Jaeyeon Lee

Jaeyeon Lee, Guantong Qi, Matthew Brady Neeley et al.

Recent advancements in large language models (LLMs) integrating explicit reasoning, such as OpenAI's o3-mini, DeepSeek-R1, and QWQ-32B, enable smaller models to solve complex tasks by generating intermediate reasoning steps prior to providing answers. However, this approach significantly increases computational costs, both monetarily and environmentally. The widely-used self-consistency method further exacerbates these costs by aggregating multiple reasoning paths to improve accuracy, often requiring between 40 to 64 samples per task. Although aggregation effectively reduces variance and bias, additional sampling can lead to diminishing returns when early samples yield consistent results. To address inefficiencies, we propose leveraging Sequential Probability Ratio Testing (SPRT) to dynamically terminate sampling once sufficient consistency is achieved. We calibrate SPRT parameters specifically for LLM applications, accounting for sensitivity to detect the mode of the distribution. Our experiments demonstrate that incorporating SPRT significantly enhances token efficiency, achieving comparable accuracy to self-consistency methods but at a substantially reduced computational cost. To promote transparency and facilitate reproducibility, we have made the source code and datasets used in our experiments publicly available at our GitHub repository: https://github.com/LiuzLab/consol, or available as a PyPI package: pip install consol. We hope that this resource will support further research and encourage the development of new methods building upon our work.

Youngwoo Yoon, Bok Cha, Joo-Haeng Lee et al.

For human-like agents, including virtual avatars and social robots, making proper gestures while speaking is crucial in human--agent interaction. Co-speech gestures enhance interaction experiences and make the agents look alive. However, it is difficult to generate human-like gestures due to the lack of understanding of how people gesture. Data-driven approaches attempt to learn gesticulation skills from human demonstrations, but the ambiguous and individual nature of gestures hinders learning. In this paper, we present an automatic gesture generation model that uses the multimodal context of speech text, audio, and speaker identity to reliably generate gestures. By incorporating a multimodal context and an adversarial training scheme, the proposed model outputs gestures that are human-like and that match with speech content and rhythm. We also introduce a new quantitative evaluation metric for gesture generation models. Experiments with the introduced metric and subjective human evaluation showed that the proposed gesture generation model is better than existing end-to-end generation models. We further confirm that our model is able to work with synthesized audio in a scenario where contexts are constrained, and show that different gesture styles can be generated for the same speech by specifying different speaker identities in the style embedding space that is learned from videos of various speakers. All the code and data is available at https://github.com/ai4r/Gesture-Generation-from-Trimodal-Context.

Woo-Ri Ko, Minsu Jang, Jaeyeon Lee et al.

To better interact with users, a social robot should understand the users' behavior, infer the intention, and respond appropriately. Machine learning is one way of implementing robot intelligence. It provides the ability to automatically learn and improve from experience instead of explicitly telling the robot what to do. Social skills can also be learned through watching human-human interaction videos. However, human-human interaction datasets are relatively scarce to learn interactions that occur in various situations. Moreover, we aim to use service robots in the elderly-care domain; however, there has been no interaction dataset collected for this domain. For this reason, we introduce a human-human interaction dataset for teaching non-verbal social behaviors to robots. It is the only interaction dataset that elderly people have participated in as performers. We recruited 100 elderly people and two college students to perform 10 interactions in an indoor environment. The entire dataset has 5,000 interaction samples, each of which contains depth maps, body indexes and 3D skeletal data that are captured with three Microsoft Kinect v2 cameras. In addition, we provide the joint angles of a humanoid NAO robot which are converted from the human behavior that robots need to learn. The dataset and useful python scripts are available for download at https://github.com/ai4r/AIR-Act2Act. It can be used to not only teach social skills to robots but also benchmark action recognition algorithms.

