Research map / Label-efficient and robust learning / Semi-supervised learning
LLM-Guided Probabilistic Fusion for Label-Efficient Document Layout Analysis
Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LLM structural priors via fusion to teacher pseudo-labels (ablation) | SwiftFormer-DETR semi-supervised · PubLayNet (5% labels) | AP: 82.3 -> 87.3 | Improved |
| remove LLM guidance (ablation) | Full method · PubLayNet (5% labels) | AP: 84.3 vs 87.3 (-3.0) | Improved |
| add probabilistic fusion (ablation) | Teacher+LLM pseudo-labels · PubLayNet (5% labels) | AP: 85.9 -> 86.7 | Improved |
| add cross-modal consistency loss (ablation) | Fusion framework · PubLayNet (5% labels) | AP: 86.7 -> 87.3 | Improved |
| use learned instance-adaptive gating instead of fixed weights (ablation) | Lightweight SwiftFormer · PubLayNet (5% labels) | AP: 87.3 -> 88.2 | Improved |
| use learned gating instead of fixed weights (ablation) | LayoutLMv3 variant · PubLayNet (5% labels) | AP: 89.4 -> 89.7 | Improved |
| add LLM fusion to LayoutLMv3 semi-supervised learning | LayoutLMv3 · PubLayNet (5% labels) | AP: 89.1 -> 89.7 | Improved |
| use MLP fusion instead of inverse-variance plus gate (ablation) | Fusion · PubLayNet (5% labels) | AP: 88.2 -> 87.6 | Hurt |
| use open-source Llama-3-70B instead of GPT-4o-mini (ablation) | LLM component · PubLayNet (5% labels, LayoutLMv3) | AP: 89.7 -> 89.4 | No clear effect |
| use regex text patterns instead of LLM (ablation) | Text prior · PubLayNet (5% labels) | AP: 88.2 vs 84.9 for regex | Hurt |
| add LLM guidance to semi-supervised baseline on rare classes | SoftTeacher · DocLayNet (5% labels) | Caption AP: 70.1 -> 76.8 | Improved |
Closest papers by meaning
- M2Doc: A Multi-Modal Fusion Approach for Document Layout Analysis (AAAI 2024 · similarity 0.90)
- LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding (CVPR 2024 · similarity 0.89)
- A Simple yet Effective Layout Token in Large Language Models for Document Understanding (CVPR 2025 · similarity 0.89)
- DocLayLLM: An Efficient Multi-modal Extension of Large Language Models for Text-rich Document Understanding (CVPR 2025 · similarity 0.88)
- Relation-Rich Visual Document Generator for Visual Information Extraction (CVPR 2025 · similarity 0.87)
- Graph-based Document Structure Analysis (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
