arXiv subjects

Jaejin Lee

Publications and source records attributed to Jaejin Lee.

At least 19 recordsLinked to original sources

arXiv2026

What Makes Video World Model Latents Action-Relevant: Prediction over Reconstruction

Video world models are increasingly used to provide predictive visual representations, yet it remains unclear which pretraining signals induce action-relevant structure in their latent spaces. We study this question through a unified probe-based evaluation across diverse encoder families, including image-only self-supervision, video pretraining with and without latent prediction, reconstruction-based autoencoders, diffusion models, and shortcut-forcing dynamics models. Using a common inverse-dynamics probing objective, we find that action-relevant structure is driven primarily by temporal video pretraining rather than pixel reconstruction fidelity: models with strong pixel decoding quality can exhibit near-zero action recoverability, while video-pretrained self-supervised encoders consistently achieve the best Pareto trade-off between visual fidelity and action prediction. Comparing V-JEPA and VideoMAE further shows that most gains arise from natural-video temporal context, with feature-level latent prediction providing a smaller additional benefit. These trends transfer across robotic benchmarks, though CALVIN reveals that static-environment tasks can partially mask the importance of temporal structure by allowing strong image priors to suffice. Finally, inverse-dynamics supervision substantially improves robustness to visual corruption, suggesting that action-aware objectives regularize latent geometry beyond clean-setting performance. Our results identify temporal predictive structure -- not reconstruction fidelity -- as the primary ingredient underlying action-relevant video representations.

cs.CV↗

arXiv2026

KBS: A Korean Benchmark Suite for Foundation Model Evaluation

We present a Korean Benchmark Suite (KBS) comprising Ko-ARC, Ko-GSM8K, Ko-EQ-Bench, Ko-WinoGrande, Ko-LAMBADA, and Ko-IFEval, covering six capabilities across 9,396 items. The suite combines review, localization, direct Korean construction, and redesign for Korean, avoiding a translation-only approach to benchmark adaptation. We evaluate Korean-developed models, multilingual open models, and closed API models under protocols that account for model access, prompt format, and scoring differences. Results show that Korean benchmark scores depend on task construction, scoring rules, prompting, and model post-training. These findings suggest that Korean evaluation should report both final scores and the settings used to obtain them.

cs.CL↗

arXiv2026

Thunder-DeID+: Boundary-Aware Korean De-identification from Court Judgments to Broader Sensitive Documents

This study examines the transferability of a de-identification model developed for Korean court judgments to a broader range of sensitive documents, focusing on boundary precision, label granularity, and recall. The agglutinative structure of Korean causes Byte Pair Encoding (BPE) tokenizers to frequently merge Personally Identifiable Information (PII) with adjacent particles, which complicates accurate span recovery. Furthermore, current PII taxonomies lack sufficient granularity for legal texts. To address these issues, this study introduces Thunder-DeID+, which integrates a refined three-tier PII taxonomy with a boundary-aware tokenizer. The tokenizer enhances a pretrained BPE vocabulary by incorporating N-gram and boundary-scoring components to better align token and PII span boundaries. The proposed approach is evaluated using a tokenizer-independent word-segment-level protocol. On full-length judgments, the tokenizer consistently improves W. Strict Typed F1-score from 0.76-0.77 to 0.87-0.88 across four model sizes (150M, 395M, 750M, and 1B), mainly by increasing precision rather than model scale. With fine-tuning on 1,000 broader sensitive documents, the judgment-trained model further achieves 0.93 W. Strict Typed F1-score, indicating effective transfer beyond court judgments.

