AI
Improving OCR Faithfulness via Gated and Attenuated On-Policy Distillation
Publication: cs.AI updates on arXiv.org
Author: Baode Wang, Zuming Huang, Kexuan Ren, Jun Huang, Wei Chu
Source published: 01/10/2026, 04:00:00 UTC
Briefing created: 01/10/2026, 04:16:21 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 04:03:24 UTC.
Briefings
This research presents GAD-RL, a method to improve OCR transcription faithfulness in vision-language models by adaptively regulating teacher supervision. The study analyzes that fixed teacher guidance becomes progressively less effective as the student improves, leading to the development of a dynamic system. On Qwen3.5-2B, GAD-RL achieved a Micro Recall of 59.92% on CHAOS-Bench, outperforming fixed-weight GRPO and GRPO+OPD by 8.45 and 4.43 percentage points respectively, while maintaining an Overall score of 91.18 on OmniDocBench v1.6.
Key points
- GAD-RL dynamically regulates teacher supervision strength based on the student's current task performance and local distributions.
- The method disables distillation for response groups containing an output with task reward at least 0.95 and continuously attenuates distillation strength as group-mean reward increases.
Context and limitations
The study covers only Qwen3.5-2B and uses a snapshot from October 1, 2026, with limited coverage across different training checkpoints and response groups.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 04:16:21 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeieygv7pnrep6ck47d25d57zaqd6y2fizdig66he4fdkujvzd6qqhu