DongGeon Lee

Bum Jun Kim, Hyeyeon Choi, Hyeonah Jang et al.

For the stable optimization of deep neural networks, regularization methods such as dropout and batch normalization have been used in various tasks. Nevertheless, the correct position to apply dropout has rarely been discussed, and different positions have been employed depending on the practitioners. In this study, we investigate the correct position to apply dropout. We demonstrate that for a residual network with batch normalization, applying dropout at certain positions increases the performance, whereas applying dropout at other positions decreases the performance. Based on theoretical analysis, we provide the following guideline for the correct position to apply dropout: apply one dropout after the last batch normalization but before the last weight layer in the residual branch. We provide detailed theoretical explanations to support this claim and demonstrate them through module tests. In addition, we investigate the correct position of dropout in the head that produces the final prediction. Although the current consensus is to apply dropout after global average pooling, we prove that applying dropout before global average pooling leads to a more stable output. The proposed guidelines are validated through experiments using different datasets and models.

Dasol Choi, DongGeon Lee, Brigitta Jesica Kartono et al.

As large language models are deployed in high-stakes enterprise applications, from healthcare to finance, ensuring adherence to organization-specific policies has become essential. Yet existing safety evaluations focus exclusively on universal harms. We present COMPASS (Company/Organization Policy Alignment Assessment), the first systematic framework for evaluating whether LLMs comply with organizational allowlist and denylist policies. We apply COMPASS to eight diverse industry scenarios, generating and validating 5,920 queries that test both routine compliance and adversarial robustness through strategically designed edge cases. Evaluating seven state-of-the-art models, we uncover a fundamental asymmetry: models reliably handle legitimate requests (>95% accuracy) but catastrophically fail at enforcing prohibitions, refusing only 13-40% of adversarial denylist violations. These results demonstrate that current LLMs lack the robustness required for policy-critical deployments, establishing COMPASS as an essential evaluation framework for organizational AI safety.

DongGeon Lee, Joonwon Jang, Jihae Jeong et al.

Rapid deployment of vision-language models (VLMs) magnifies safety risks, yet most evaluations rely on artificial images. This study asks: How safe are current VLMs when confronted with meme images that ordinary users share? To investigate this question, we introduce MemeSafetyBench, a 50,430-instance benchmark pairing real meme images with both harmful and benign instructions. Using a comprehensive safety taxonomy and LLM-based instruction generation, we assess multiple VLMs across single and multi-turn interactions. We investigate how real-world memes influence harmful outputs, the mitigating effects of conversational context, and the relationship between model scale and safety metrics. Our findings demonstrate that VLMs are more vulnerable to meme-based harmful prompts than to synthetic or typographic images. Memes significantly increase harmful responses and decrease refusals compared to text-only inputs. Though multi-turn interactions provide partial mitigation, elevated vulnerability persists. These results highlight the need for ecologically valid evaluations and stronger safety mechanisms. MemeSafetyBench is publicly available at https://github.com/oneonlee/Meme-Safety-Bench.

DongGeon Lee, Ahjeong Park, Hyeri Lee et al.

Addressing non-factoid question answering (NFQA) remains challenging due to its open-ended nature, diverse user intents, and need for multi-aspect reasoning. These characteristics often reveal the limitations of conventional retrieval-augmented generation (RAG) approaches. To overcome these challenges, we propose Typed-RAG, a framework for type-aware decomposition of non-factoid questions (NFQs) within the RAG paradigm. Specifically, Typed-RAG first classifies an NFQ into a predefined type (e.g., Debate, Experience, Comparison). It then decomposes the question into focused sub-queries, each focusing on a single aspect. This decomposition enhances both retrieval relevance and answer quality. By combining the results of these sub-queries, Typed-RAG produces more informative and contextually aligned responses. Additionally, we construct Wiki-NFQA, a benchmark dataset for NFQA covering a wide range of NFQ types. Experiments show that Typed-RAG consistently outperforms existing QA approaches based on LLMs or RAG methods, validating the effectiveness of type-aware decomposition for improving both retrieval quality and answer generation in NFQA. Our code and dataset are available on https://github.com/TeamNLP/Typed-RAG.

