DIVE-Doc: Downscaling Foundational Image Visual Encoder into Hierarchical Architecture for DocVQA
Bibliographic record
Abstract
In the DocVQA context, current end-to-end models ei-ther use lightweight architectures that run efficiently on small devices but have limited performance or rely on LVLMs that achieve high performance at significant computational cost. Thus, we present DIVE-Doc, an end-to-end model that bridges this gap by distilling a 400M-parameter SigLIP visual encoder into a small hierarchical Swin transformer, preserving LVLM performance with only one-fifth of the visual encoder's parameters. We investi-gate two distillation strategies: Fixed-Resolution Distillation (FRD), which matches teacher-student patch counts by forcing student input resolution, and Adaptive-Resolution Distillation (ARD), which aligns mismatched sequences via parameter-free interpolation, enabling various input reso-lutions. Fine-tuned with QLoRA, DIVE-Doc attains 82.7% ANLS, outperforming lightweight models and sitting within 2 ANLS of its teacher PaliGEMMA on DocVQA, while halving the teacher visual encoder's latency and supporting higher input resolutions. Analysis on RVL-CDIP and Do-cLayNet shows that the visual encoder captures document-level structure but delegates fine-grained layout reasoning to the language model decoder. The code is available at https://github.com/JayRay5/DIVE-Doc.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.002 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.015 | 0.006 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".