LLMs on Tabular Data with Limited Semantics: Evidence from Industrial Car Retrofit Prediction
About
Industrial retrofit planning depends on structured operational data rather than free text: planners must estimate whether a newly registered prototype will require a retrofit, which retrofit package it will need, and how long the work will take. We study an industrial dataset linking a prototype-registration system (284,271 vehicles) with a retrofit-management system (48,716 cleaned visits), and compare strong tabular machine learning baselines with three LLM-based strategies on row-serialized inputs: embedding features (Amazon Titan), direct prompted classification (Claude Sonnet 4), and an ML+LLM stacking approach. Across binary occurrence prediction, 15-way retrofit-type classification, per-visit duration regression, and an aggregated monthly benchmark, classical tree ensembles remain the strongest standalone models. However, the LLM results reveal a consistent pattern: embeddings remain useful on tables (binary AUC = 0.982), direct prompting collapses once semantic signal is stripped by hashing (binary AUC = 0.500; multiclass weighted F1 = 0.018), and hybrid stacking yields the best manually built multiclass model (weighted F1 = 0.626). On the monthly benchmark, lag-based machine learning outperforms time-series foundation models, though Chronos-small remains competitive in zero-shot forecasting. The results suggest that on privacy-constrained industrial tables, LLMs are more effective as complementary components than as replacements for strong tabular baselines.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Binary Classification | Stage 1 binary (left join) | ROC-AUC99.6 | 11 | |
| Multiclass Classification | Retrofit system Stage 2 multiclass inner join | Accuracy68 | 9 | |
| Multiclass Classification | retrofit-type prediction Stage 2 | Accuracy68 | 6 | |
| Binary Classification | Vehicle retrofit occurrence | ROC-AUC0.982 | 4 | |
| Temporal forecasting | Enterprise Tables Stage 3 (Monthly) | -- | 4 | |
| Binary Classification | Vehicle retrofit occurrence 200-row | ROC-AUC0.996 | 2 | |
| Duration Prediction | Enterprise Tables Stage 3 (Per-visit) | -- | 2 |