Tae-Ho Kim

Hancheol Park, Geonho Lee, Tairen Piao et al.

Mixture-of-Experts (MoE) models scale foundation models efficiently by activating only a subset of experts for each token, but their large number of expert parameters still makes quantization essential for practical deployment. Unlike dense models, however, MoE models are sensitive to routing instability: small quantization-induced perturbations can change the top-$k$ expert selection, altering the computation path and degrading model quality. We propose Value-and-Structure Routing Alignment for Quantization (VSRAQ), a MoE-specific post-training quantization objective that preserves pre-quantization expert-selection behavior under quantization. VSRAQ combines two complementary objectives that jointly preserve expert-selection behavior: value alignment, which matches routing-relevant logits or scores, and structure alignment, which preserves expert ordering and top-$k$ decision boundaries. By maintaining routing consistency, VSRAQ reduces quantization-induced degradation without introducing any inference-time overhead and can be integrated into existing quantization frameworks. Experiments on recent MoE foundation models show that VSRAQ improves expert-selection consistency and consistently outperforms reconstruction-only and router-aware baselines.

Fachrina Dewi Puspitasari, Chaoning Zhang, Jiaquan Zhang et al.

The demand for powerful instruction following and reasoning capability of large language models (LLMs) has promoted rapid development of retrieval-augmented generation (RAG). The RAG system assists LLM generation by retrieving chunks of query-fit supplementary knowledge from an external database. Conventional RAG systems, however, suffer from information insufficiency due to two factors, which are intent-agnostic retrieval and information fragmentation. Our work proposes a RAG framework, termed InSemRAG, that addresses these challenges via an iterative retrieve-and-check mechanism with two supporting modules, an intention-aware retriever (IAR) and semantics-preserving chunking (SPC). IAR implements a dynamic hybrid retrieval method that adaptively weights the retrieval channels based on the query intent, while SPC performs detection and reparation to the damaged evidence chunks to preserve the semantic integrity. To alleviate the computational latency brought by our iterative mechanism, we leverage small language models (SLMs). Extensive experiments across several benchmark datasets consistently demonstrate the competitiveness of our method against recent state-of-the-art RAG mechanisms. Particularly, our method achieves significant gains on multi-hop and evidence-sensitive tasks, with a 2.65-point improvement in F1 on HotPotQA and a 1.5-point increase in accuracy on FEVER. Our method also achieves competitive performance to Multi-Hop RAG with 4.32$\times$ lower latency with the utilization of SLM.

Pengcheng Zheng, Chaoning Zhang, Ya Wen et al.

Large Language Models (LLMs) have demonstrated strong capabilities in natural language understanding and reasoning, while recent extensions that incorporate visual inputs enable them to process multimodal information. Despite these advances, Large Vision-Language Models (LVLMs) incur substantial computational and memory costs, hindering deployment in resource-constrained scenarios. Existing layer pruning methods typically rely on local similarity metrics or static proxy signals, failing to capture the global and dynamic evolution of representations across model depth, which often leads to the removal of transition-critical layers. To address this limitation, we propose a topology-aware layer pruning framework for LVLMs. Specifically, we represent layer wise hidden states as point clouds and models their evolution using \textit{simplicial complexes}. By leveraging \textit{zigzag persistent homology}, we quantify inter-layer topological consistency and enable adaptive pruning that preserves critical representational transitions. Extensive experiments on diverse multimodal benchmarks demonstrate that the proposed framework consistently outperforms existing pruning methods across a wide range of sparsity ratios. Our code is available at https://github.com/zpc456/TopoVLM.

Seongmin Park, Beomseok Kwon, Jieun Lim et al.

Uniform-precision neural network quantization has gained popularity since it simplifies densely packed arithmetic unit for high computing capability. However, it ignores heterogeneous sensitivity to the impact of quantization errors across the layers, resulting in sub-optimal inference accuracy. This work proposes a novel neural architecture search called neural channel expansion that adjusts the network structure to alleviate accuracy degradation from ultra-low uniform-precision quantization. The proposed method selectively expands channels for the quantization sensitive layers while satisfying hardware constraints (e.g., FLOPs, PARAMs). Based on in-depth analysis and experiments, we demonstrate that the proposed method can adapt several popular networks channels to achieve superior 2-bit quantization accuracy on CIFAR10 and ImageNet. In particular, we achieve the best-to-date Top-1/Top-5 accuracy for 2-bit ResNet50 with smaller FLOPs and the parameter size.

Bo-Kyeong Kim, Geonmin Kim, Tae-Ho Kim et al.

