Squint OPSD
An on-policy student-teacher distillation training method targeting multimodal models' deficiencies in low-resolution text recognition. During training, privileged context guides the student to establish correct spatial attention strategies; no additional auxiliary input is required at inference time.
Timeline (last 90 days)
Squint OPSD adopts a teacher-student distillation paradigm, where the teacher model receives a downsampled image view and is given the ground-truth text transcription as privileged context.
In Squint OPSD, the teacher's token-level distribution supervises the student model's learning along the on-policy trajectory.
After Squint OPSD training is complete, the student model faces only the original image during inference, requiring no auxiliary views or additional inputs.
On SquintBench, Squint OPSD achieves an ANLS score of 88.61.
All Facts (4)
Squint OPSD adopts a teacher-student distillation paradigm, where the teacher model receives a downsampled image view and is given the ground-truth text transcription as privileged context.
50%UnverifiedIn Squint OPSD, the teacher's token-level distribution supervises the student model's learning along the on-policy trajectory.
50%UnverifiedAfter Squint OPSD training is complete, the student model faces only the original image during inference, requiring no auxiliary views or additional inputs.
50%UnverifiedOn SquintBench, Squint OPSD achieves an ANLS score of 88.61.
50%