Jaeyeon Lee, Hyun-Hwan Jeong, Zhandong Liu

Diagnosing rare diseases requires linking gene findings with often unstructured reference text. Current pipelines collect many candidate genes, but clinicians still spend a lot of time filtering false positives and combining evidence from papers and databases. A key challenge is language: phenotype descriptions and inheritance patterns are written in prose, not fully captured by tables. Large language models (LLMs) can read such text, but clinical use needs grounding in citable knowledge and stable, repeatable behavior. We explore a knowledge-grounded and language-aware reranking layer on top of a high-recall first-stage pipeline. The goal is to improve precision and explainability, not to replace standard bioinformatics steps. We use expert-built context and a consensus method to reduce LLM variability, producing shorter, better-justified gene lists for expert review. LA-MARRVEL achieves the highest accuracy, outperforming other methods -- including traditional bioinformatics diagnostic tools (AI-MARRVEL, Exomiser, LIRICAL) and naive large language models (e.g., Anthropic Claude) -- with an average Recall@5 of 94.10%, a +3.65 percentage-point improvement over AI-MARRVEL. The LLM-generated reasoning provides clear prose on phenotype matching and inheritance patterns, making clinical review faster and easier. LA-MARRVEL has three parts: expert-engineered context that enriches phenotype and disease information; a ranked voting algorithm that combines multiple LLM runs to choose a consensus ranked gene list; and the AI-MARRVEL pipeline that provides first-stage ranks and gene annotations, already known as a state-of-the-art method in Rare Disease Diagnosis on BG, DDD, and UDN cohorts. The online AI-MARRVEL includes LA-MARRVEL as an LLM feature at https://ai.marrvel.org . We evaluate LA-MARRVEL on three datasets from independent cohorts of real-world diagnosed patients.

Jaeyeon Lee, Dong-Wan Choi

Although Vision Transformers (ViTs) have become the standard architecture in computer vision, their massive sizes lead to significant computational overhead. Token compression techniques have attracted considerable attention to address this issue, but they often suffer from severe information loss, requiring extensive additional training to achieve practical performance. In this paper, we propose Adaptive Token Merging (ATM), a novel method that ensures lossless token merging, eliminating the need for fine-tuning while maintaining competitive performance. ATM adaptively reduces tokens across layers and batches by carefully adjusting layer-specific similarity thresholds, thereby preventing the undesirable merging of less similar tokens with respect to each layer. Furthermore, ATM introduces a novel token matching technique that considers not only similarity but also merging sizes, particularly for the final layers, to minimize the information loss incurred from each merging operation. We empirically validate our method across a wide range of pretrained models, demonstrating that ATM not only outperforms all existing training-free methods but also surpasses most training-intensive approaches, even without additional training. Remarkably, training-free ATM achieves over a 30% reduction in FLOPs for the DeiT-T and DeiT-S models without any drop in their original accuracy.

Hyemin Lim, Jaeyeon Lee, Dong-Wan Choi

Large pretrained language models such as BERT suffer from slow inference and high memory usage, due to their huge size. Recent approaches to compressing BERT rely on iterative pruning and knowledge distillation, which, however, are often too complicated and computationally intensive. This paper proposes a novel semi-structured one-shot pruning method for BERT, called $\textit{Permutation and Grouping for BERT}$ (PGB), which achieves high compression efficiency and sparsity while preserving accuracy. To this end, PGB identifies important groups of individual weights by permutation and prunes all other weights as a structure in both multi-head attention and feed-forward layers. Furthermore, if no important group is formed in a particular layer, PGB drops the entire layer to produce an even more compact model. Our experimental results on BERT$_{\text{BASE}}$ demonstrate that PGB outperforms the state-of-the-art structured pruning methods in terms of computational cost and accuracy preservation.

Minsu Jang, Sangwon Seo, Dohyung Kim et al.

This paper introduces a large-scale Korean speech dataset, called VOTE400, that can be used for analyzing and recognizing voices of the elderly people. The dataset includes about 300 hours of continuous dialog speech and 100 hours of read speech, both recorded by the elderly people aged 65 years or over. A preliminary experiment showed that speech recognition system trained with VOTE400 can outperform conventional systems in speech recognition of elderly people's voice. This work is a multi-organizational effort led by ETRI and MINDs Lab Inc. for the purpose of advancing the speech recognition performance of the elderly-care robots.

Jinhyeok Jang, Dohyung Kim, Cheonshu Park et al.

Deep learning, based on which many modern algorithms operate, is well known to be data-hungry. In particular, the datasets appropriate for the intended application are difficult to obtain. To cope with this situation, we introduce a new dataset called ETRI-Activity3D, focusing on the daily activities of the elderly in robot-view. The major characteristics of the new dataset are as follows: 1) practical action categories that are selected from the close observation of the daily lives of the elderly; 2) realistic data collection, which reflects the robot's working environment and service situations; and 3) a large-scale dataset that overcomes the limitations of the current 3D activity analysis benchmark datasets. The proposed dataset contains 112,620 samples including RGB videos, depth maps, and skeleton sequences. During the data acquisition, 100 subjects were asked to perform 55 daily activities. Additionally, we propose a novel network called four-stream adaptive CNN (FSA-CNN). The proposed FSA-CNN has three main properties: robustness to spatio-temporal variations, input-adaptive activation function, and extension of the conventional two-stream approach. In the experiment section, we confirmed the superiority of the proposed FSA-CNN using NTU RGB+D and ETRI-Activity3D. Further, the domain difference between both groups of age was verified experimentally. Finally, the extension of FSA-CNN to deal with the multimodal data was investigated.

Woo-han Yun, Taewoo Kim, Jaeyeon Lee et al.

Training an object instance detector where only a few training object images are available is a challenging task. One solution is a cut-and-paste method that generates a training dataset by cutting object areas out of training images and pasting them onto other background images. A detector trained on a dataset generated with a cut-and-paste method suffers from the conventional domain shift problem, which stems from a discrepancy between the source domain (generated training dataset) and the target domain (real test dataset). Though state-of-the-art domain adaptation methods are able to reduce this gap, it is limited because they do not consider the difference of domain gaps of foreground and background. In this study, we present that the conventional domain gap can be divided into two sub-domain gaps for foreground and background. Then, we show that the original cut-and-paste approach suffers from a new domain gap problem, an unbalanced domain gaps, because it has two separate source domains for foreground and background, unlike the conventional domain shift problem. Then, we introduce an advanced cut-and-paste method to balance the unbalanced domain gaps by diversifying the foreground with GAN (generative adversarial network)-generated seed images and simplifying the background using image processing techniques. Experimental results show that our method is effective for balancing domain gaps and improving the accuracy of object instance detection in a cluttered indoor environment using only a few seed images. Furthermore, we show that balancing domain gaps can improve the detection accuracy of state-of-the-art domain adaptation methods.

Jinhyeok Jang, Jaehong Kim, Jaeyeon Lee et al.

This work presents an adaptive activation method for neural networks that exploits the interdependency of features. Each pixel, node, and layer is assigned with a polynomial activation function, whose coefficients are provided by an auxiliary activation network. The activation of a feature depends on the features of neighboring pixels in a convolutional layer and other nodes in a dense layer. The dependency is learned from data by the activation networks. In our experiments, networks with activation networks provide significant performance improvement compared to the baseline networks on which they are built. The proposed method can be used to improve the network performance as an alternative to increasing the number of nodes and layers.

Youngwoo Yoon, Woo-Ri Ko, Minsu Jang et al.

Co-speech gestures enhance interaction experiences between humans as well as between humans and robots. Existing robots use rule-based speech-gesture association, but this requires human labor and prior knowledge of experts to be implemented. We present a learning-based co-speech gesture generation that is learned from 52 h of TED talks. The proposed end-to-end neural network model consists of an encoder for speech text understanding and a decoder to generate a sequence of gestures. The model successfully produces various gestures including iconic, metaphoric, deictic, and beat gestures. In a subjective evaluation, participants reported that the gestures were human-like and matched the speech content. We also demonstrate a co-speech gesture with a NAO robot working in real time.

Jinhyeok Jang, Hyunjoong Cho, Jaehong Kim et al.

This work presents deep asymmetric networks with a set of node-wise variant activation functions. The nodes' sensitivities are affected by activation function selections such that the nodes with smaller indices become increasingly more sensitive. As a result, features learned by the nodes are sorted by the node indices in the order of their importance. Asymmetric networks not only learn input features but also the importance of those features. Nodes of lesser importance in asymmetric networks can be pruned to reduce the complexity of the networks, and the pruned networks can be retrained without incurring performance losses. We validate the feature-sorting property using both shallow and deep asymmetric networks as well as deep asymmetric networks transferred from famous networks.