Structured pruning of modern large language models (LLMs) has emerged as a way of decreasing their high computational needs. Width pruning reduces the size of projection weight matrices (e.g., by removing attention heads) while maintaining the number of layers. Depth pruning, in contrast, removes entire layers or blocks, while keeping the size of the remaining weights unchanged. Most current research focuses on either width-only or a blend of width and depth pruning, with little comparative analysis between the two units (width vs. depth) concerning their impact on LLM inference efficiency. In this work, we show that simple depth pruning can effectively compress LLMs while achieving comparable or superior performance to recent width pruning studies. Our pruning method boosts inference speeds, especially under memory-constrained conditions that require limited batch sizes for running LLMs, where width pruning is ineffective. In retraining pruned models for quality recovery, continued pretraining on a large corpus markedly outperforms LoRA-based tuning, particularly at severe pruning ratios. We hope this work can help build compact yet capable LLMs. Code and models can be found at: https://github.com/Nota-NetsPresso/shortened-llm

Chaoning Zhang, Dongshen Han, Sheng Zheng et al.

Segment anything model (SAM) addresses two practical yet challenging segmentation tasks: \textbf{segment anything (SegAny)}, which utilizes a certain point to predict the mask for a single object of interest, and \textbf{segment everything (SegEvery)}, which predicts the masks for all objects on the image. What makes SegAny slow for SAM is its heavyweight image encoder, which has been addressed by MobileSAM via decoupled knowledge distillation. The efficiency bottleneck of SegEvery with SAM, however, lies in its mask decoder because it needs to first generate numerous masks with redundant grid-search prompts and then perform filtering to obtain the final valid masks. We propose to improve its efficiency by directly generating the final masks with only valid prompts, which can be obtained through object discovery. Our proposed approach not only helps reduce the total time on the mask decoder by at least 16 times but also achieves superior performance. Specifically, our approach yields an average performance boost of 3.6\% (42.5\% \textit{v.s.} 38.9\%) for zero-shot object proposal on the LVIS dataset with the mask AR@$K$ metric. Qualitative results show that our approach generates fine-grained masks while avoiding over-segmenting things. This project targeting faster SegEvery than the original SAM is termed MobileSAMv2 to differentiate from MobileSAM which targets faster SegAny. Moreover, we demonstrate that our new prompt sampling is also compatible with the distilled image encoders in MobileSAM, contributing to a unified framework for efficient SegAny and SegEvery. The code is available at the same link as MobileSAM Project \href{https://github.com/ChaoningZhang/MobileSAM}{\textcolor{red}{https://github.com/ChaoningZhang/MobileSAM}}. \end{abstract}

Xudong Wang, Chaoning Zhang, Chenghao Li et al.

Large Language Models (LLMs) have demonstrated strong capabilities in complex reasoning tasks, while recent prompting strategies such as Chain-of-Thought (CoT) have further elevated their performance in handling complex logical problems. Despite these advances, high-quality reasoning remains heavily reliant on manual static prompts and is sensitive to decoding configurations and task distributions, leading to performance fluctuations and limited transferability. Existing automatic prompt optimization methods typically adopt single-agent local search, failing to simultaneously optimize prompts and decoding hyperparameters within a unified framework to achieve stable global improvements. To address this limitation, we propose Agent-GWO, a dynamic prompt optimization framework for complex reasoning. Specifically, we unify prompt templates and decoding hyperparameters as inheritable agent configurations. By leveraging the leader-follower mechanism of the Grey Wolf Optimizer (GWO), we automatically select three leader agents ($α$, $β$, and $δ$) to guide the collaborative updates of the remaining agents, enabling iterative convergence toward robust optimal reasoning configurations that can be seamlessly integrated for inference. Extensive experiments on multiple mathematical and hybrid reasoning benchmarks across diverse LLM backbones show that Agent-GWO consistently improves accuracy and stability over existing prompt optimization methods. The code will be released publicly.

Jewon Lee, Wooksu Shin, Seungmin Yang et al.