cs.CL↗

arXiv2026

Efficient Expert-Parallel Communication on PCIe-Connected Consumer GPUs

Expert parallelism (EP) enables inference of large Mixture-of-Experts (MoE) models by placing their experts across multiple GPUs, but requires substantial communication between GPUs at every MoE layer. As contemporary MoE models activate more experts per token, this communication accounts for a growing fraction of inference time. The cost becomes particularly pronounced on PCIe-based consumer GPU systems, where all inter-GPU transfers traverse CPU memory. However, existing MoE-specialized EP communication libraries assume that direct GPU-to-GPU access is available, largely overlooking consumer GPUs. Therefore, most LLM frameworks instead rely on NCCL, whose CPU-staged communication incurs redundant PCIe transfers and competes with expert computation for GPU resources, limiting their overlap. We present ThunderEP, a novel communication design for such systems that removes the relay hops of traditional ring algorithm, moves data through DMA engines to avoid compute resource contention, and minimizes synchronization latency by reducing the polling overhead of completion flags in CPU memory. We integrate the proposed design into vLLM and evaluate it on three widely used MoE models. Experiments on two PCIe systems equipped with RTX 4090 and RTX 5090 GPUs show that ThunderEP achieves average speedups of 2.00$\times$ and 1.53$\times$ over NCCL for dispatch and combine, respectively, and up to 1.66$\times$ end-to-end speedup over state-of-the-art MoE inference frameworks.

cs.DC↗

arXiv2026

Less Is More: Reducing Token Counts Without Compromising Performance

Tokenization directly affects the inference efficiency of large language models, since fragmented tokenization increases sequence length and generation cost. Although longer, multi-word tokens can reduce fertility, naively adding them often degrades language model performance. We propose Thunder-Tok, a subword tokenizer that reduces fertility while preserving downstream performance. Thunder-Tok first constructs a large seed vocabulary from corpus substrings and filters structurally incomplete candidates, including invalid Unicode byte fragments and word-boundary violations. It then prunes the seed vocabulary using a likelihood-based token score derived from a uniform Jensen lower bound of the training-data probability. Experiments show that Thunder-Tok reduces fertility by approximately 25% in English and 9% in Korean compared with the standard BPE tokenizer while maintaining competitive performance.

cs.CL↗

arXiv2026

Thunder-KoNUBench: A Corpus-Aligned Benchmark for Korean Negation Understanding

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 negation understanding benchmark that reflects the empirical distribution of Korean negation phenomena. Evaluating 47 LLMs on Thunder-KoNUBench, we analyze the effects of model size and instruction tuning, and perform error analysis to better understand model behavior. We further show that fine-tuning on Thunder-KoNUBench improves negation understanding and broader contextual comprehension in Korean.

cs.CL↗

arXiv2026

Polar: A Benchmark for Evaluating Political Bias in LLMs

Political bias in large language models (LLMs) is increasingly significant, but difficult to measure reproducibly across political and linguistic contexts. We introduce Polar, a 4,026-instance multiple-choice benchmark that measures political bias through option-level likelihoods rather than prompt-based generation. Polar covers two ideological axes and eight issue categories derived from the Manifesto Project, and evaluates models in parallel across U.S. and South Korean political contexts. Across 38 LLMs, measured bias varies systematically with political context, issue category, model group, and presentation language. All models lean left-progressive on U.S. political content, but show more centered and mixed patterns on South Korean content. Translation experiments further show that presentation language alone can shift measured bias. These findings highlight the need for multilingual and cross-contextual evaluation of political bias in LLMs.

cs.CL↗

arXiv2026

SEDD: Scalable and Efficient Dataset Deduplication with GPUs

Dataset deduplication is widely recognized as a crucial preprocessing step that enhances data quality and improves the performance of large language models. A commonly used method for this process is the MinHash Locality-Sensitive Hashing (LSH) algorithm. Recently, GPU-accelerated frameworks such as NVIDIA NeMo Curator have been introduced to handle large-scale corpora; however, they remain suboptimal due to high communication overhead from physical data shuffling and underutilization of GPU resources. In this paper, we propose SEDD, a high-performance GPU-accelerated deduplication framework optimized for distributed cluster environments. SEDD introduces a computationally efficient, partially reusable hash function, alongside highly optimized GPU kernels and a hardware-aware automatic parameter selection mechanism. By replacing traditional data shuffling with a streaming-based approach, SEDD significantly mitigates communication bottlenecks. Our framework outperforms the CPU-based deduplication tool in SlimPajama by up to 158$\times$ and the GPU-based tool in NVIDIA NeMo Curator by up to 7.8$\times$ when processing 30 million documents on a node with four GPUs. Notably, SEDD dramatically accelerates the previously time-consuming MinHash signature generation phase, achieving speedups of up to 375$\times$ over the CPU baseline. Despite these gains in efficiency, SEDD maintains high deduplication fidelity, with duplicate document sets achieving Jaccard similarities of over 0.95 compared to those identified by the standard MinHash algorithm. In large-scale experiments, the deduplication of 1.2 trillion tokens is completed in just 3 hours on an 8-node 32-GPU V100 cluster. The related code is publicly available on GitHub (https://github.com/mcrl/SEDD).