DongGeon Lee, Hwanjo Yu

Hallucinations in large language model (LLM) outputs severely limit their reliability in knowledge-intensive tasks such as question answering. To address this challenge, we introduce REFIND (Retrieval-augmented Factuality hallucINation Detection), a novel framework that detects hallucinated spans within LLM outputs by directly leveraging retrieved documents. As part of the REFIND, we propose the Context Sensitivity Ratio (CSR), a novel metric that quantifies the sensitivity of LLM outputs to retrieved evidence. This innovative approach enables REFIND to efficiently and accurately detect hallucinations, setting it apart from existing methods. In the evaluation, REFIND demonstrated robustness across nine languages, including low-resource settings, and significantly outperformed baseline models, achieving superior IoU scores in identifying hallucinated spans. This work highlights the effectiveness of quantifying context sensitivity for hallucination detection, thereby paving the way for more reliable and trustworthy LLM applications across diverse languages. Our code is available at https://github.com/oneonlee/REFIND.

Jihae Jeong, DaeYeop Lee, DongGeon Lee et al.

Existing physical commonsense reasoning benchmarks predominantly focus on Western contexts, overlooking cultural variations in physical problem-solving. To address this gap, we introduce EPiK (Everyday Physics in Korean Contexts), a novel benchmark comprising 181 binary-choice problems that test physical reasoning within Korean cultural contexts, ranging from kimchi (Korean food) to traditional fermentation. EPiK is constructed using a two-stage generation and verification pipeline to create culturally-authentic problems across 9 reasoning subtasks and 84 scenarios. Unlike approaches based on simple translation, our method generates problems organically from Korean contexts while upholding rigorous physical reasoning standards. Our evaluations show that Korean-specialized models consistently outperform general-purpose models of comparable size. This performance gap highlights the limitations of culturally-agnostic models and demonstrates the critical need for culturally-aware benchmarks to truly measure language understanding. Our EPiK is publicly available at https://huggingface.co/datasets/jjae/EPiK.

Wonpyo Park, Woonggi Chang, Donggeon Lee et al.

We propose a novel positional encoding for learning graph on Transformer architecture. Existing approaches either linearize a graph to encode absolute position in the sequence of nodes, or encode relative position with another node using bias terms. The former loses preciseness of relative position from linearization, while the latter loses a tight integration of node-edge and node-topology interaction. To overcome the weakness of the previous approaches, our method encodes a graph without linearization and considers both node-topology and node-edge interaction. We name our method Graph Relative Positional Encoding dedicated to graph representation learning. Experiments conducted on various graph datasets show that the proposed method outperforms previous approaches significantly. Our code is publicly available at https://github.com/lenscloth/GRPE.

Sungjoon Park, Jaesub Yun, Donggeon Lee et al.

Because videos in the wild can be out of sync for various reasons, a sync-net is used to bring the video back into sync for tasks that require synchronized videos. Previous state-of-the-art (SOTA) sync-nets use InfoNCE loss, rely on the transformer architecture, or both. Unfortunately, the former makes the model's output difficult to interpret, and the latter is unfriendly with large images, thus limiting the usefulness of sync-nets. In this work, we train a convolutional sync-net using the balanced BCE loss (BBCE), a loss inspired by the binary cross entropy (BCE) and the InfoNCE losses. In contrast to the InfoNCE loss, the BBCE loss does not require complicated sampling schemes. Our model can better handle larger images, and its output can be given a probabilistic interpretation. The probabilistic interpretation allows us to define metrics such as probability at offset and offscreen ratio to evaluate the sync quality of audio-visual (AV) speech datasets. Furthermore, our model achieves SOTA accuracy of $96.5\%$ on the LRS2 dataset and $93.8\%$ on the LRS3 dataset.

Bodam Kim, Hiskias Dingeto, Taeyoun Kwon et al.