Efficient processing of high-resolution images is crucial for real-world vision-language applications. However, existing Large Vision-Language Models (LVLMs) incur substantial computational overhead due to the large number of vision tokens. With the advent of "thinking with images" models, reasoning now extends beyond text to the visual domain. This capability motivates our two-stage "coarse-to-fine" reasoning pipeline: first, a downsampled image is analyzed to identify task-relevant regions; then, only these regions are cropped at full resolution and processed in a subsequent reasoning stage. This approach reduces computational cost while preserving fine-grained visual details where necessary. A major challenge lies in inferring which regions are truly relevant to a given query. Recent related methods often fail in the first stage after input-image downsampling, due to perception-driven reasoning, where clear visual information is required for effective reasoning. To address this issue, we propose ERGO (Efficient Reasoning & Guided Observation) that performs reasoning-driven perception-leveraging multimodal context to determine where to focus. Our model can account for perceptual uncertainty, expanding the cropped region to cover visually ambiguous areas for answering questions. To this end, we develop simple yet effective reward components in a reinforcement learning framework for coarse-to-fine perception. Across multiple datasets, our approach delivers higher accuracy than the original model and competitive methods, with greater efficiency. For instance, ERGO surpasses Qwen2.5-VL-7B on the V* benchmark by 4.7 points while using only 23% of the vision tokens, achieving a 3x inference speedup. The code and models can be found at: https://github.com/nota-github/ERGO.

Joseph Cho, Fachrina Dewi Puspitasari, Sheng Zheng et al.

The evolution of video generation from text, starting with animating MNIST numbers to simulating the physical world with Sora, has progressed at a breakneck speed over the past seven years. While often seen as a superficial expansion of the predecessor text-to-image generation model, text-to-video generation models are developed upon carefully engineered constituents. Here, we systematically discuss these elements consisting of but not limited to core building blocks (vision, language, and temporal) and supporting features from the perspective of their contributions to achieving a world model. We employ the PRISMA framework to curate 97 impactful research articles from renowned scientific databases primarily studying video synthesis using text conditions. Upon minute exploration of these manuscripts, we observe that text-to-video generation involves more intricate technologies beyond the plain extension of text-to-image generation. Our additional review into the shortcomings of Sora-generated videos pinpoints the call for more in-depth studies in various enabling aspects of video generation such as dataset, evaluation metric, efficient architecture, and human-controlled generation. Finally, we conclude that the study of the text-to-video generation may still be in its infancy, requiring contribution from the cross-discipline research community towards its advancement as the first step to realize artificial general intelligence (AGI).

Thibault Castells, Hyoung-Kyu Song, Tairen Piao et al.

The intensive computational burden of Stable Diffusion (SD) for text-to-image generation poses a significant hurdle for its practical application. To tackle this challenge, recent research focuses on methods to reduce sampling steps, such as Latent Consistency Model (LCM), and on employing architectural optimizations, including pruning and knowledge distillation. Diverging from existing approaches, we uniquely start with a compact SD variant, BK-SDM. We observe that directly applying LCM to BK-SDM with commonly used crawled datasets yields unsatisfactory results. It leads us to develop two strategies: (1) leveraging high-quality image-text pairs from leading generative models and (2) designing an advanced distillation process tailored for LCM. Through our thorough exploration of quantization, profiling, and on-device deployment, we achieve rapid generation of photo-realistic, text-aligned images in just two steps, with latency under one second on resource-limited edge devices.

Jewon Lee, Ki-Ung Song, Seungmin Yang et al.

Visual token reduction lowers inference costs caused by extensive image features in large vision-language models (LVLMs). Unlike relevant studies that prune tokens in self-attention-only LVLMs, our work uniquely addresses cross-attention-based models, which achieve superior performance. We identify that the key-value (KV) cache size for image tokens in cross-attention layers significantly exceeds that of text tokens in self-attention layers, posing a major compute bottleneck. To mitigate this issue, we exploit the sparse nature in cross-attention maps to selectively prune redundant visual features. Our Trimmed Llama effectively reduces KV cache demands without requiring additional training. By benefiting from 50%-reduced visual features, our model can reduce inference latency and memory usage while achieving benchmark parity.

Geonmin Kim, Jaeyeon Kim, Hancheol Park et al.

Thanks to unprecedented language understanding and generation capabilities of large language model (LLM), Retrieval-augmented Code Generation (RaCG) has recently been widely utilized among software developers. While this has increased productivity, there are still frequent instances of incorrect codes being provided. In particular, there are cases where plausible yet incorrect codes are generated for queries from users that cannot be answered with the given queries and API descriptions. This study proposes a task for evaluating answerability, which assesses whether valid answers can be generated based on users' queries and retrieved APIs in RaCG. Additionally, we build a benchmark dataset called Retrieval-augmented Code Generability Evaluation (RaCGEval) to evaluate the performance of models performing this task. Experimental results show that this task remains at a very challenging level, with baseline models exhibiting a low performance of 46.7%. Furthermore, this study discusses methods that could significantly improve performance.

