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.

Controllable text generation is the primary technique for controlling specific attributes such as topic, keywords and obtaining augmented data. This work proposes a novel controllable text generation framework to improve the controllability of generation models. First, we introduce semantic control grammar, a controllable input format to generate sentences that satisfy the constraints. Second, we adopt a generation and rerank method to obtain semantically reranked controlled sentences. Extensive experiments and analyses are conducted on benchmark, natural language understanding, data-to-text generation, and text classification datasets. Through automatic evaluations, we show that our method leads to improvement over strong baselines. The results show that our model can control sentence and word attributes and semantically generate natural sentences. Furthermore, our techniques improve the generation quality of the model.

In recent years, methods that selectively fine-tune or reduce the number of layers in large language models (LLMs) have garnered attention as an efficient alternative to traditional fine-tuning, where all layers are trained. In this study, we revisit the classical concept of layer freezing and propose a simple, effective strategy that selectively fine-tunes only a portion of transformer layers. We show that freezing the bottom 25% or 50% of layers in small-scale LLMs with sub-3 billion parameters yields significant improvements in memory efficiency and training speed while maintaining, or even surpassing, the performance of full fine-tuning and Low-Rank Adaptation (LoRA). Through experiments on Natural Language Inference (NLI) tasks using LLMs with fewer than 3 billion parameters, our approach achieves up to 50% memory savings and 30% faster training. Notably, our method does not require architectural modifications or additional parameters, making it particularly suitable for resource-constrained environments.

Natural language understanding (NLU) is a core technique for implementing natural user interfaces. In this study, we propose a neural network architecture to learn syntax vector representation by employing the correspondence between texts and syntactic structures. For representing the syntactic structures of sentences, we used three methods: dependency trees, phrase structure trees, and part of speech tagging. A pretrained transformer is used to propose text-to-vector and syntax-to-vector projection approaches. The texts and syntactic structures are projected onto a common vector space, and the distances between the two vectors are minimized according to the correspondence property to learn the syntax representation. We conducted massive experiments to verify the effectiveness of the proposed methodology using Korean corpora, i.e., Weather, Navi, and Rest, and English corpora, i.e., the ATIS, SNIPS, Simulated Dialogue-Movie, Simulated Dialogue-Restaurant, and NLU-Evaluation datasets. Through the experiments, we concluded that our model is quite effective in capturing a syntactic representation and the learned syntax vector representations are useful.

Recent deep learning models, including GPT-4, have achieved remarkable performance using the back-propagation (BP) algorithm. However, the mechanism of BP is fundamentally different from how the human brain processes learning. To address this discrepancy, the Forward-Forward (FF) algorithm was introduced. Although FF enables deep learning without backward passes, it suffers from instability, dependence on artificial input construction, and limited generalizability. To overcome these challenges, we propose Local Back-Propagation (LBP), a method that integrates layer-wise unsupervised learning with standard inputs and conventional loss functions. Specifically, LBP demonstrates high training stability and competitive accuracy, significantly outperforming FF-based training methods. Moreover, LBP reduces memory usage by up to 48% compared to convolutional neural networks trained with back-propagation, making it particularly suitable for resource-constrained environments such as federated learning. These results suggest that LBP is a promising biologically inspired training method for decentralized deep learning.

Recent deep learning models, such as GPT-4, use the back-propagation algorithm (BP) and have achieved impressive performance. However, there is a noticeable difference between how BP operates and how the human brain learns. In response to this, the Forward-Forward algorithm (FF) was introduced. FF trains deep learning models using only forward passes. Although FF cannot fully replace BP due to its need for specialized inputs and loss functions, it remains promising in situations where BP is difficult to use, such as federated learning. To address these limitations and demonstrate the practical value of FF, we propose a Local Back-Propagation method that incorporates unsupervised FF. By using an unsupervised learning model, our approach allows training with standard inputs and common loss functions, thereby avoiding the special requirements of FF. This not only leads to more stable learning but also enables a wider range of possible applications than FF alone. Furthermore, because our method allows each layer to be physically separated, we have tested its effectiveness in scenarios like federated learning, where individual models are trained separately and then combined. Our results confirm that this approach expands the usability and scope of FF-based training methods.