Jaejin Lee

Youngjun Son, Chaewon Kim, Jaejin Lee

Dataset deduplication plays a crucial role in enhancing data quality, ultimately improving the training performance and efficiency of large language models. A commonly used method for data deduplication is the MinHash LSH algorithm. Recently, NVIDIA introduced a GPU-based MinHash LSH deduplication method, but it remains suboptimal, leaving room for further improvement in processing efficiency. This paper proposes a GPU-accelerated deduplication framework, FED, that optimizes MinHash LSH for GPU clusters and leverages computationally efficient, partially reusable non-cryptographic hash functions. FED significantly outperforms the CPU-based deduplication tool in SlimPajama (using 64 logical CPU cores) by up to 107.2 times and the GPU-based tool in NVIDIA NeMo Curator by up to 6.3 times when processing 30 million documents on a node with four GPUs. Notably, our method dramatically accelerates the previously time-consuming MinHash signature generation phase, achieving speed-ups of up to 260 compared to the CPU baseline. Despite these gains in efficiency, FED maintains high deduplication quality, with the duplicate document sets reaching a Jaccard similarity of over 0.96 compared to those identified by the standard MinHash algorithm. In large-scale experiments, the deduplication of 1.2 trillion tokens is completed in just 6 hours in a four-node, 16-GPU environment. The related code is publicly available on GitHub (\href{https://github.com/mcrl/FED}{https://github.com/mcrl/FED}).

Jinpyo Kim, Wooekun Jung, Hyungmo Kim et al.

Deep Convolutional Neural Networks (CNNs) are empirically known to be invariant to moderate translation but not to rotation in image classification. This paper proposes a deep CNN model, called CyCNN, which exploits polar mapping of input images to convert rotation to translation. To deal with the cylindrical property of the polar coordinates, we replace convolution layers in conventional CNNs to cylindrical convolutional (CyConv) layers. A CyConv layer exploits the cylindrically sliding windows (CSW) mechanism that vertically extends the input-image receptive fields of boundary units in a convolutional layer. We evaluate CyCNN and conventional CNN models for classification tasks on rotated MNIST, CIFAR-10, and SVHN datasets. We show that if there is no data augmentation during training, CyCNN significantly improves classification accuracies when compared to conventional CNN models. Our implementation of CyCNN is publicly available on https://github.com/mcrl/CyCNN.

Heehoon Kim, Jaehwan Lee, Taejeoung Kim et al.

The rapid growth of large language models is driving organizations to expand their GPU clusters, often with GPUs from multiple vendors. However, current deep learning frameworks lack support for collective communication across heterogeneous GPUs, leading to inefficiency and higher costs. We present HetCCL, a collective communication library that unifies vendor-specific backends and enables RDMA-based communication across GPUs without requiring driver modifications. HetCCL introduces two novel mechanisms that enable cross-vendor communication while leveraging optimized vendor libraries, NVIDIA NCCL and AMD RCCL. Evaluations on a multi-vendor GPU cluster show that HetCCL matches NCCL and RCCL performance in homogeneous setups while uniquely scaling in heterogeneous environments, enabling practical, high-performance training with both NVIDIA and AMD GPUs without changes to existing deep learning applications.

Seunghun Baek, Jaejin Lee, Jaeyoon Sim et al.

Various neuroimaging studies suffer from small sample size problem which often limit their reliability. Meta-analysis addresses this challenge by aggregating findings from different studies to identify consistent patterns of brain activity. However, traditional approaches based on keyword retrieval or linear mappings often overlook the rich hierarchical structure in the brain. In this work, we propose a novel framework that leverages hyperbolic geometry to bridge the gap between neuroscience literature and brain activation maps. By embedding text from research articles and corresponding brain images into a shared hyperbolic space via the Lorentz model, our method captures both semantic similarity and hierarchical organization inherent in neuroimaging data. In the hyperbolic space, our method performs multi-level neuroimaging meta-analysis (MNM) by 1) aligning brain and text embeddings for semantic correspondence, 2) guiding hierarchy between text and brain activations, and 3) preserving the hierarchical relationships within brain activation patterns. Experimental results demonstrate that our model outperforms baselines, offering a robust and interpretable paradigm of multi-level neuroimaging meta-analysis via hyperbolic brain-text representation.