cs.CL↗

arXiv2026

Reasoning by Commented Code for Table Question Answering

Table Question Answering (TableQA) poses a significant challenge for large language models (LLMs) because conventional linearization of tables often disrupts the two-dimensional relationships intrinsic to structured data. Existing methods, which depend on end-to-end answer generation or single-line program queries, typically exhibit limited numerical accuracy and reduced interpretability. This work introduces a commented, step-by-step code-generation framework that incorporates explicit reasoning into the Python program-generation process. The approach decomposes TableQA reasoning into multi-line executable programs with concise natural language comments, thereby promoting clearer reasoning and increasing the likelihood of generating correct code. On the WikiTableQuestions benchmark, the proposed method achieves 70.9\% accuracy using Qwen2.5-Coder-7B-Instruct, surpassing the Repanda baseline (67.6\%). Integrating the proposed framework with a robust end-to-end TableQA model via a lightweight answer-selection mechanism yields further improvements. This combined approach achieves up to 84.3\% accuracy on the WikiTableQuestions benchmark.

cs.CL↗

arXiv2026

HetCCL: Accelerating LLM Training with Heterogeneous GPUs

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.

cs.DC↗

arXiv2026

Models Know Models Best: Evaluation via Model-Preferred Formats

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.

cs.CL↗

arXiv2026

Thunder-NUBench: A Benchmark for LLMs' Sentence-Level Negation Understanding

Negation is a fundamental linguistic phenomenon that poses ongoing challenges for Large Language Models (LLMs), particularly in tasks requiring deep semantic understanding. Current benchmarks often treat negation as a minor detail within broader tasks, such as natural language inference. Consequently, there is a lack of benchmarks specifically designed to evaluate comprehension of negation. In this work, we introduce Thunder-NUBench, a novel benchmark explicitly created to assess sentence-level understanding of negation in LLMs. Thunder-NUBench goes beyond merely identifying surface-level cues by contrasting standard negation with structurally diverse alternatives, such as local negation, contradiction, and paraphrase. This benchmark includes manually curated sentence-negation pairs and a multiple-choice dataset, allowing for a comprehensive evaluation of models' understanding of negation.

cs.CL↗

arXiv2026

Choices Speak Louder than Questions

Recent findings raise concerns about whether the evaluation of Multiple-Choice Question Answering (MCQA) accurately reflects the comprehension abilities of large language models. This paper explores the concept of choice sensitivity, which refers to the tendency for model decisions to be more influenced by the answer options than by a genuine understanding of the question. We introduce a new scoring method called Normalized Probability Shift by the Question (NPSQ), designed to isolate the impact of the question itself and provide a more reliable assessment of comprehension. Through experiments involving various input formats, including cloze, symbols, and hybrid formats, we find that traditional scoring methods - such as those based on log-likelihood or its length-normalized variant - are vulnerable to superficial characteristics of the answer choices. In contrast, NPSQ remains stable even when modifications are made to the answer options.

cs.CL↗

arXiv2025

MNM : Multi-level Neuroimaging Meta-analysis with Hyperbolic Brain-Text Representations

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.

cs.LG↗

arXiv2025

PPMI: Privacy-Preserving LLM Interaction with Socratic Chain-of-Thought Reasoning and Homomorphically Encrypted Vector Databases

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.

cs.CR↗

arXiv2025

Beyond Line-Level Filtering for the Pretraining Corpora of LLMs

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.

cs.CL↗

arXiv2025

Ko-MuSR: A Multistep Soft Reasoning Benchmark for LLMs Capable of Understanding Korean

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.

cs.CL↗

arXiv2025

Thunder-DeID: Accurate and Efficient De-identification Framework for Korean Court Judgments

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.

cs.CL↗