Seul-Ki Yeom, Marcel Simon, Eunbin Lee et al.

Self-supervised learning (SSL) has made rapid progress, yet learned features often over-rely on contextual shortcuts-background textures and co-occurrence statistics. While video provides rich temporal variation, dense in-the-wild streams with strong ego-motion create a co-occurrence trap: foreground objects and background context move coherently, encouraging representations to collapse into scene encoders. To address this, we propose VINO (Video-driven Invariance for Non-Contextual Objects), a teacher-student framework that learns robust image encoders from dense video by imposing a structural information bottleneck. Using a class-agnostic structural prior solely to generate views-not as semantic pseudo-labels-VINO forms an asymmetric distillation problem. The teacher predicts from a foreground-union view with the background suppressed, while the student observes object-conditioned scene views that retain surrounding context but remove competing instances. Matching these targets via masked distillation makes background cues unreliable, pushing the representation toward object-centric invariances. We further enforce temporal object permanence via teacher-anchored cross-time distillation over track-matched objects, and stabilize part-to-whole consistency with mask-guided local views. Through attention visualization and unsupervised object discovery on PASCAL VOC, we demonstrate that VINO effectively disentangles foreground from background. Pretrained on the dense Walking Tours Venice video, VINO achieves 34.8 CorLoc, yielding highly focused, shape-biased representations that substantially outperform prior dense-video and motion-guided SSL baselines.

Marcel Simon, Tae-Ho Kim, Seul-Ki Yeom

Self-supervised image encoders such as DINO have recently gained significant interest for learning robust visual features without labels. However, most SSL methods train on static images and miss the temporal cues inherent in videos. We introduce a video-distilled single-image encoder trained to predict the next-frame representation from the current frame. This simple objective injects 3D spatial and temporal priors without optical flow or tracking. When pre-training on a single 2-hour video, our approach raises the mean Intersection-over-Union (mIoU) on ADE20K from 35.0 (DoRA) to 36.4 while remaining a drop-in replacement for image-only pipelines. Our results highlight video self-distillation as a lightweight route to geometry-aware perception an essential ingredient for physically plausible world models and Physical AI.

SoYoung Park, Hyewon Lee, Mingyu Choi et al.

Anomaly segmentation is essential for industrial quality, maintenance, and stability. Existing text-guided zero-shot anomaly segmentation models are effective but rely on fixed prompts, limiting adaptability in diverse industrial scenarios. This highlights the need for flexible, context-aware prompting strategies. We propose Image-Aware Prompt Anomaly Segmentation (IAP-AS), which enhances anomaly segmentation by generating dynamic, context-aware prompts using an image tagging model and a large language model (LLM). IAP-AS extracts object attributes from images to generate context-aware prompts, improving adaptability and generalization in dynamic and unstructured industrial environments. In our experiments, IAP-AS improves the F1-max metric by up to 10%, demonstrating superior adaptability and generalization. It provides a scalable solution for anomaly segmentation across industries

Seul-Ki Yeom, Tae-Ho Kim

Transformer-based architectures have demonstrated remarkable success across various domains, but their deployment on edge devices remains challenging due to high memory and computational demands. In this paper, we introduce a novel Reuse Attention mechanism, tailored for efficient memory access and computational optimization, enabling seamless operation on resource-constrained platforms without compromising performance. Unlike traditional multi-head attention (MHA), which redundantly computes separate attention matrices for each head, Reuse Attention consolidates these computations into a shared attention matrix, significantly reducing memory overhead and computational complexity. Comprehensive experiments on ImageNet-1K and downstream tasks show that the proposed UniForm models leveraging Reuse Attention achieve state-of-the-art imagenet classification accuracy while outperforming existing attention mechanisms, such as Linear Attention and Flash Attention, in inference speed and memory scalability. Notably, UniForm-l achieves a 76.7% Top-1 accuracy on ImageNet-1K with 21.8ms inference time on edge devices like the Jetson AGX Orin, representing up to a 5x speedup over competing benchmark methods. These results demonstrate the versatility of Reuse Attention across high-performance GPUs and edge platforms, paving the way for broader real-time applications

Tae-Ho Kim, Sungjae Cho, Shinkook Choi et al.

