Large Language Models for Missing Data Imputation: Understanding Behavior, Hallucination Effects, and Control Mechanisms
Authors: Arthur Dantas Mangussi, Ricardo Cardoso Pereira, Ana Carolina Lorena, Pedro Henriques Abreu
Organizations: Computer Science Division, Aeronautics Institute of Technology, Praça Marechal Eduardo Gomes, 50, São José dos Campos, 12228-900, São Paulo, Brazil · Science and Technology Institute, Federal University of São Paulo, Avenue Cesare Monsueto Giulio Lattes, 1201, São José dos Campos, 12231-280, São Paulo, Brazil · University of Coimbra, CISUC/LASI – Centre for Informatics and Systems of the University of Coimbra, Department of Informatics Engineering, Polo II, Pinhal de Marrocos, Coimbra 3030-290, Portugal
Data imputation is a cornerstone technique for handling missing values in real-world datasets, which are often plagued by missingness. Despite recent progress, prior studies on Large Language Models-based imputation remain limited by scalability challenges, restricted cross-model comparisons, and evaluations conducted on small or domain-specific datasets. Furthermore, heterogeneous experimental protocols and inconsistent treatment of missingness mechanisms (MCAR, MAR, and MNAR) hinder systematic benchmarking across methods. This work investigates the robustness of Large Language Models for missing data imputation in tabular datasets using a zero-shot prompt engineering approach. To this end, we present a comprehensive benchmarking study comparing five widely used LLMs against six state-of-the-art imputation baselines. The experimental design evaluates these methods across 29 datasets (including nine synthetic datasets) under MCAR, MAR, and MNAR mechanisms, with missing rates of up to 20%. The results demonstrate that leading LLMs, particularly Gemini 3.0 Flash and Claude 4.5 Sonnet, consistently achieve superior performance on real-world open-source datasets compared to traditional methods. However, this advantage appears to be closely tied to the models' prior exposure to domain-specific patterns learned during pre-training on internet-scale corpora. In contrast, on synthetic datasets, traditional methods such as MICE outperform LLMs, suggesting that LLM effectiveness is driven by semantic context rather than purely statistical reconstruction. Furthermore, we identify a clear trade-off: while LLMs excel in imputation quality, they incur significantly higher computational time and monetary costs. Overall, this study provides a large-scale comparative analysis, positioning LLMs as promising semantics-driven imputers for complex tabular data.
Figures & tables
Figure 1: Overview of methodology applied in this work.
Dataset Name
Samples
Features
Feature Type
synthetic-cat
500
4
Categorical
synthetic-cont
500
4
Continuous
synthetic-cat-cont
500
4
Categorical/Continuous
synthetic-one
381
4
Continuous
synthetic-two
131
8
Continuous
synthetic-three
342
11
Continuous
Table 1: Overview of synthetic dataset characteristics.
Dataset
Instances
Features
Subject Area
Continuous
Categorical
pima-diabetes
768
8
1
Healthcare
heart-cleveland
303
6
8
Healthcare
wiscosin
569
31
1
Healthcare
parkinsons
195
22
1
Healthcare
hepatitis
80
6
14
Healthcare
Table 2: Overview of the datasets characteristics selected by the Gemini model.
Model Identifier
API String
Input
Output
Provider
Access
$/Mtk
$/Mtk
MiMo-V2-Flash
xiaomi/mimo-v2-flash
0.09
0.29
Xiaomi
OpenRouter
Devstral 2 2512
mistralai/devstral-2512
0.40
2.00
Mistral
OpenRouter
Gemini 3.0 Flash
gemini-3-flash
0.00*
0.00*
Google
AI Studio
GPT-4.1-Nano
openai/gpt-4.1-nano
0.10
0.40
OpenAI
OpenRouter
Claude 4.5 Sonnet
claude-sonnet-4.5
3.00
6.00
Anthropic
Anthropic
Table 3: Comparative analysis of LLMs and cost structures. Pricing is denoted in USD per million tokens (Mtk). Asterisks (*) denote tier-dependent pricing or free-tier availability.
Imputation method
Parameters
K NN
k = 5 (neighbors), Euclidean distance metric
MICE
100 iterations, default parameters from Scikit-learn
SAEI
epochs = 200, batch size = 64, optimizer = “Adam"
missForest
criterion = “absolute_error", n_estimators=10,
default parameters from Scikit-learn
TabPFN
default parameters
Table 4: Parameters of each imputation method.
Figure 2: Illustration of the complete prompt structure used to perform data imputation via prompt engineering.
Figure 3: Density plots of the Normalized Root Mean Square Error results obtained for each imputation strategy. The analysis is presented by missing data mechanisms. The average NRMSE for each setting is also displayed.
MD Mechanisms
MNAR
MCAR
MAR
Missing Rate
5%
10%
20%
5%
10%
20%
5%
10%
20%
SoftImpute
0.654
0.644
0.649
0.273
0.294
0.320
0.311
0.325
0.351
kNN
0.485
0.496
0.509
0.203
0.228
0.256
0.236
0.249
0.284
missForest
0.418
0.440
0.453
0.192
0.218
0.242
0.233
0.242
0.283
MICE
0.426
0.439
0.475
0.174
0.212
0.292
0.211
0.227
0.298
SAEI
0.518
0.482
0.418
0.295
0.313
0.320
0.330
0.333
0.335
Table 5: Overall Normalized Root Mean Square Error (NRMSE) divided by missing rate on real-world datasets under MNAR, MCAR, and MAR mechanisms for all large language models. Bold values indicate the best-performing method, while underlined values denote the second best.