Joonhak Lee, Sungmok Jung, Jongyeon Park et al.

Performance of Large Language Models (LLMs) on multiple-choice tasks differs markedly between symbol-based and cloze-style evaluation formats. The observed discrepancies are systematically attributable to task characteristics: natural language continuation benefits from likelihood scoring, whereas explicit comparison is better suited to symbol-based selection. These trends are consistent across various decoder-based LLMs, indicating model-agnostic effects. To address these inconsistencies, a dynamic format-alignment strategy is introduced that employs a lightweight classifier trained on latent model-preference signals. In contrast to human-designed heuristics, which often degrade performance, this approach uses model-generated signals to determine the optimal format for each problem instance. The proposed method achieves substantial and consistent improvements in zero-shot accuracy across reasoning and knowledge benchmarks, better revealing the models' latent capabilities.

Sungmok Jung, Yeonkyoung So, Joonhak Lee et al.

Although negation is known to challenge large language models (LLMs), benchmarks for evaluating negation understanding, especially in Korean, are scarce. We conduct a corpus-based analysis of Korean negation and show that LLM performance degrades under negation. We then introduce Thunder-KoNUBench, a sentence-level benchmark that reflects the empirical distribution of Korean negation phenomena. Evaluating 47 LLMs, we analyze the effects of model size and instruction tuning, and show that fine-tuning on Thunder-KoNUBench improves negation understanding and broader contextual comprehension in Korean.

Jaejin Lee, Seho Kee, Mani Janakiram et al.

Job shop scheduling problems represent a significant and complex facet of combinatorial optimization problems, which have traditionally been addressed through either exact or approximate solution methodologies. However, the practical application of these solutions is often challenged due to the complexity of real-world problems. Even when utilizing an approximate solution approach, the time required to identify a near-optimal solution can be prohibitively extensive, and the solutions derived are generally not applicable to new problems. This study proposes an innovative attention-based reinforcement learning method specifically designed for the category of job shop scheduling problems. This method integrates a policy gradient reinforcement learning approach with a modified transformer architecture. A key finding of this research is the ability of our trained learners within the proposed method to be repurposed for larger-scale problems that were not part of the initial training set. Furthermore, empirical evidence demonstrates that our approach surpasses the results of recent studies and outperforms commonly implemented heuristic rules. This suggests that our method offers a promising avenue for future research and practical application in the field of job shop scheduling problems.

Yubeen Bae, Minchan Kim, Jaejin Lee et al.

Large language models (LLMs) are increasingly used as personal agents, accessing sensitive user data such as calendars, emails, and medical records. Users currently face a trade-off: They can send private records, many of which are stored in remote databases, to powerful but untrusted LLM providers, increasing their exposure risk. Alternatively, they can run less powerful models locally on trusted devices. We bridge this gap. Our Socratic Chain-of-Thought Reasoning first sends a generic, non-private user query to a powerful, untrusted LLM, which generates a Chain-of-Thought (CoT) prompt and detailed sub-queries without accessing user data. Next, we embed these sub-queries and perform encrypted sub-second semantic search using our Homomorphically Encrypted Vector Database across one million entries of a single user's private data. This represents a realistic scale of personal documents, emails, and records accumulated over years of digital activity. Finally, we feed the CoT prompt and the decrypted records to a local language model and generate the final response. On the LoCoMo long-context QA benchmark, our hybrid framework, combining GPT-4o with a local Llama-3.2-1B model, outperforms using GPT-4o alone by up to 7.1 percentage points. This demonstrates a first step toward systems where tasks are decomposed and split between untrusted strong LLMs and weak local ones, preserving user privacy.

