mlsanityData to 7 Oct 2026

Improving OCR Faithfulness via Gated and Attenuated On-Policy Distillation

arXiv 2609.38282 ↗First posted 29 Sep 2026cs.AIcited by 0

On the map: Multimodal LLMs / vision-language models → Prompt learning and adaptation of CLIP also fits: MLLM jailbreaks and adversarial robustness

Abstract (arXiv)

Vision-language models may rewrite anomalous text in images into linguistically plausible expressions, compromising OCR transcription faithfulness. Sequence-level task rewards and local teacher guidance are complementary, but guidance from the same teacher may not remain equally effective as the student improves. Offline analysis shows that supervision from a fixed teacher becomes progressively less favorable as the student improves, both across training checkpoints and across response groups with different task rewards. Motivated by this observation, we introduce GAD-RL, which adaptively regulates teacher supervision during joint post-training according to the student's current task performance and local distributions. A frozen teacher conditions on reference transcriptions and student-generated prefixes. GAD-RL 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. It also weights forward KL by the student's probability of the teacher's Top-1 token, moderating local auxiliary updates when student support for that candidate is low. On Qwen3.5-2B, GAD-RL achieves 59.92% Micro Recall on CHAOS-Bench, surpassing GRPO and GRPO+OPD (fixed-weight) by 8.45 and 4.43 percentage points, respectively, while achieving an Overall score of 91.18 on OmniDocBench v1.6.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Prompt learning and adaptation of CLIP, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 32 of its 34 references that are in the map. First posted on arXiv 29 Sep 2026.

Builds on

32 of 34 references.

References · 32

  1. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  2. 2025Qwen2.5-VL Technical Reportcited by 5,418
  3. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,678
  4. 2023On-Policy Distillation of Language Models: Learning from Self-Generated Mistakescited by 840
  5. 2023Aligning Large Multimodal Models with Factually Augmented RLHFcited by 703
  6. 2023Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decodingcited by 671
  7. 2019LayoutLM: Pre-training of Text and Layout for Document Image Understandingcited by 549
  8. 2023OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocationcited by 468
  9. 2023RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedbackcited by 404
  10. 2021OCR-free Document Understanding Transformercited by 400
  11. 2026Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Modelscited by 386
  12. 2022Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understandingcited by 360
20 more
  1. 2023Nougat: Neural Optical Understanding for Academic Documentscited by 240
  2. 2023Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimizationcited by 223
  3. 2024OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive Annotationscited by 144
  4. 2025MonkeyOCR: Document Parsing with a Structure-Recognition-Relation Triplet Paradigmcited by 81
  5. 2025The Hidden Life of Tokens: Reducing Hallucination of Large Vision-Language Models via Visual Information Steeringcited by 62
  6. 2026Stable On-Policy Distillation through Adaptive Target Reformulationcited by 47
  7. 2024Mitigating Object Hallucination in MLLMs via Data-augmented Phrase-level Alignmentcited by 42
  8. 2025olmOCR 2: Unit Test Rewards for Document OCRcited by 41
  9. 2025KDRL: Post-Training Reasoning LLMs via Unified Knowledge Distillation and Reinforcement Learningcited by 35
  10. 2026PACED: Distillation and On-Policy Self-Distillation at the Frontier of Student Competencecited by 26
  11. 2026PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Trainingcited by 20
  12. 2026HunyuanOCR-1.5: Making Lightweight OCR VLMs Faster and Bettercited by 14
  13. 2026Reward-Gated On-Policy Distillationcited by 11
  14. 2026Teacher-Guided Policy Optimization for On-Policy Reasoning Distillation under Large Policy Divergencecited by 9
  15. 2026Distill Where You Fail: Recovering Learning Signals of Negative RL-Groups from Adaptive Teacher Guidancecited by 8
  16. 2026Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvementcited by 8
  17. 2026OvisOCR2 Technical Reportcited by 5
  18. 2026Infinity-Parser2 Technical Reportcited by 4
  19. 2026Do VLMs Read or Rewrite? On Transcription Faithfulness in Vision-Language Modelscited by 2
  20. 2026I-SDPO: Instance-Level Adaptive Self-Distillation Policy Optimizationcited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026CA-OPD: Confidence-Aware On-Policy Distillation for Structured Visual Prediction
  2. 2026When Teachers Mislead: Spurious-Signal-Aware On-Policy Distillation
  3. 2026Distill What the Student Can See: Fisher-Projected On-Policy Distillation for Vision-Language Models
  4. 2026Your Teacher Can't Help You Here: Combating Supervision Fidelity Decay in On-Policy Distillation
  5. 2026DiffGate: Difficulty-Gated Teacher Guidance for On-Policy Distillation
  6. 2026Dense Supervision, Sparse Updates: On the Sparsity and Geometry of On-Policy Distillation
  7. 2025Masking Teacher and Reinforcing Student for Distilling Vision-Language Models
  8. 2026Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training