Research map / Vision-language and multimodal / Cross-modal learning
TabGLM: Tabular Graph Language Model for Learning Transferable Representations Through Multi-Modal Consistency Minimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TabGLM, a multi-modal tabular model that encodes each row as a fully connected graph (GNN, trained) and serialized text (frozen TAPAS), aligned via a contrastive consistency loss (MuCosa) plus supervised loss, evaluated on 25 tabular classification benchmarks.
Problem
Deep learning models for tabular data, especially with heterogeneous numerical and categorical columns, often underperform linear and tree-based models; uni-modal table transformations (image, text, or graph) capture only one type of information.
Approach
Jointly encode each row as a fully connected feature graph (trained GNN) and as serialized text (frozen pretrained TAPAS encoder), and align the two embeddings with a symmetric InfoNCE-style consistency loss while training a classifier on graph embeddings only.
Key results
- TabGLM averages 89.47 AUC-ROC across 25 datasets versus 88.22 for NODE and 84.69 for LR, per Table 1.
- On the 9 TabLLM benchmark datasets TabGLM averages 89.26 vs 87.91 for TabLLM and 81.30 for IGNNet.
- Ablation: TabGLM with TAPAS encoder (129M params) reaches 82.82 on pc3 vs 77.04 (graph only) and 78.24 (text only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine graph and text encoders instead of using either encoder alone (ablation) | TabGLM · pc3 | AUCROC: 77.04 -> 82.82; 78.24 -> 82.82 | Helped |
| combine graph and text encoders instead of using either encoder alone (ablation) | TabGLM · bank | AUCROC: 91.11 -> 92.07; 90.52 -> 92.07 | Helped |
| combine graph and text encoders instead of using either encoder alone (ablation) | TabGLM · Creditg | AUCROC: 71.99 -> 79.32; 77.36 -> 79.32 | Helped |
| replace TAPEX with TAPAS as the text encoder (ablation) | TabGLM · blood | AUC-ROC: 77.57 -> 78.48 | Helped |
| replace TAPEX with TAPAS as the text encoder (ablation) | TabGLM · calhousing | AUC-ROC: 95.29 -> 95.47 | Helped |
| replace TAPEX with TAPAS as the text encoder (ablation) | TabGLM · creditg | AUC-ROC: 78.72 -> 79.32 | Helped |
| replace LabelEncoder with One-Hot Encoder (ablation) | TabGLM · blood | AUC-ROC: 78.29 -> 78.49 | Helped |
| replace LabelEncoder with One-Hot Encoder (ablation) | TabGLM · calhousing | AUC-ROC: 95.34 -> 95.47 | Helped |
| replace LabelEncoder with One-Hot Encoder (ablation) | TabGLM · coil2000 | AUC-ROC: 73.68 -> 74.17 | Helped |
| replace LabelEncoder with One-Hot Encoder (ablation) | TabGLM · diabetes | AUC-ROC: 83.14 -> 83.70 | Helped |
| use the text-only pipeline instead of the full multimodal design (ablation) | TabGLM · car | AUC-ROC: 98.37 -> 99.40 | Helped |
| use TabGLM instead of TabLLM | tabular benchmark · TabLLM benchmark (9 datasets) | AUC-ROC (average): 87.91 -> 89.26 | Helped |
Surprising findings
- CatBoost and XGBoost outperform TabGLM on several datasets such as kr-vs-kp, pc3-like simple tables, and tic-tac-toe (XGBoost 99.98 vs 99.85).
- Text-only pipeline beat the full multi-modal model on the car dataset (99.40 vs 98.37 in appendix table), contradicting the multi-modal premise for categorical-only tables.
Limitations
- Number of input columns is capped by the LLM token limit (512 tokens for TAPAS).
- Graph construction uses column indices and statistical relationships but ignores column names.
- Graph-only IGNNet performs comparably or better on some purely numerical tables such as blood and calhousing in some comparisons.
Experimental setup
Datasets: bank blood calhousing car coil2000 creditg diabetes heart kr-vs-kp mfeat-fourier pc3 income
Benchmarks: bank blood calhousing car coil2000 creditg diabetes heart kr-vs-kp mfeat-fourier pc3 income
Metrics: AUC-ROC
Baselines: CatBoost XGBoost Gradient Boosting Random Forest Logistic Regression TabTransformer FT-Transformer NODE IGNNet TabLLM TabPFN
Builds on: IGNNet TAPAS TabLLM CLIP
Closest papers by meaning
- Large Scale Transfer Learning for Tabular Data via Language Modeling (NeurIPS 2024 · similarity 0.90)
- TabSTAR: A Tabular Foundation Model for Tabular Data with Text Fields (NeurIPS 2025 · similarity 0.90)
- TabM: Advancing tabular deep learning with parameter-efficient ensembling (ICLR 2025 · similarity 0.89)
- Making Pre-trained Language Models Great on Tabular Prediction (ICLR 2024 · similarity 0.89)
- Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings (NeurIPS 2025 · similarity 0.88)
- TabFlex: Scaling Tabular Learning to Millions with Linear Attention (ICML 2025 · similarity 0.88)
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.