Vision
11d agoGAD-RL Reduces OCR Hallucinations In Vision Models
The GAD-RL framework prevents vision-language models from rewriting anomalous image text into plausible but incorrect phrases. It adaptively regulates teacher supervision based on the student's real-time performance and local distributions. This approach stops fixed teacher guidance from hindering progress as models improve.
The Signal
Practitioners can now achieve higher transcription faithfulness during joint post-training.