Voice conversion (VC) is a task to transform a person's voice to different style while conserving linguistic contents. Previous state-of-the-art on VC is based on sequence-to-sequence (seq2seq) model, which could mislead linguistic information. There was an attempt to overcome it by using textual supervision, it requires explicit alignment which loses the benefit of using seq2seq model. In this paper, a voice converter using multitask learning with text-to-speech (TTS) is presented. The embedding space of seq2seq-based TTS has abundant information on the text. The role of the decoder of TTS is to convert embedding space to speech, which is same to VC. In the proposed model, the whole network is trained to minimize loss of VC and TTS. VC is expected to capture more linguistic information and to preserve training stability by multitask learning. Experiments of VC were performed on a male Korean emotional text-speech dataset, and it is shown that multitask learning is helpful to keep linguistic contents in VC.

Hyoung-Kyu Song, Ebrahim AlAlkeem, Jaewoong Yun et al.

Identification using biometrics is an important yet challenging task. Abundant research has been conducted on identifying personal identity or gender using given signals. Various types of biometrics such as electrocardiogram (ECG), electroencephalogram (EEG), face, fingerprint, and voice have been used for these tasks. Most research has only focused on single modality or a single task, while the combination of input modality or tasks is yet to be investigated. In this paper, we propose deep identification and gender classification using multimodal biometrics. Our model uses ECG, fingerprint, and facial data. It then performs two tasks: gender identification and classification. By engaging multi-modality, a single model can handle various input domains without training each modality independently, and the correlation between domains can increase its generalization performance on the tasks.

Azam Rabiee, Tae-Ho Kim, Soo-Young Lee

Emotion is not limited to discrete categories of happy, sad, angry, fear, disgust, surprise, and so on. Instead, each emotion category is projected into a set of nearly independent dimensions, named pleasure (or valence), arousal, and dominance, known as PAD. The value of each dimension varies from -1 to 1, such that the neutral emotion is in the center with all-zero values. Training an emotional continuous text-to-speech (TTS) synthesizer on the independent dimensions provides the possibility of emotional speech synthesis with unlimited emotion categories. Our end-to-end neural speech synthesizer is based on the well-known Tacotron. Empirically, we have found the optimum network architecture for injecting the 3D PADs. Moreover, the PAD values are adjusted for the speech synthesis purpose.

Dae-Woong Jeong, Jaehun Kim, Youngseok Kim et al.

Existing high-performance deep learning models require very intensive computing. For this reason, it is difficult to embed a deep learning model into a system with limited resources. In this paper, we propose the novel idea of the network compression as a method to solve this limitation. The principle of this idea is to make iterative pruning more effective and sophisticated by simulating the reduced network. A simple experiment was conducted to evaluate the method; the results showed that the proposed method achieved higher performance than existing methods at the same pruning level.

Azam Rabiee, Geonmin Kim, Tae-Ho Kim et al.

This paper introduces a deep neural network model for subband-based speech synthesizer. The model benefits from the short bandwidth of the subband signals to reduce the complexity of the time-domain speech generator. We employed the multi-level wavelet analysis/synthesis to decompose/reconstruct the signal into subbands in time domain. Inspired from the WaveNet, a convolutional neural network (CNN) model predicts subband speech signals fully in time domain. Due to the short bandwidth of the subbands, a simple network architecture is enough to train the simple patterns of the subbands accurately. In the ground truth experiments with teacher-forcing, the subband synthesizer outperforms the fullband model significantly in terms of both subjective and objective measures. In addition, by conditioning the model on the phoneme sequence using a pronunciation dictionary, we have achieved the fully time-domain neural model for subband-based text-to-speech (TTS) synthesizer, which is nearly end-to-end. The generated speech of the subband TTS shows comparable quality as the fullband one with a slighter network architecture for each subband.

Myungsu Chae, Tae-Ho Kim, Young Hoon Shin et al.

Multi-task learning is a method for improving the generalizability of multiple tasks. In order to perform multiple classification tasks with one neural network model, the losses of each task should be combined. Previous studies have mostly focused on multiple prediction tasks using joint loss with static weights for training models, choosing the weights between tasks without making sufficient considerations by setting them uniformly or empirically. In this study, we propose a method to calculate joint loss using dynamic weights to improve the total performance, instead of the individual performance, of tasks. We apply this method to design an end-to-end multimodal emotion and gender recognition model using audio and video data. This approach provides proper weights for the loss of each task when the training process ends. In our experiments, emotion and gender recognition with the proposed method yielded a lower joint loss, which is computed as the negative log-likelihood, than using static weights for joint loss. Moreover, our proposed model has better generalizability than other models. To the best of our knowledge, this research is the first to demonstrate the strength of using dynamic weights for joint loss for maximizing overall performance in emotion and gender recognition tasks.