Kyusu Ahn, Jinpyo Kim, Chanwoo Park et al.

Under-Display Camera (UDC) houses a digital camera lens under a display panel. However, UDC introduces complex degradations such as noise, blur, decrease in transmittance, and flare. Despite the remarkable progress, previous research on UDC mainly focuses on eliminating diffraction in the spatial domain and rarely explores its potential in the frequency domain. It is essential to consider both the spatial and frequency domains effectively. For example, degradations, such as noise and blur, can be addressed by local information (e.g., CNN kernels in the spatial domain). At the same time, tackling flares may require leveraging global information (e.g., the frequency domain). In this paper, we revisit the UDC degradations in the Fourier space and figure out intrinsic frequency priors that imply the presence of the flares. Based on this observation, we propose a novel multi-level DNN architecture called SFIM. It efficiently restores UDC-distorted images by integrating local and global (the collective contribution of all points in the image) information. The architecture exploits CNNs to capture local information and FFT-based models to capture global information. SFIM comprises a spatial domain block (SDB), a Frequency Domain Block (FDB), and an Attention-based Multi-level Integration Block (AMIB). Specifically, SDB focuses more on detailed textures such as noise and blur, FDB emphasizes irregular texture loss in extensive areas such as flare, and AMIB enables effective cross-domain interaction. SFIM's superior performance over state-of-the-art approaches is demonstrated through rigorous quantitative and qualitative assessments across three UDC benchmarks.

Sungeun Hahm, Heejin Kim, Gyuseong Lee et al.

To ensure a balance between open access to justice and personal data protection, the South Korean judiciary mandates the de-identification of court judgments before they can be publicly disclosed. However, the current de-identification process is inadequate for handling court judgments at scale while adhering to strict legal requirements. Additionally, the legal definitions and categorizations of personal identifiers are vague and not well-suited for technical solutions. To tackle these challenges, we propose a de-identification framework called Thunder-DeID, which aligns with relevant laws and practices. Specifically, we (i) construct and release the first Korean legal dataset containing annotated judgments along with corresponding lists of entity mentions, (ii) introduce a systematic categorization of Personally Identifiable Information (PII), and (iii) develop an end-to-end deep neural network (DNN)-based de-identification pipeline. Our experimental results demonstrate that our model achieves state-of-the-art performance in the de-identification of court judgments.

Yeonkyoung So, Gyuseong Lee, Sungmok Jung et al.

Negation is a fundamental linguistic phenomenon that poses persistent challenges for Large Language Models (LLMs), particularly in tasks requiring deep semantic understanding. Existing benchmarks often treat negation as a side case within broader tasks like natural language inference, resulting in a lack of benchmarks that exclusively target negation understanding. In this work, we introduce Thunder-NUBench, a novel benchmark explicitly designed to assess sentence-level negation understanding in LLMs. Thunder-NUBench goes beyond surface-level cue detection by contrasting standard negation with structurally diverse alternatives such as local negation, contradiction, and paraphrase. The benchmark consists of manually curated sentence-negation pairs and a multiple-choice dataset that enables in-depth evaluation of models' negation understanding.

Chanwoo Park, Suyoung Park, JiA Kang et al.

We present Ko-MuSR, the first benchmark to comprehensively evaluate multistep, soft reasoning in long Korean narratives while minimizing data contamination. Built following MuSR, Ko-MuSR features fully Korean narratives, reasoning chains, and multiple-choice questions verified by human annotators for logical consistency and answerability. Evaluations of four large language models -- two multilingual and two Korean-specialized -- show that multilingual models outperform Korean-focused ones even in Korean reasoning tasks, indicating cross-lingual generalization of reasoning ability. Carefully designed prompting strategies, which combine few-shot examples, reasoning traces, and task-specific hints, further boost accuracy, approaching human-level performance. Ko-MuSR offers a solid foundation for advancing Korean NLP by enabling systematic evaluation of long-context reasoning and prompting strategies.

