated human motion remains plausible and consistent with the underlying intentions, which matters for applications such as embodied simulation, animation and human-robot interaction.
Timeline (last 90 days)
After scaling the training data by 100 times, both agent fidelity and object fidelity improved, but object fidelity is far lower than agent fidelity and improves much more slowly.
A carefully designed visual conditioning scheme can saturate agent fidelity using only a small portion of the data.
Agent fidelity measures whether the model can accurately reproduce a person's pose, motion trajectory, and behavioral intent; object fidelity measures whether the model can correctly predict changes in an object's state under external forces or interactions.
All Facts (3)
After scaling the training data by 100 times, both agent fidelity and object fidelity improved, but object fidelity is far lower than agent fidelity and improves much more slowly.
50%UnverifiedA carefully designed visual conditioning scheme can saturate agent fidelity using only a small portion of the data.
50%UnverifiedAgent fidelity measures whether the model can accurately reproduce a person's pose, motion trajectory, and behavioral intent; object fidelity measures whether the model can correctly predict changes in an object's state under external forces or interactions.
50%