Hierarchical Synthetic Tabular Data Generation: A Hybrid Top-Down and Bottom-Up Framework

Conference: ICML 2026
arXiv: 2605.28198
Code: None
Area: Synthetic Tabular Data Generation / Multimodal Financial Data / Weak Multimodal Alignment / Controllable Generation
Keywords: Tabular Synthesis, Top-Down/Bottom-Up, Rule Constraints, LLM as Rule Generator, XGBoost Conditional Generation

TL;DR

This paper proposes the H-TDBU framework: utilizing LLMs or human-written rules in a top-down path to generate a "logical skeleton" \(\mathcal{S}\), while using lightweight bottom-up generators such as RandomForest, XGBoost, or CTGAN to learn "statistical texture" \(z\). These components are integrated via a conditional generator \(G(z\in\mathcal{Z}\mid\mathcal{S})\) and iteratively refined through a TSTR + XModal feedback loop. On weak multimodal financial benchmarks, it achieves TSTR AUROC superior to pure neural network baselines while maintaining cross-modal consistency.

Background & Motivation

Background: Synthetic tabular data generation is a primary solution for mitigating data scarcity, privacy constraints, and multi-source heterogeneity. Two main paradigms exist: pure generative models like CTGAN, TVAE, TabDDPM, and STaSy, which directly learn and sample from the real data distribution; and LLM-based approaches like GReaT, REaLTabFormer, and TabuLa, which serialize table rows into strings for in-context generation.

Limitations of Prior Work: Pure generative models are prone to overfitting and mode collapse in data-scarce scenarios (e.g., finance, healthcare), often smoothing out long-tail events like fraud. LLM-based approaches, due to autoregressive token decoding, cannot explicitly enforce structural consistency or functional dependencies, frequently producing "plausible but logically invalid" samples with weak modeling of long-range feature dependencies and numerical-categorical interactions. Additionally, recent research indicates that recursive training on synthetic data leads to model collapse (Shumailov et al. 2024 Nature).

Key Challenge: Synthetic tabular data inherently requires two capabilities: logical controllability (satisfying business rules, cross-modal alignment, and coverage of rare events) and statistical authenticity (marginal distributions, feature correlations, and indistinguishability from real data). Compressing both into a single end-to-end model often compromises both: either relying on expensive LLM reasoning for logic at high cost (e.g., Davidson et al. 2026) or relying on generative models for statistics while sacrificing controllability.

Goal: (1) Decouple "logical structure" from "statistical texture" at the framework level; (2) Use LLMs only for generating structural rules rather than row-by-row sampling to minimize costs; (3) Validate that controllability and downstream utility can be simultaneously preserved on weak multimodal (tabular + text sentiment) financial benchmarks.

Key Insight: LLMs excel at writing JSON alignment rules for data schemas rather than batch-generating thousands of rows. Thus, the LLM can serve as a rule provider rather than a data provider, delegating the heavy lifting of per-row generation to efficient tree-based models.

Core Idea: Reconcile logical skeletons \(\mathcal{S}\) (from LLM/human rules) and latent spaces \(z\) (from lightweight generators) within a conditional generator \(G(z\mid\mathcal{S})\). A TSTR + XModal feedback loop automatically determines whether to retrain the bottom-up model or adjust the top-down constraints.

Method

Overall Architecture

H-TDBU consists of a three-stage pipeline. Input: Real data \(\mathcal{D}_{\text{real}}\) and a schema description (human or LLM prompt). Output: \(\mathcal{D}_{\text{syn}}\) — a synthetic table satisfying business rules while statistically matching the real distribution (optionally with other modalities).

