Large Language Models for Coagulant Dosage Prediction: A Systematic Investigation of Generalization, Rationale Patterns and Hallucination

Precise coagulant dosage control is essential for stable drinking-water treatment, yet conventional machine learning (ML) methods can be sensitive to data conditions. This study evaluates a large language model (LLM)-based in-context workflow for this tabular prediction task using the DeepSeek family, benchmarked against XGBoost, ANN, SVM, and k-NN under a shared chronological protocol. We examined performance across feature configurations, training-pool conditions, and outlier subsets. On the HQ-WTP case dataset, full-feature input outperformed temperature-only input, indicating the value of multivariate information. Performance responses to training-pool condition were model-dependent, with no universal optimum. Under the fixed protocol, in the full-feature test setting, the strongest tabular baseline showed the strongest test performance, while DeepSeek-Reasoner and DeepSeek-Chat showed intermediate performance, and DeepSeek-R1-Distill-Qwen-32B showed relatively lower stability. DeepSeek-Reasoner reached its best test performance at the 1/2 condition. We also coded narrative rationale patterns in generated responses and performed a protocol-based hallucination audit. DeepSeek-Reasoner showed comparatively lower hallucination incidence and more stable error behavior in this benchmark. These analyses are interpreted as response-level reliability evidence rather than verification of internal computational mechanisms. Overall, the study provides transparent, case-specific benchmark evidence for LLM-assisted decision support, while broader deployment claims require strict hallucination control, human-in-the-loop safeguards, and independent multi-site validation.