Chanwoo Park, Suyoung Park, Yelim Ahn et al.

While traditional line-level filtering techniques, such as line-level deduplication and trailing-punctuation filters, are commonly used, these basic methods can sometimes discard valuable content, negatively affecting downstream performance. In this paper, we introduce two methods-pattern-aware line-level deduplication (PLD) and pattern-aware trailing punctuation filtering (PTF)-by enhancing the conventional filtering techniques. Our approach not only considers line-level signals but also takes into account their sequential distribution across documents, enabling us to retain structurally important content that might otherwise be removed. We evaluate these proposed methods by training small language models (1 B parameters) in both English and Korean. The results demonstrate that our methods consistently improve performance on multiple-choice benchmarks and significantly enhance generative question-answering accuracy on both SQuAD v1 and KorQuAD v1.

Kyubyung Chae, Gihoon Kim, Gyuseong Lee et al.

Recent trends in LLMs development clearly show growing interest in the use and application of sovereign LLMs. The global debate over sovereign LLMs highlights the need for governments to develop their LLMs, tailored to their unique socio-cultural and historical contexts. However, there remains a shortage of frameworks and datasets to verify two critical questions: (1) how well these models align with users' socio-cultural backgrounds, and (2) whether they maintain safety and technical robustness without exposing users to potential harms and risks. To address this gap, we construct a new dataset and introduce an analytic framework for extracting and evaluating the socio-cultural elements of sovereign LLMs, alongside assessments of their technical robustness. Our experimental results demonstrate that while sovereign LLMs play a meaningful role in supporting low-resource languages, they do not always meet the popular claim that these models serve their target users well. We also show that pursuing this untested claim may lead to underestimating critical quality attributes such as safety. Our study suggests that advancing sovereign LLMs requires a more extensive evaluation that incorporates a broader range of well-grounded and practical criteria.

Yelim Ahn, Jaejin Lee

As large language models (LLMs) are increasingly deployed across diverse domains, ensuring their safety has become a critical concern. In response, studies on jailbreak attacks have been actively growing. Existing approaches typically rely on iterative prompt engineering or semantic transformations of harmful instructions to evade detection. In this work, we introduce PUZZLED, a novel jailbreak method that leverages the LLM's reasoning capabilities. It masks keywords in a harmful instruction and presents them as word puzzles for the LLM to solve. We design three puzzle types-word search, anagram, and crossword-that are familiar to humans but cognitively demanding for LLMs. The model must solve the puzzle to uncover the masked words and then proceed to generate responses to the reconstructed harmful instruction. We evaluate PUZZLED on five state-of-the-art LLMs and observe a high average attack success rate (ASR) of 88.8%, specifically 96.5% on GPT-4.1 and 92.3% on Claude 3.7 Sonnet. PUZZLED is a simple yet powerful attack that transforms familiar puzzles into an effective jailbreak strategy by harnessing LLMs' reasoning capabilities.

Seorin Kim, Dongyoung Lee, Jaejin Lee

Large language models (LLMs) often exhibit societal biases in their outputs, prompting ethical concerns regarding fairness and harm. In this work, we propose KLAAD (KL-Attention Alignment Debiasing), an attention-based debiasing framework that implicitly aligns attention distributions between stereotypical and anti-stereotypical sentence pairs without directly modifying model weights. KLAAD introduces a composite training objective combining Cross-Entropy, KL divergence, and Triplet losses, guiding the model to consistently attend across biased and unbiased contexts while preserving fluency and coherence. Experimental evaluation of KLAAD demonstrates improved bias mitigation on both the BBQ and BOLD benchmarks, with minimal impact on language modeling quality. The results indicate that attention-level alignment offers a principled solution for mitigating bias in generative language models.

