The recent advancements in large language models (LLMs) have brought significant changes to the field of education, particularly in the generation and evaluation of feedback. LLMs are transforming education by streamlining tasks like content creation, feedback generation, and assessment, reducing teachers’ workload and improving online education efficiency. This study aimed to verify the consistency and reliability of LLMs as evaluators by conducting automated evaluations using various LLMs based on five educational evaluation criteria. The analysis revealed that while LLMs were capable of performing consistent evaluations under certain conditions, a lack of consistency was observed both among evaluators and across models for other criteria. Notably, low agreement among human evaluators correlated with reduced reliability in LLM evaluations. Furthermore, variations in evaluation results were influenced by factors such as prompt strategies and model architecture, highlighting the complexity of achieving reliable assessments using LLMs. These findings suggest that while LLMs have the potential to transform educational systems, careful selection and combination of models are essential to improve their consistency and align their performance with human evaluators in educational settings.

Intrinsic clearance is a key pharmacokinetic parameter in drug development because it influences systemic exposure and dose selection. Conventional in vitro-in vivo extrapolation (IVIVE) approaches are useful but often require additional experimental resources and may be difficult to apply in the earliest stages of compound screening. This study aimed to develop a biologically informed computational framework for intrinsic clearance prediction by integrating drug-target interaction (DTI)-derived features from 23 hepatic metabolic proteins with physicochemical properties, using an IVIVE-based endpoint harmonization approach to unify heterogeneous endpoint labels. A multi-source human clearance dataset was assembled and harmonized into a common intrinsic clearance endpoint. A pretrained DTI model based on ChemBERTa and ProtBERT was used to generate precomputed interaction features for CYP, UGT, and SULT family proteins. These features were combined with compound structure and physicochemical descriptors in downstream clearance prediction models using either a multilayer perceptron (MLP) or a transformer encoder. Internal evaluation included repeated random 5-fold cross-validation and scaffold-split analysis, and an independent external evaluation set of 185 compounds was additionally examined. Feature settings incorporating DTI-derived information showed improved regression performance compared with the descriptor-only baseline across most evaluated metrics. Under the augmented setting, the MLP model with DTI + LogP/Fup showed the highest overall regression performance among the evaluated models (r²_m = 0.2505 ± 0.0792, r² = 0.2364 ± 0.0879), while the highest CI was observed in the transformer encoder model with DTI + LogP/Fup (0.6915 ± 0.0144). On the external evaluation set, the MLP model showed an AFE of 1.116, with 52.970% and 70.792% of predictions falling within 2-fold and 3-fold error, respectively. Comparisons with published IVIVE-related studies should be interpreted as contextual rather than direct head-to-head validation. The proposed framework supports intrinsic clearance estimation by combining biologically informed DTI-derived features with physicochemical information in an interpretable prediction setting. Although the predictive performance remains modest, the results suggest potential utility as a supportive early-stage prioritization tool when standardized experimental clearance data are limited. Further work is needed to expand the dataset, strengthen external validation with independent cohorts, and incorporate additional elimination pathways such as renal clearance and transporter-mediated disposition.

Mathematical problem transformation is a teaching-and-learning strategy that extends conceptual understanding and problem-solving ability by expressing the same concept across diverse situations. It has recently attracted attention in artificial intelligence as a tool for data augmentation, difficulty control, and model evaluation. However, existing approaches struggle to jointly represent and control how core mathematical elements—such as operational structure, quantitative relations, and conditions—are preserved or modified. This limitation is particularly evident in natural-language problems, where intertwined components make it difficult to perform targeted partial transformations or verify structural validity. To address these challenges, we propose the Symbolic Math Problem Representation Framework (SyMPRep), which represents the relationships among sentences, conditions, questions, quantities, units, and operations in a symbolic structure. It classifies free-form instructions into predefined categories and decomposes problems into constituent elements, enabling transformation over an explicit abstraction structure. This allows problem transformation to be treated as a controllable, traceable, and recoverable structural operation rather than surface rewriting. Experiments on GSM8K and Math500 show that SyMPRep achieves stable alignment and recoverability, and confirm that the main challenge lies in structural control rather than surface fluency. Ablation results highlight the importance of symbolic schema and show that different metrics capture distinct aspects of transformation quality. In downstream applications, answer-invariant transformations yield modest improvements on easier problems, while human evaluation indicates that the generated problems are coherent and suitable for educational use. These findings suggest that SyMPRep serves as a representation-driven interface for controllable structural transformation.