AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Geometry-Aware Tabular Diffusion

David Turtora Zagardo

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 0 citations (OpenAlex)

Read the paper

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 changedWhereMeasuredEffect
add supervised geometric losses (angle, length, consistency) to MLP denoiser (ablation)Residual MLP diffusion · 10 tabular datasetsShape error: 1.279 -> 0.862 (InputsOnly -> Full)Improved
add geometric inputs and heads without supervision (ablation)Residual MLP diffusion · 10 tabular datasetsShape error: 1.229 -> 1.279 (d = -0.08)No clear effect
add geometric supervision to GNN denoiserGNN+LE diffusion · 10 tabular datasetsShape error: 1.23 -> 1.17Mixed
add geometric supervision to Transformer denoiserColumn-wise Transformer diffusion · 10 tabular datasetsShape error: 1.23 -> 0.959Improved
add geometric supervision to GNN denoiserGNN+LE diffusion · 10 tabular datasetsTrend error: 5.72 -> 4.19Improved
replace TabDiff transformer with MLP+GeomTabDiff · 10 tabular datasetsShape error: 1.19 -> 0.862Improved
replace TabDiff transformer with MLP+GeomTabDiff · 10 tabular datasetsMLE-1 gap to real: 11.4% -> 4.8%Improved
match TabDiff training budget at 8,000 epochs (ablation)GATD-MLP · 10 tabular datasetsShape error: 0.862 -> 0.854No clear effect
reduce sampling steps from 1000 to 50 (ablation)GATD-MLP · 10 tabular datasetsShape error: 0.862 -> 1.546Hurt
reduce sampling steps from 1000 to 100 (ablation)GATD-MLP · 10 tabular datasetsShape error: 0.862 -> 1.125Hurt
use geometric supervision on California with MLPGATD-MLP · CaliforniaShape error: 22.8 percentage points worse than TabDiffMixed
raise geometric loss weights to dominate diffusion loss (ablation)GATD · 10 tabular datasetsfidelityImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy