Sungmin Lee

Junghwan Lim, Wai Ting Cheung, Minsu Ha et al.

Training strong video generation models usually requires massive datasets, large parameter counts, and substantial compute. In this work, we ask whether strong text-to-video quality is possible at a much smaller budget: fewer than 10M clips and less than 100,000 H200 GPU hours. Our core claim is that part of the answer lies in how model capacity is organized, not only in how much of it is used. In video generation, prompt alignment, temporal consistency, and fine-detail recovery can interfere with one another when they are handled through the same pathway. Motif-Video 2B addresses this by separating these roles architecturally, rather than relying on scale alone. The model combines two key ideas. First, Shared Cross-Attention strengthens text control when video token sequences become long. Second, a three-part backbone separates early fusion, joint representation learning, and detail refinement. To make this design effective under a limited compute budget, we pair it with an efficient training recipe based on dynamic token routing and early-phase feature alignment to a frozen pretrained video encoder. Our analysis shows that later blocks develop clearer cross-frame attention structure than standard single-stream baselines. On VBench, Motif-Video~2B reaches 83.76\%, surpassing Wan2.1 14B while using 7$\times$ fewer parameters and substantially less training data. These results suggest that careful architectural specialization, combined with an efficiency-oriented training recipe, can narrow or exceed the quality gap typically associated with much larger video models.

Yujin Shin, Kichang Lee, Sungmin Lee et al.

While federated learning leverages distributed client resources, it faces challenges due to heterogeneous client capabilities. This necessitates allocating models suited to clients' resources and careful parameter aggregation to accommodate this heterogeneity. We propose HypeMeFed, a novel federated learning framework for supporting client heterogeneity by combining a multi-exit network architecture with hypernetwork-based model weight generation. This approach aligns the feature spaces of heterogeneous model layers and resolves per-layer information disparity during weight aggregation. To practically realize HypeMeFed, we also propose a low-rank factorization approach to minimize computation and memory overhead associated with hypernetworks. Our evaluations on a real-world heterogeneous device testbed indicate that \system enhances accuracy by 5.12% over FedAvg, reduces the hypernetwork memory requirements by 98.22%, and accelerates its operations by 1.86x compared to a naive hypernetwork approach. These results demonstrate HypeMeFed's effectiveness in leveraging and engaging heterogeneous clients for federated learning.

Junghwan Lim, Sungmin Lee, Dongseok Kim et al.

We introduce Motif-2-12.7B, a new open-weight foundation model that pushes the efficiency frontier of large language models by combining architectural innovation with system-level optimization. Designed for scalable language understanding and robust instruction generalization under constrained compute budgets, Motif-2-12.7B builds upon Motif-2.6B with the integration of Grouped Differential Attention (GDA), which improves representational efficiency by disentangling signal and noise-control attention pathways. The model is pre-trained on 5.5 trillion tokens spanning diverse linguistic, mathematical, scientific, and programming domains using a curriculum-driven data scheduler that gradually changes the data composition ratio. The training system leverages the MuonClip optimizer alongside custom high-performance kernels, including fused PolyNorm activations and the Parallel Muon algorithm, yielding significant throughput and memory efficiency gains in large-scale distributed environments. Post-training employs a three-stage supervised fine-tuning pipeline that successively enhances general instruction adherence, compositional understanding, and linguistic precision. Motif-2-12.7B demonstrates competitive performance across diverse benchmarks, showing that thoughtful architectural scaling and optimized training design can rival the capabilities of much larger models.

Sungmin Lee

Scientific publishing systematically filters out negative results. We argue that this long-standing asymmetry has become an urgent problem in the era of large language models, which inherit the positive bias of the literature they are trained on, face an impending shortage of high-quality training data, and are increasingly deployed as both research tools and peer reviewers. We analyze three ways in which LLMs have changed the value of failure data and show that the systematic absence of such data degrades their utility as research tools, training data consumers, and peer reviewers alike. We outline experimental protocols to validate these claims and discuss the structural conditions under which a failure-inclusive publishing culture could emerge.

Junghwan Lim, Sungmin Lee, Dongseok Kim et al.

The self-attention mechanism, while foundational to modern Transformer architectures, suffers from a critical inefficiency: it frequently allocates substantial attention to redundant or noisy context. Differential Attention addressed this by using subtractive attention maps for signal and noise, but its required balanced head allocation imposes rigid constraints on representational flexibility and scalability. To overcome this, we propose Grouped Differential Attention (GDA), a novel approach that introduces unbalanced head allocation between signal-preserving and noise-control groups. GDA significantly enhances signal focus by strategically assigning more heads to signal extraction and fewer to noise-control, stabilizing the latter through controlled repetition (akin to GQA). This design achieves stronger signal fidelity with minimal computational overhead. We further extend this principle to group-differentiated growth, a scalable strategy that selectively replicates only the signal-focused heads, thereby ensuring efficient capacity expansion. Through large-scale pretraining and continual training experiments, we demonstrate that moderate imbalance ratios in GDA yield substantial improvements in generalization and stability compared to symmetric baselines. Our results collectively establish that ratio-aware head allocation and selective expansion offer an effective and practical path toward designing scalable, computation-efficient Transformer architectures.

Sungmin Lee, Minju Kang, Joonhee Lee et al.

Question-answering (QA) interfaces powered by large language models (LLMs) present a promising direction for improving interactivity with HVAC system insights, particularly for non-expert users. However, enabling accurate, real-time, and context-aware interactions with HVAC systems introduces unique challenges, including the integration of frequently updated sensor data, domain-specific knowledge grounding, and coherent multi-stage reasoning. In this paper, we present JARVIS, a two-stage LLM-based QA framework tailored for sensor data-driven HVAC system interaction. JARVIS employs an Expert-LLM to translate high-level user queries into structured execution instructions, and an Agent that performs SQL-based data retrieval, statistical processing, and final response generation. To address HVAC-specific challenges, JARVIS integrates (1) an adaptive context injection strategy for efficient HVAC and deployment-specific information integration, (2) a parameterized SQL builder and executor to improve data access reliability, and (3) a bottom-up planning scheme to ensure consistency across multi-stage response generation. We evaluate JARVIS using real-world data collected from a commercial HVAC system and a ground truth QA dataset curated by HVAC experts to demonstrate its effectiveness in delivering accurate and interpretable responses across diverse queries. Results show that JARVIS consistently outperforms baseline and ablation variants in both automated and user-centered assessments, achieving high response quality and accuracy.

Junghwan Lim, Sungmin Lee, Dongseok Kim et al.

We introduce Motif-2-12.7B-Reasoning, a 12.7B parameter language model designed to bridge the gap between open-weight systems and proprietary frontier models in complex reasoning and long-context understanding. Addressing the common challenges of model collapse and training instability in reasoning adaptation, we propose a comprehensive, reproducible training recipe spanning system, data, and algorithmic optimizations. Our approach combines memory-efficient infrastructure for 64K-token contexts using hybrid parallelism and kernel-level optimizations with a two-stage Supervised Fine-Tuning (SFT) curriculum that mitigates distribution mismatch through verified, aligned synthetic data. Furthermore, we detail a robust Reinforcement Learning Fine-Tuning (RLFT) pipeline that stabilizes training via difficulty-aware data filtering and mixed-policy trajectory reuse. Empirical results demonstrate that Motif-2-12.7B-Reasoning achieves performance comparable to models with significantly larger parameter counts across mathematics, coding, and agentic benchmarks, offering the community a competitive open model and a practical blueprint for scaling reasoning capabilities under realistic compute constraints.

Yesung Cho, Sungmin Lee, Geongyu Lee et al.

Recent studies in pathology foundation models have shown that scaling training data, diversifying cancer types, and increasing model size consistently improve their performance. However, giga-scale foundation models, which are trained on hundreds of thousands of slides covering tens of cancer types and contain billions of parameters, pose significant challenges for practical use due to their tremendous computational costs in both development and deployment. In this work, we present a novel strategy, named the G2L framework, to increase the performance of large-scale foundation models, which consist of only $15\%$ of the parameters of giga-scale models, to a comparable performance level of giga-scale models in cancer-specific tasks. Our approach applies knowledge distillation, transferring the capabilities of a giga-scale model to a large-scale model, using just 1K pathology slides of a target cancer (e.g., breast, prostate, etc.). The resulting distilled model not only outperformed state-of-the-art models of the same size (i.e., large-scale) across several benchmarks but also, interestingly, surpassed the giga-scale teacher and huge-scale models in some benchmarks. In addition, the distilled model exhibited a higher robustness index, indicating improved resilience to image variations originating from multiple institutions. These findings suggest that the proposed distillation approach for a large-scale model is a data- and parameter-efficient way to achieve giga-scale-level performance for cancer-specific applications without prohibitive computational burden.

Junghwan Lim, Gangwon Jo, Sungmin Lee et al.

We introduce Llama-3-Motif, a language model consisting of 102 billion parameters, specifically designed to enhance Korean capabilities while retaining strong performance in English. Developed on the Llama 3 architecture, Llama-3-Motif employs advanced training techniques, including LlamaPro and Masked Structure Growth, to effectively scale the model without altering its core Transformer architecture. Using the MoAI platform for efficient training across hyperscale GPU clusters, we optimized Llama-3-Motif using a carefully curated dataset that maintains a balanced ratio of Korean and English data. Llama-3-Motif shows decent performance on Korean-specific benchmarks, outperforming existing models and achieving results comparable to GPT-4.

Junghwan Lim, Sungmin Lee, Dongseok Kim et al.

Recent advancements in Large Language Models (LLMs) have revolutionized artificial intelligence, yet developing an effective foundational LLM that balances high performance with computational efficiency remains challenging, especially for emerging research groups. To address this gap, we introduce Motif-2.6B, a 2.6-billion-parameter foundation model designed to democratize advanced LLM capabilities. Motif-2.6B incorporates several innovative architectural enhancements, including Differential Attention and PolyNorm activation functions, which improve long-context comprehension, reduce hallucination, and enhance in-context learning capabilities. We rigorously tested multiple novel architectural components through extensive experimentation to determine the optimal architecture for Motif-2.6B. Comprehensive evaluations demonstrate that Motif-2.6B consistently meets or exceeds the performance of similarly sized state-of-the-art models across diverse benchmarks, showcasing its effectiveness, scalability, and real-world applicability. Through detailed experiments and tailored techniques, Motif-2.6B significantly advances the landscape of efficient, scalable, and powerful foundational LLMs, offering valuable insights and a robust foundation for future research and deployment.

Sara Akbarzadeh, Sungmin Lee, Chin-Tuan Tan

In multi speakers environments, cochlear implant (CI) users may attend to a target sound source in a different manner from the normal hearing (NH) individuals during a conversation. This study attempted to investigate the effect of conversational sound levels on the mechanisms adopted by CI and NH listeners in selective auditory attention and how it affects their daily conversation. Nine CI users (five bilateral, three unilateral, and one bimodal) and eight NH listeners participated in this study. The behavioral speech recognition scores were collected using a matrix sentences test and neural tracking to speech envelope was recorded using electroencephalography (EEG). Speech stimuli were presented at three different levels (75, 65, and 55 dB SPL) in the presence of two maskers from three spatially separated speakers. Different combinations of assisted/impaired hearing modes were evaluated for CI users and the outcomes were analyzed in three categories: electric hearing only, acoustic hearing only, and electric+acoustic hearing. Our results showed that increasing the conversational sound level degraded the selective auditory attention in electrical hearing. On the other hand, increasing the sound level improved the selective auditory attention for the acoustic hearing group. In NH listeners, however, increasing the sound level did not cause a significant change in the auditory attention. Our result implies that the effect of the sound level on the selective auditory attention varies depending on hearing modes and the loudness control is necessary for the ease of attending to the conversation by CI users.

Sara Akbarzadeh, Sungmin Lee, Fei Chen et al.

Electrical hearing by cochlear implants (CIs) may be fundamentally different from acoustic hearing by normal-hearing (NH) listeners, presumably showing unequal speech quality perception in various noise environments. Noise reduction (NR) algorithms used in CI reduce the noise in favor of signal-to-noise ratio (SNR), regardless of plausible accompanying distortions that may degrade the speech quality perception. To gain better understanding of CI speech quality perception, the present work aimed investigating speech quality perception in a diverse noise conditions, including factors of speech/noise levels, type of noise, and distortions caused by NR models. Fifteen NH and seven CI subjects participated in this study. Speech sentences were set to two different levels (65 and 75 dB SPL). Two types of noise (Cafeteria and Babble) at three levels (55, 65, and 75 dB SPL) were used. Sentences were processed using two NR algorithms to investigate the perceptual sensitivity of CI and NH listeners to the distortion. All sentences processed with the combinations of these sets were presented to CI and NH listeners, and they were asked to rate the sound quality of speech as they perceived. The effect of each factor on the perceived speech quality was investigated based on the group averaged quality rated by CI and NH listeners. Consistent with previous studies, CI listeners were not as sensitive as NH to the distortion made by NR algorithms. Statistical analysis showed that the speech level has significant effect on quality perception. At the same SNR, the quality of 65 dB speech was rated higher than that of 75 dB for CI users, but vice versa for NH listeners. Therefore, the present study showed that the perceived speech quality patterns were different between CI and NH listeners in terms of their sensitivity to distortion and speech level in complex listening environment.

Naeem Ul Islam, Sungmin Lee, Jaebyung Park

Modifying the facial images with desired attributes is important, though challenging tasks in computer vision, where it aims to modify single or multiple attributes of the face image. Some of the existing methods are either based on attribute independent approaches where the modification is done in the latent representation or attribute dependent approaches. The attribute independent methods are limited in performance as they require the desired paired data for changing the desired attributes. Secondly, the attribute independent constraint may result in the loss of information and, hence, fail in generating the required attributes in the face image. In contrast, the attribute dependent approaches are effective as these approaches are capable of modifying the required features along with preserving the information in the given image. However, attribute dependent approaches are sensitive and require a careful model design in generating high-quality results. To address this problem, we propose an attribute dependent face modification approach. The proposed approach is based on two generators and two discriminators that utilize the binary as well as the real representation of the attributes and, in return, generate high-quality attribute modification results. Experiments on the CelebA dataset show that our method effectively performs the multiple attribute editing with preserving other facial details intactly.

Jungbeom Lee, Eunji Kim, Sungmin Lee et al.

When a deep neural network is trained on data with only image-level labeling, the regions activated in each image tend to identify only a small region of the target object. We propose a method of using videos automatically harvested from the web to identify a larger region of the target object by using temporal information, which is not present in the static image. The temporal variations in a video allow different regions of the target object to be activated. We obtain an activated region in each frame of a video, and then aggregate the regions from successive frames into a single image, using a warping technique based on optical flow. The resulting localization maps cover more of the target object, and can then be used as proxy ground-truth to train a segmentation network. This simple approach outperforms existing methods under the same level of supervision, and even approaches relying on extra annotations. Based on VGG-16 and ResNet 101 backbones, our method achieves the mIoU of 65.0 and 67.4, respectively, on PASCAL VOC 2012 test images, which represents a new state-of-the-art.

Jungbeom Lee, Eunji Kim, Sungmin Lee et al.

The main obstacle to weakly supervised semantic image segmentation is the difficulty of obtaining pixel-level information from coarse image-level annotations. Most methods based on image-level annotations use localization maps obtained from the classifier, but these only focus on the small discriminative parts of objects and do not capture precise boundaries. FickleNet explores diverse combinations of locations on feature maps created by generic deep neural networks. It selects hidden units randomly and then uses them to obtain activation scores for image classification. FickleNet implicitly learns the coherence of each location in the feature maps, resulting in a localization map which identifies both discriminative and other parts of objects. The ensemble effects are obtained from a single network by selecting random hidden unit pairs, which means that a variety of localization maps are generated from a single image. Our approach does not require any additional training steps and only adds a simple layer to a standard convolutional neural network; nevertheless it outperforms recent comparable techniques on the Pascal VOC 2012 benchmark in both weakly and semi-supervised settings.

Sungmin Lee, Jangho Lee, Jungbeom Lee et al.

A meningioma is a type of brain tumor that requires tumor volume size follow ups in order to reach appropriate clinical decisions. A fully automated tool for meningioma detection is necessary for reliable and consistent tumor surveillance. There have been various studies concerning automated lesion detection. Studies on the application of convolutional neural network (CNN)-based methods, which have achieved a state-of-the-art level of performance in various computer vision tasks, have been carried out. However, the applicable diseases are limited, owing to a lack of strongly annotated data being present in medical image analysis. In order to resolve the above issue we propose pyramid gradient-based class activation mapping (PG-CAM) which is a novel method for tumor localization that can be trained in weakly supervised manner. PG-CAM uses a densely connected encoder-decoder-based feature pyramid network (DC-FPN) as a backbone structure, and extracts a multi-scale Grad-CAM that captures hierarchical features of a tumor. We tested our model using meningioma brain magnetic resonance (MR) data collected from the collaborating hospital. In our experiments, PG-CAM outperformed Grad-CAM by delivering a 23 percent higher localization accuracy for the validation set.

Jungbeom Lee, Jangho Lee, Sungmin Lee et al.

Video prediction has been considered a difficult problem because the video contains not only high-dimensional spatial information but also complex temporal information. Video prediction can be performed by finding features in recent frames, and using them to generate approximations to upcoming frames. We approach this problem by disentangling spatial and temporal features in videos. We introduce a mutual suppression network (MSnet) which are trained in an adversarial manner and then produces spatial features which are free of motion information, and motion features with no spatial information. MSnet then uses motion-guided connection within an encoder-decoder-based architecture to transform spatial features from a previous frame to the time of an upcoming frame. We show how MSnet can be used for video prediction using disentangled representations. We also carry out experiments to assess the effectiveness of our method to disentangle features. MSnet obtains better results than other recent video prediction methods even though it has simpler encoders.