TAGAL: Agentic Tabular Data Generation

• TAGAL is a training-free framework for synthetic tabular data generation that uses iterative self-critique and external knowledge injection to refine data quality.
• The method employs an agentic workflow with a generation agent and a critic agent that iteratively improve synthetic data via feedback loops and prompt refinement.
• Empirical evaluations show TAGAL achieves competitive downstream classifier performance and data similarity compared to state-of-the-art models without further LLM training.

Searching arXiv for TAGAL and closely related tabular synthetic data generation work. TAGAL, short for Tabular data Generation using Agentic LLM Methods, is a collection of methods for generating synthetic tabular data through an agentic workflow in which LLMs generate rows, critique their own outputs, and iteratively refine subsequent generations without any further gradient-based training of the LLMs themselves. The framework emphasizes automatic feedback, prompt-injected external knowledge, and training-free operation, and it is evaluated through both downstream classifier utility and similarity between real and synthetic data. The reported results show that TAGAL is able to perform on par with state-of-the-art approaches that require LLM training and generally outperforms other training-free approaches across diverse datasets (Ronval et al., 4 Sep 2025).

1. Concept and scope

TAGAL is designed for synthetic tabular data generation in settings where one wants to improve the performance of machine learning tasks, among which is the training of models for classification. Its central premise is that an LLM can be used not merely as a one-pass generator, but as part of an automatic and iterative process that uses feedback to improve the generated data. The framework also uses prompt-time external knowledge rather than fine-tuning, which allows the addition of domain information without modifying model weights (Ronval et al., 4 Sep 2025).

The framework is explicitly organized around three core components. The Prompt Engine (Generation Agent) is an LLM instance prompted to produce new rows in CSV style given a compact schema description plus a handful of real examples. The Feedback Loop (Critic Agent) is a second LLM instance that reads the generated batch, performs a step-by-step critique, and issues recommendations for improvement. The External Knowledge Module injects domain knowledge at prompt time through summary statistics or value-lists for each feature and user-supplied instructions such as class-balance or feature constraints (Ronval et al., 4 Sep 2025).

A recurrent point in the formulation is that TAGAL is training-free in the sense used by the paper: it does not require further LLM training. This distinguishes it from fine-tuned LLM approaches such as GReaT and Tabula, and from tabular generative models such as CTGAN and TabDDPM, while still aiming at competitive downstream performance (Ronval et al., 4 Sep 2025).

2. Agentic architecture and iterative generation

The operational structure of TAGAL is cyclical. In its simplest form, one run consists of Initial Prompt Construction, Data Generation, Automatic Analysis (Feedback), Prompt Refinement and Re-Generation, and Iteration. The initial system prompt enumerates each feature name, type, and either, for categorical variables, the full list of levels with their empirical frequencies or, for numerical variables, their mean, median, and standard deviation. It also instructs the LLM to “think step by step,” follow observed patterns in the few-shot examples, and produce rows in CSV format grouped by target class. The user prompt includes 20 real “few-shot” examples per class and a request to generate new examples (Ronval et al., 4 Sep 2025).

After generation, a second LLM analyzes the synthetic batch. The feedback system prompt again restates schema information and asks the critic to list strengths and weaknesses of the batch, including distribution gaps, duplicates, and unrealistic values, and then to conclude with precise instructions for improvement. Those recommendations are appended to the next user prompt sent to the generation agent, which produces a new batch under the guidance of the critique. The paper gives a standard run length of 3 iterations for a complete SynthLoop run (Ronval et al., 4 Sep 2025).

This architecture treats critique as an explicit computational resource rather than as an implicit latent capability of a single prompt. The authors attribute TAGAL’s performance to iterative self-critique, external knowledge prompts, and model-agnosticism, arguing that these mechanisms let the LLM correct distributional blind spots that pure few-shot prompting cannot catch (Ronval et al., 4 Sep 2025).

3. Prompt design, operational modes, and external knowledge

TAGAL includes three named operating modes: SynthLoop, ReducedLoop, and Prompt-Refine.

Mode	Procedure	Distinguishing property
SynthLoop	Repeats generation and feedback for a fixed number of iterations	Resets with fresh few-shots after a run
ReducedLoop	Continues generation after one run with the final conversation history	Reuses accumulated feedback
Prompt-Refine	Summarizes a run into a concise refined prompt, then reuses it	Uses a summary LLM to reduce later prompting cost

In SynthLoop, after the fixed number of iterations, one may reset both conversation histories and repeat with a fresh random sample of few-shots. In ReducedLoop, one keeps the fully built-up conversation history and simply asks the generation agent for more examples until the quota is met. In Prompt-Refine, after one SynthLoop or ReducedLoop run, the entire generation conversation is passed to a summary LLM which produces a single concise refined prompt encapsulating learned recommendations plus a small set of few-shots; that refined prompt is then repeatedly submitted to the generation model with newly sampled few-shots (Ronval et al., 4 Sep 2025).

External knowledge enters the system through two specific mechanisms. First, the prompt provides schema statistics: for each feature, either a full list of categorical levels plus their frequencies or numerical summary statistics $\{\mu,\mathrm{median},\sigma\}$. Second, the user may append user-supplied guidelines, for example “ensure at least 15% of rows have feature X in value v” or “never exceed value Z for feature w.” The paper characterizes this as a prompt-level “straightjacket” that steers the LLM toward realistic marginal and conditional distributions without any back-propagation or retraining (Ronval et al., 4 Sep 2025).

