GAD-RL
A reinforcement learning training method for improving OCR transcription faithfulness. It dynamically regulates teacher distillation intensity through three mechanisms—gating, attenuation, and probabilistically weighted KL divergence—built on the GRPO framework and combining on-policy distillation to co-optimize task reward and teacher guidance.
Timeline (last 90 days)
The arXiv paper "Improving OCR Faithfulness via Gated and Attenuated On-Policy Distillation" proposes the GAD-RL method
GAD-RL, short for Gated and Attenuated Distillation RL, dynamically adjusts the intensity of teacher supervision based on the student's current task performance and local distribution during joint post-training
The paper observes that supervision signals from a fixed teacher become increasingly less worthwhile as the student model's capabilities improve
GAD-RL's gating mechanism: when a response group contains outputs with a task reward of 0.95 or above, distillation for that group is directly disabled
GAD-RL's attenuation mechanism: as the average reward within a group increases, distillation intensity is continuously weakened, maintaining a smooth inverse relationship between supervision intensity and student capability
GAD-RL weights the forward KL divergence using the student's predicted probability for the teacher's Top-1 token, suppressing local auxiliary updates when the student's support for the teacher's recommended token is low
On the Qwen3.5-2B model, GAD-RL achieved a Micro Recall of 59.92% on CHAOS-Bench, surpassing GRPO and fixed-weight GRPO+OPD by 8.45 and 4.43 percentage points respectively
GAD-RL achieved an Overall total score of 91.18 on OmniDocBench v1.6
The paper's validation is mainly focused on OCR faithfulness scenarios and the Qwen3.5-2B scale model, and generalization to larger-scale models and more diverse tasks has not yet been verified
All Facts (9)
The arXiv paper "Improving OCR Faithfulness via Gated and Attenuated On-Policy Distillation" proposes the GAD-RL method
50%UnverifiedGAD-RL, short for Gated and Attenuated Distillation RL, dynamically adjusts the intensity of teacher supervision based on the student's current task performance and local distribution during joint post-training
50%UnverifiedThe paper observes that supervision signals from a fixed teacher become increasingly less worthwhile as the student model's capabilities improve
50%UnverifiedGAD-RL's gating mechanism: when a response group contains outputs with a task reward of 0.95 or above, distillation for that group is directly disabled
50%UnverifiedGAD-RL's attenuation mechanism: as the average reward within a group increases, distillation intensity is continuously weakened, maintaining a smooth inverse relationship between supervision intensity and student capability
50%UnverifiedGAD-RL weights the forward KL divergence using the student's predicted probability for the teacher's Top-1 token, suppressing local auxiliary updates when the student's support for the teacher's recommended token is low
50%UnverifiedOn the Qwen3.5-2B model, GAD-RL achieved a Micro Recall of 59.92% on CHAOS-Bench, surpassing GRPO and fixed-weight GRPO+OPD by 8.45 and 4.43 percentage points respectively
50%UnverifiedGAD-RL achieved an Overall total score of 91.18 on OmniDocBench v1.6
50%UnverifiedThe paper's validation is mainly focused on OCR faithfulness scenarios and the Qwen3.5-2B scale model, and generalization to larger-scale models and more diverse tasks has not yet been verified
50%