Dataset
5%
10%
20%
MNAR
MCAR
MAR
MNAR
MCAR
MAR
MNAR
MCAR
MAR
pima
✓
✓
-
✓
✓
✓
✓
-
✓
cleveland
✓
✓
✓
✓
✓
✓
✓
✓
✓
wiscosin
-
-
-
-
-
-
-
-
-
parkinsons
✓
✓
✓
✓
✓
✓
✓
✓
✓
hepatitis
-
-
-
-
-
-
-
-
-
Table 6: Scenarios where at least one Large Language Model achieves superior performance compared to traditional data imputation methods.
LLMs
Fallback Usage (%)
MiMo-V2-Flash
0%
Devstral 2 2512
1.85%
Gemini 3.0 Flash
0%
Claude 4.5 Sonnet
0%
GPT-4.1-Nano
14.94%
Table 7: Comparative analysis of the percentage of cases in which LLMs relied on the fallback mechanism for missing data imputation. The results are based on synthetic datasets, where LLM-based methods exhibited poorer performance.
Figure 4: NRMSE vs. computational time across imputation methods. Each point corresponds to a dataset. The x-axis reports imputation accuracy (NRMSE), while the y-axis shows runtime in seconds (log scale). The figure highlights the trade-off between predictive performance and computational cost across LLM-based and traditional imputation approaches with Pareto frontier.
Dataset
NBF
pima
4
cleveland
4
wiscosin
9
parkinsons
3
hepatitis
2
mathernal-risk
6
Table 8: Number of batches per fold in real-world datasets used for LLMs-based to perform data imputation task.
Imputation Methods
MD Mechanism
MAR
MNAR
MCAR
missForest
0 .224 ± 0.122
0 .425 ± 0.096
0 .188 ± 0.140
MICE
0.160 ± 0.146
0.429 ± 0.115
0.169 ± 0.137
Gemini 3.0 Flash
0.239 ± 0.144
0.422 ± 0.153
0.208 ± 0.134
Claude 4.5 Sonnet
0.257 ± 0.134
0.433 ± 0.152
0.225 ± 0.143
Table 9: Overall Normalized Root Mean Square Error (NRMSE) grouped by missing rate on synthetic datasets under MNAR, MCAR, and MAR mechanisms for Gemini 3.0 Flash, Claude 4.5 Sonnet, MICE, and missForest. Bold values indicate the best-performing method, while underlined values denote the second best.
Missing data imputation remains a fundamental challenge in modern data science, especially when uncertainty quantification is essential. In this work, we propose MissBGM, an AI-powered missing data imputation method via Bayesian generative modeling that bridges the expressive flexibility of neural networks with the statistical rigor of Bayesian inference. Unlike existing methods that often focus on point estimates or treat the missingness mechanism implicitly, MissBGM explicitly and jointly models the data-generating and missingness mechanisms, providing principled posterior uncertainty over imputations rather than a single point estimate. We develop a stochastic optimization framework with alternating updates among missing values, model parameters, and latent variables until convergence. Our theoretical analysis shows that estimates of missing values from MissBGM converge consistently under mild assumptions. Empirically, we demonstrate that MissBGM achieves superior performance over traditional imputers and recent neural network-based methods across extensive experimental settings. These results establish MissBGM as a principled and scalable solution for modern missing data imputation. The code for MissBGM is open sourced at https://github.com/liuq-lab/MissBGM.
Large language models have been widely evaluated as simulators of individual survey responses. In practice, however, fully unobserved responses are rare; the dominant problem is partial non-response. Imputation aims to restore the overall structure of a survey dataset by filling in these missing values. It has its own well-defined evaluation criteria and differs fundamentally from prediction. We propose to impute missing survey data through in-context learning (ICL). We systematically evaluate ICL design choices across different missingness mechanisms (MCAR, MAR, MNAR) on 150 opinion variables spanning 15 waves of the American Trends Panel. Compared to well-established statistical methods for data imputation like MICE PMM, our ICL approach consistently reduces absolute error across all missingness mechanisms, with the largest gains under non-random missingness (MNAR). Notably, the best-performing specification (gpt-oss-120b with 100 in-context examples) achieves near-nominal aggregate coverage (approaching the 95% level) with confidence intervals two to five times narrower than MICE PMM. We publish a Python package with an sklearn-like API to enable easy deployment of our method using local and proprietary LLMs.
Tobias Holtdirk, Georg Ahnert, Joseph W Sakshaug +1
LMU Munich · Munich Center for Machine Learning · University of Mannheim +2
In real-world machine learning applications, incomplete observations create a fundamental challenge. Researchers have come up with several ideas to address this crucial problem. However, current models still face challenges in balancing scalability and structural consistency. This study proposes a feature imputation method, called FILLER, that deliberately searches the two-dimensional latent space produced by a generative model and fills the missing values with appropriate entries. The generative model is trained on fully observed data to generate samples from the latent space, and FILLER uses this trained model to impute the values missing in the corrupted test samples. In this study, G-NeuroDAVIS serves the purpose of the generative model. This work also presents a mathematical proof on the convergence of the iterative search. Finally, FILLER has been evaluated on several image datasets under random and structured missingness patterns with varying levels of imputation complexities. In order to justify the efficacy of FILLER, it has been compared against existing state-of-the-art solution strategies in terms of RMSE, PSNR, and SSIM. In addition, Wilcoxon signed-rank test has been carried out to validate statistical significance. Moreover, downstream analyses (classification and clustering) have also established the quality of imputation in terms of standard metrics.
Santu Mondal, Chayan Maitra, Rajat K. De
Machine Intelligence Unit Indian Statistical Institute 203 Barrackpore Trunk Road Kolkata 700108