posterior-probing

GridProbe is a training-free inference paradigm for Long-Video VLMs that adaptively selects relevant frames using posterior probing, achieving sub-quadratic attention costs with minimal accuracy loss.