Jinpyo Kim, Gyeongje Cho, Chanwoo Park et al.

Since state-of-the-art LLMs often underperform in languages other than English or Chinese, improving the capability of LLMs in new languages has become an essential task. Moreover, LLMs' entire end-to-end training process remains largely unknown to the public due to proprietary reasons, technical complexity, inconsistent documentation, and ethical considerations. The complete picture remains a closely guarded secret within the industry. This paper presents methods to adapt an existing English-based LLM to Korean in a low-budget scenario. We describe the entire end-to-end process: collecting Korean datasets, preprocessing the data, training the model, creating downstream benchmarks, and conducting evaluations. The evaluation results indicate that our method can effectively and cost-efficiently add new language capabilities to existing LLMs. Our new bilingual models, Thunder-LLM and Thunder-LLM-Ins, achieve superior Korean performance compared to state-of-the-art models while utilizing minimal data and computational resources. We share our comprehensive experience and make the code publicly available.

Gyeongje Cho, Yeonkyoun So, Chanwoo Park et al.

This paper introduces Thunder-Tok, a new Korean tokenizer designed to reduce token fertility without compromising model performance. Our approach uses a rule-based pre-tokenization method that aligns with the linguistic structure of the Korean language. We also create a seed vocabulary containing tokens that resemble linguistic units and employ a branching entropy-based selection algorithm. These techniques increase the average token length, thus lowering fertility while preserving linguistic information. Experimental results indicate that Thunder-Tok reduces fertility by approximately 10% (i.e., reduces the number of tokens by 10%, improving the inference speed by 10%) compared to BPE without compromising performance across various downstream tasks. These findings demonstrate that our linguistically informed approach is effective and practical for designing efficient tokenizers for language models.

Gyeongje Cho, Yeonkyoung So, Jaejin Lee

Multiple-choice benchmarks, consisting of various prompts and choices, are among the most widely used methods to assess a language model's natural language understanding capability. Given a specific prompt, we typically compute $P(Choice|Prompt)$ to evaluate how likely a language model is to generate the correct choice compared to incorrect ones. However, we observe that performance measured using this approach reflects not only the model's comprehension of the prompt but also its inherent biases for certain choices regardless of the prompt. This issue makes it challenging to accurately measure a model's natural language understanding, as models may select the answer without fully understanding the prompt. To address this limitation, we propose a novel metric called ANPMI, which normalizes Pointwise Mutual Information (PMI) by $-\log P(Choice)$. ANPMI provides a more accurate assessment of the model's natural language understanding by ensuring that it is challenging to answer a question without properly understanding the prompt.

Kyusu Ahn, JiSoo Kim, Sangik Lee et al.

Even though an Under-Display Camera (UDC) is an advanced imaging system, the display panel significantly degrades captured images or videos, introducing low transmittance, blur, noise, and flare issues. Tackling such issues is challenging because of the complex degradation of UDCs, including diverse flare patterns. However, no dataset contains videos of real-world UDC degradation. In this paper, we propose a real-world UDC video dataset called UDC-VIT. Unlike existing datasets, UDC-VIT exclusively includes human motions for facial recognition. We propose a video-capturing system to acquire clean and UDC-degraded videos of the same scene simultaneously. Then, we align a pair of captured videos frame by frame, using discrete Fourier transform (DFT). We compare UDC-VIT with six representative UDC still image datasets and two existing UDC video datasets. Using six deep-learning models, we compare UDC-VIT and an existing synthetic UDC video dataset. The results indicate the ineffectiveness of models trained on earlier synthetic UDC video datasets, as they do not reflect the actual characteristics of UDC-degraded videos. We also demonstrate the importance of effective UDC restoration by evaluating face recognition accuracy concerning PSNR, SSIM, and LPIPS scores. UDC-VIT is available at our official GitHub repository.