Mechanism:

1. Top-Down Path: Humans or LLMs produce a structural template \(\mathcal{S}\) after reading data summaries, formatted as a JSON describing "Factor \(f_i\) → Samplable Attributes." For example, specifying "target=1 must align with positive financial text."
2. Bottom-Up Path: An ensemble (RandomForest/XGBoost) or generative model (CTGAN/TVAE) fits the latent space \(\mathcal{Z}\) from \(\mathcal{D}_{\text{real}}\), capturing complex inter-column correlations.
3. Synthesis & Reconciliation: The skeleton \(\mathcal{S}\) and latent noise \(z\in\mathcal{Z}\) are fed into \(X_{\text{Syn}}:=G(z\in\mathcal{Z}\mid\mathcal{S})\), where every generated row is constrained by \(\mathcal{S}\). Post-generation is evaluated via TSTR (downstream utility) and XModal (alignment). Failures trigger either Retrain Model (for statistics) or Adjust Constraints (for logic).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    IN["Real Data + Schema Summary"]
    IN --> TD["LLM/Human Rule Provider<br/>One-shot Structural Template 𝒮"]
    IN --> BU["Lightweight Generator (Stats)<br/>RF/XGBoost/CTGAN → Latent Space z"]
    subgraph DECOUP["Conditional Decoupling"]
        direction TB
        BU -->|Statistical Texture z| SYN["Conditional Generator G(z∈𝒵∣𝒮)<br/>Constrained Sampling"]
    end
    TD -->|Logical Skeleton 𝒮| SYN
    SYN --> OUT["Synthetic Data D_syn"]
    OUT --> EVAL["TSTR + XModal Assessment"]
    EVAL -->|XModal Failure<br/>Retrain Model| BU
    EVAL -->|Rule Violation<br/>Adjust Constraints| TD
    EVAL -->|Pass| DONE["Validated Synthetic Table"]

Key Designs

1. Downgrading LLM from "Per-row Generator" to "One-shot Rule Provider"

Generating entire tables with LLMs is expensive and sensitive to inference quality. This work observes that LLMs are better at generating a single alignment JSON based on a schema. By using Gemini 3.1 Pro to output a rule-provider JSON (e.g., mapping target=1 to positive sentiments in FinancialPhraseBank) and delegating row generation to tree-based models, high LLM costs are reduced from \(O(N_{\text{rows}})\) to \(O(1)\).

2. Conditional Decoupling of Top-Down Skeleton and Bottom-Up Texture

To avoid the optimization difficulties of embedding rules directly into GAN/Diffusion losses, \(G(z\in\mathcal{Z}\mid\mathcal{S})\) separates the two. The bottom-up generator learns an unconstrained latent representation \(z\) (e.g., via XGBoost conditional sampling). During sampling, \(z\) is clipped/filtered by \(\mathcal{S}\) so that \(X_{\text{Syn}}\) preserves both learned correlations and alignment rules. This allows swapping rules without retraining models.

3. TSTR + XModal Feedback Loop: Distinguishing Learning Failure from Rule Failure

Conventional evaluation relies only on TSTR, failing to pinpoint why synthesis fails. This framework adds XModal (Total Variation Distance on joint distributions like "target × sentiment"). If XModal is high but TSTR is acceptable, the bottom-up model needs retraining. If TSTR is poor or violations are high, the rules \(\mathcal{S}\) are revised.

Loss & Training

The bottom-up phase uses native training schemes: RandomForest (30 trees, 5k samples), XGBoost (80 estimators, depth 6, 8k samples), and CTGAN/TVAE (SDV defaults, 50 epochs). Synthesis generates 12k rows across multiple seeds. The top-down phase requires no training, only rule generation.

Key Experimental Results

Main Results

Evaluation on weak multimodal financial benchmarks (Bank Marketing + FinancialPhraseBank):

Dataset	Method	TSTR Acc ↑	F1 ↑	AUROC ↑	XModal ↓
Manual	Independent	0.8830	0.0000	0.4905	0.1094
Manual	Gaussian copula	0.8949	0.2034	0.7738	0.0485
Manual	RandomForest	0.9281	0.6122	0.9188	0.0555
Manual	XGBoost	0.9139	0.5621	0.9190	0.1127
Manual	CTGAN	0.8992	0.2694	0.8358	0.0646
Manual	TVAE	0.8622	0.5581	0.8827	0.0533
Gemini	Independent	0.8830	0.0000	0.5359	0.3320
Gemini	Gaussian copula	0.9423	0.6749	0.9968	0.1605
Gemini	RandomForest	0.9971	0.9878	0.9998	0.1437
Gemini	XGBoost	0.9746	0.9003	0.9948	0.3234
Gemini	CTGAN	0.9574	0.7863	0.9903	0.1225
Gemini	TVAE	0.9881	0.9476	0.9885	0.0193

