Selecting a layer normalization (LN) strategy that stabilizes training and speeds convergence in Transformers remains difficult, even for today’s large language models (LLM). We present a comprehensive analytical foundation for understanding how different LN strategies influence training dynamics in large-scale Transformers. Until recently, Pre-LN and Post-LN have long dominated practices despite their limitations in large-scale training. However, several open-source models have recently begun silently adopting a third strategy without much explanation. This strategy places normalization layer peripherally around sublayers, a design we term Peri-LN. While Peri-LN has demonstrated promising performance, its precise mechanisms and benefits remain almost unexplored. Our in-depth analysis delineates the distinct behaviors of LN strategies, showing how each placement shapes activation variance and gradient propagation. To validate our theoretical insight, we conduct extensive experiments on Transformers up to $3.2$B parameters, showing that Peri-LN consistently achieves more balanced variance growth, steadier gradient flow, and convergence stability. Our results suggest that Peri-LN warrants broader consideration for large-scale Transformer architectures, providing renewed insights into the optimal placement of LN.

Guijin Son, Hanwool Lee, Sungdong Kim, Seungone Kim, Niklas Muennighoff, Taekyoon Choi, Cheonbok Park, Kang Min Yoo, Stella Biderman. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Hyuhng Joon Kim, Youna Kim, Cheonbok Park, Junyeob Kim, Choonghyun Park, Kang Min Yoo, Sang-goo Lee, Taeuk Kim. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

With deep learning (DL) outperforming conventional methods for different tasks, much effort has been devoted to utilizing DL in various domains. Researchers and developers in the traffic domain have also designed and improved DL models for forecasting tasks such as estimation of traffic speed and time of arrival. However, there exist many challenges in analyzing DL models due to the black-box property of DL models and complexity of traffic data (i.e., spatio-temporal dependencies). Collaborating with domain experts, we design a visual analytics system, AttnAnalyzer, that enables users to explore how DL models make predictions by allowing effective spatio-temporal dependency analysis. The system incorporates dynamic time warping (DTW) and Granger causality tests for computational spatio-temporal dependency analysis while providing map, table, line chart, and pixel views to assist user to perform dependency and model behavior analysis. For the evaluation, we present three case studies showing how AttnAnalyzer can effectively explore model behaviors and improve model performance in two different road networks. We also provide domain expert feedback.

Predicting traffic conditions is tremendously challenging since every road is highly dependent on each other, both spatially and temporally. Recently, to capture this spatial and temporal dependency, specially designed architectures such as graph convolutional networks and temporal convolutional networks have been introduced. While there has been remarkable progress in traffic forecasting, we found that deep-learning-based traffic forecasting models still fail in certain patterns, mainly in event situations (e.g., rapid speed drops). Although it is commonly accepted that these failures are due to unpredictable noise, we found that these failures can be corrected by considering previous failures. Specifically, we observe autocorrelated errors in these failures, which indicates that some predictable information remains. In this study, to capture the correlation of errors, we introduce ResCAL, a residual estimation module for traffic forecasting, as a widely applicable add-on module to existing traffic forecasting models. Our ResCAL calibrates the prediction of the existing models in real time by estimating future errors using previous errors and graph signals. Extensive experiments on METR-LA and PEMS-BAY demonstrate that our ResCAL can correctly capture the correlation of errors and correct the failures of various traffic forecasting models in event situations.

Multi-domain Neural Machine Translation (NMT) trains a single model with multiple domains.It is appealing because of its efficacy in handling multiple domains within one model.An ideal multi-domain NMT should learn distinctive domain characteristics simultaneously, however, grasping the domain peculiarity is a non-trivial task.In this paper, we investigate domain-specific information through the lens of mutual information (MI) and propose a new objective that penalizes low MI to become higher.Our method achieved the state-of-theart performance among the current competitive multi-domain NMT models.Also, we empirically show our objective promotes low MI to be higher resulting in domain-specialized multidomain NMT.

Cheonbok Park, Yunwon Tae, TaeHee Kim, Soyoung Yang, Mohammad Azam Khan, Lucy Park, Jaegul Choo. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

To tackle ever-increasing city traffic congestion problems, researchers have proposed deep learning models to aid decision-makers in the traffic control domain. Although the proposed models have been remarkably improved in recent years, there are still questions that need to be answered before deploying models. For example, it is difficult to figure out which models provide state-of-the-art performance, as recently proposed models have often been evaluated with different datasets and experiment environments. It is also difficult to determine which models would work when traffic conditions change abruptly (e.g., rush hour). In this work, we conduct two experiments to answer the two questions. In the first experiment, we conduct an experiment with the state-of-the-art models and the identical public datasets to compare model performance under a consistent experiment environment. We then extract a set of temporal regions in the datasets, whose speeds change abruptly and use these regions to explore model performance with difficult intervals. The experiment results indicate that Graph-WaveNet and GMAN show better performance in general. We also find that prediction models tend to have varying performances with data and intervals, which calls for in-depth analysis of models on difficult intervals for real-world deployment.

Predicting road traffic speed is a challenging task due to different types of roads, abrupt speed change and spatial dependencies between roads; it requires the modeling of dynamically changing spatial dependencies among roads and temporal patterns over long input sequences. This paper proposes a novel spatio-temporal graph attention (ST-GRAT) that effectively captures the spatio-temporal dynamics in road networks. The novel aspects of our approach mainly include spatial attention, temporal attention, and spatial sentinel vectors. The spatial attention takes the graph structure information (e.g., distance between roads) and dynamically adjusts spatial correlation based on road states. The temporal attention is responsible for capturing traffic speed changes, and the sentinel vectors allow the model to retrieve new features from spatially correlated nodes or preserve existing features. The experimental results show that ST-GRAT outperforms existing models, especially in difficult conditions where traffic speeds rapidly change (e.g., rush hours). We additionally provide a qualitative study to analyze when and where ST-GRAT tended to make accurate predictions during rush-hour times.

Document labeling is a critical step in building various machine learning applications. However, the step can be time-consuming and arduous, requiring a significant amount of human efforts. To support an efficient document labeling environment, we present a system called Attentive Interactive Labeling Assistant (AILA). In its core, AILA uses Interactive Attention Module (IAM), a novel module that visually highlights words in a document that labelers may pay attention to when labeling a document. IAM utilizes attention-based Deep Neural Networks which not only support a prediction of which words to highlight but also enable labelers to indicate words that should be assigned a high attention weight while labeling to improve the future quality of word prediction.We evaluated the labeling efficiency and the accuracy by comparing the conditions with and without IAM in our study. The results showed that participants' labeling efficiency increased significantly under the condition with IAM than the condition without IAM, while the two conditions maintained roughly the same labeling accuracy.