A concrete illustration is given for the Thyroid dataset, which was released after most LLMs’ training cut-offs. TAGAL still attains high utility there because prompt-injected statistics provide missing domain knowledge. This suggests that TAGAL is not confined to domains already well represented in LLM pre-training, provided that the prompt can supply the relevant tabular constraints (Ronval et al., 4 Sep 2025).

4. Evaluation formalism and quality criteria

The paper evaluates two broad families of metrics: utility for downstream classifiers and data similarity. For downstream utility, TAGAL trains three off-the-shelf classifiers—Random Forest, LightGBM, and XGBoost—under two regimes: TSTR (Train on Synthetic, Test on Real) and Combined (50% Real + 50% Synthetic train, Test on Real). Performance is reported as ROC-AUC averaged over 5 runs (Ronval et al., 4 Sep 2025).

For data similarity, the paper uses precision and recall on manifolds following Naeem et al. (2020). Let real rows be $\{X_i\}_{i=1}^N$ and synthetic rows be $\{Y_j\}_{j=1}^M$, and let $NND_k(X)$ denote the Euclidean distance from $X$ to its $k$-th nearest neighbor among real samples. The real manifold is defined as

$$\mathcal M_{\rm real}=\bigcup_i B(X_i,\,NND_k(X_i)),$$

with an analogous definition for $\mathcal M_{\rm synth}$. The reported similarity metrics are

$$\mathrm{Precision} \;=\;\frac1M\sum_{j=1}^M\mathbf 1\bigl(Y_j\in \mathcal M_{\rm real}\bigr),\quad \mathrm{Recall} \;=\;\frac1N\sum_{i=1}^N\mathbf 1\bigl(X_i\in \mathcal M_{\rm synth}\bigr).$$

The study also reports Collision Rate for exact overlap between synthetic and real rows (Ronval et al., 4 Sep 2025).

The benchmark datasets are all binary classification problems:

Dataset	Size	Schema size
Adult	35,561 rows	15 features
Bank	45,211 rows	17 features
German	1,000 rows	10 features
Thyroid	383 rows	17 features

The baselines cover three families: a statistical sampler based on independent feature sampling; CTGAN and TabDDPM as GAN- and diffusion-based generative models; GReaT and Tabula as fine-tuned LLM approaches; and EPIC as a pure few-shot LLM baseline with no feedback (Ronval et al., 4 Sep 2025).

5. Empirical performance across datasets

On the Adult dataset, the best training-based model reported in the comparison is TabDDPM, which achieves approximately TSTR≈0.87, Combined≈0.89, Precision≈0.92, and Recall≈0.90. Against this reference point, TAGAL Prompt-Refine using Llama 3.1 8B achieves TSTR=0.85, Combined=0.88, Precision=0.94, Recall=0.86, and collisions=3.2%. The paper also reports that this substantially outperforms EPIC, which reaches TSTR=0.53 and collisions=56.5% (Ronval et al., 4 Sep 2025).

On Bank, Prompt-Refine yields TSTR=0.59, Combined=0.87, Precision=0.99, Recall=0.83, and collisions=19.1%, which the paper describes as on par with or above other no-train approaches. On the smaller datasets, especially German and Thyroid, TAGAL variants still reach Combined $\ge 0.97$–$\{X_i\}_{i=1}^N$0 AUC, TSTR up to 0.99, and manifold Precision/Recall often above 0.90, typically matching TabDDPM and outperforming EPIC or pure statistical sampling (Ronval et al., 4 Sep 2025).

The model-size analysis is also explicit. When the LLM size grows from Llama 3.1 8B to larger systems such as GPT-4o or DeepSeek-v3, TAGAL’s TSTR and precision/recall move even closer to the performance of fully trained generative models. The paper interprets this as evidence for two properties: model-agnosticism, since any LLM API may be plugged in, and the practical sufficiency of an open-source 8B-parameter LLM for strong synthetic tabular data generation (Ronval et al., 4 Sep 2025).

6. Limitations, misconceptions, and future directions

A common misconception would be to treat TAGAL as a form of LLM fine-tuning. The paper states the opposite: the workflow improves data generation without any further LLM training, relying instead on in-context prompting, critique, and external knowledge injection. Another plausible misconception is that a single-pass few-shot prompt is sufficient; the comparison with EPIC is presented precisely to show that iterative self-critique materially changes the quality regime (Ronval et al., 4 Sep 2025).

The authors note three principal limitations. Collision risk remains present, especially in ReducedLoop, because repeating few-shot examples can still lead to duplicates. Prompt brittleness means quality depends on how well system and user instructions are crafted, and Prompt-Refine can sometimes omit crucial details. Cost/time trade-off is unavoidable because full feedback loops require more LLM calls, and therefore more latency or API expense, than single-shot prompting (Ronval et al., 4 Sep 2025).

The proposed future directions extend the same agentic logic rather than abandoning it. The paper suggests exploring reinforcement-learning-enabled agentic workflows, investigating conditional generation for targeted oversampling in imbalanced settings, benchmarking on multi-class and highly high-dimensional tabular schemas, and integrating statistical distance measurements such as Wasserstein or KL directly into the critic’s checklist (Ronval et al., 4 Sep 2025).

Within the landscape of synthetic data generation, TAGAL therefore occupies a specific methodological niche: it treats prompt engineering, iterative critique, and explicit schema statistics as a substitute for model retraining. The reported findings support the narrower claim that an agentic critique–refine loop plus prompt-injected domain knowledge can make training-free LLM-based tabular synthesis competitive with approaches that require learned generative models (Ronval et al., 4 Sep 2025).