Insight: Stricter rules from Gemini improve TSTR for all generators (e.g., RF AUROC 0.9188 → 0.9998), as rules clarify decision boundaries, simplifying the bottom-up task.

Ablation Study

XGBoost ablation (training rows and conditioning columns):

Benchmark	Optimal Config	Acc ↑	F1 ↑	AUROC ↑	Note
Manual	12 cols	0.9139	0.5621	0.9190	Loose rules require more context
Gemini	4 cols	0.9925	0.9690	0.9999	Strict rules allow easier separation

Key Findings

• Stricter rules simplify generation: Gemini rules lock target=1 to positive sentiment, pushing all methods to TSTR \(\approx 0.99\).
• Utility \(\neq\) Fidelity: RandomForest achieved 0.9188 AUROC in the Manual setting but had higher XModal than Gaussian copula, justifying the dual-metric approach.
• Lightweight tree-based methods are competitive: RandomForest matched or outperformed CTGAN/TVAE at a fraction of the training time.
• Adaptive conditioning: Optimal column counts vary with rule complexity (Manual prefers 12, Gemini 4).

Highlights & Insights

• Efficiency: Reducing LLM usage to one-shot rule provision lowers costs by orders of magnitude compared to multi-stage reasoning frameworks while retaining semantic priors.
• Controlled Experimental Design: By fixing the bottom-up pipeline and varying only the JSON rules, the work quantifies the causal contribution of the "rule provider."
• Diagnostic Loops: Translating failures into structured "Retrain" or "Adjust" actions transforms ad-hoc tuning into a systematic process.

Limitations & Future Work

• The feedback loop description lacks an automated algorithmic protocol (thresholds and convergence criteria are not specified).
• Benchmarks are relatively weak, using coarse sentiment labels rather than complex financial reports or multi-table relationships.
• Controllability was tested on simple binary alignments; complex chain constraints (e.g., Age → Salary → Credit) were not assessed.
• Robustness of LLM rules across different prompts or models (variance) remains unexplored.
• Lack of head-to-head comparison with reasoning-driven synthesis works like Davidson et al. (2026).

Related Work & Insights

• vs. CTGAN/TVAE: These learn \(p(\text{table})\) directly; Ours splits it into \(p(\text{table}\mid\mathcal{S})\) and \(p(\mathcal{S})\). Ours is superior for controllability and low-data regimes.
• vs. GReaT/TabuLa: These utilize LLMs for row generation; Ours uses LLMs for metadata rules only, offering better logical enforcement at lower costs.
• vs. Model Collapse (Nature 2024): Top-down paths help mitigate collapse by forcing rare events into the data via rules rather than letting models erase them during recursive training.

Rating

• Novelty: ⭐⭐⭐⭐
• Experimental Thoroughness: ⭐⭐⭐
• Writing Quality: ⭐⭐⭐⭐
• Value: ⭐⭐⭐⭐⭐

Related Papers

• [CVPR 2026] Socratic-Geo: Synthetic Data Generation and Cross-Modal Geometric Reasoning via Multi-Agent Interaction
• [ICML 2026] Left-Right Symmetry Breaking in CLIP-style Vision-Language Models Trained on Synthetic Spatial-Relation Data
• [ACL 2025] Scaling Text-Rich Image Understanding via Code-Guided Synthetic Multimodal Data Generation
• [ICML 2026] Certified Robustness under Heterogeneous Perturbations via Hybrid Randomized Smoothing
• [CVPR 2026] Role-SynthCLIP: A Role-Play Driven Diverse Synthetic Data Approach