As large language models become increasingly integrated into daily life, audio has emerged as a key interface for human-AI interaction. However, this convenience also introduces new vulnerabilities, making audio a potential attack surface for adversaries. Our research introduces WhisperInject, a two-stage adversarial audio attack framework that can manipulate state-of-the-art audio language models to generate harmful content. Our method uses imperceptible perturbations in audio inputs that remain benign to human listeners. The first stage uses a novel reward-based optimization method, Reinforcement Learning with Projected Gradient Descent (RL-PGD), to guide the target model to circumvent its own safety protocols and generate harmful native responses. This native harmful response then serves as the target for Stage 2, Payload Injection, where we use Projected Gradient Descent (PGD) to optimize subtle perturbations that are embedded into benign audio carriers, such as weather queries or greeting messages. Validated under the rigorous StrongREJECT, LlamaGuard, as well as Human Evaluation safety evaluation framework, our experiments demonstrate a success rate exceeding 86% across Qwen2.5-Omni-3B, Qwen2.5-Omni-7B, and Phi-4-Multimodal. Our work demonstrates a new class of practical, audio-native threats, moving beyond theoretical exploits to reveal a feasible and covert method for manipulating AI behavior.

Tyler A. Chang, Catherine Arnett, Abdelrahman Eldesokey et al. · uw

To date, there exist almost no culturally-specific evaluation benchmarks for large language models (LLMs) that cover a large number of languages and cultures. In this paper, we present Global PIQA, a participatory commonsense reasoning benchmark for over 100 languages, constructed by hand by 335 researchers from 65 countries around the world. The 116 language varieties in Global PIQA cover five continents, 14 language families, and 23 writing systems. In the non-parallel split of Global PIQA, over 50% of examples reference local foods, customs, traditions, or other culturally-specific elements. We find that state-of-the-art LLMs perform well on Global PIQA in aggregate, but they exhibit weaker performance in lower-resource languages (up to a 37% accuracy gap, despite random chance at 50%). Open models generally perform worse than proprietary models. Global PIQA highlights that in many languages and cultures, everyday knowledge remains an area for improvement, alongside more widely-discussed capabilities such as complex reasoning and expert knowledge. Beyond its uses for LLM evaluation, we hope that Global PIQA provides a glimpse into the wide diversity of cultures in which human language is embedded.

Sungjoon Park, Minsik Park, Haneol Lee et al.

In this work, we revisit the effectiveness of 3DMM for talking head synthesis by jointly learning a 3D face reconstruction model and a talking head synthesis model. This enables us to obtain a FACS-based blendshape representation of facial expressions that is optimized for talking head synthesis. This contrasts with previous methods that either fit 3DMM parameters to 2D landmarks or rely on pretrained face reconstruction models. Not only does our approach increase the quality of the generated face, but it also allows us to take advantage of the blendshape representation to modify just the mouth region for the purpose of audio-based lip-sync. To this end, we propose a novel lip-sync pipeline that, unlike previous methods, decouples the original chin contour from the lip-synced chin contour, and reduces flickering near the mouth.

TaeYoon Kwack, Jisoo Kim, Ki Yong Jung et al.

Tables are a primary medium for conveying critical information in administrative domains, yet their complexity hinders utilization by Large Language Models (LLMs). This paper introduces the Theme-Explanation Structure-based Table Summarization (Tabular-TX) pipeline, a novel approach designed to generate highly interpretable summaries from tabular data, with a specific focus on Korean administrative documents. Current table summarization methods often neglect the crucial aspect of human-friendly output. Tabular-TX addresses this by first employing a multi-step reasoning process to ensure deep table comprehension by LLMs, followed by a journalist persona prompting strategy for clear sentence generation. Crucially, it then structures the output into a Theme Part (an adverbial phrase) and an Explanation Part (a predicative clause), significantly enhancing readability. Our approach leverages in-context learning, obviating the need for extensive fine-tuning and associated labeled data or computational resources. Experimental results show that Tabular-TX effectively processes complex table structures and metadata, offering a robust and efficient solution for generating human-centric table summaries, especially in low-resource scenarios.

Donggeon Lee, Seongho Choi, Youwon Jang et al.

Video question answering has recently received a lot of attention from multimodal video researchers. Most video question answering datasets are usually in the form of multiple-choice. But, the model for the multiple-choice task does not infer the answer. Rather it compares the answer candidates for picking the correct answer. Furthermore, it makes it difficult to extend to other tasks. In this paper, we challenge the existing multiple-choice video question answering by changing it to open-ended video question answering. To tackle open-ended question answering, we use the pretrained GPT2 model. The model is fine-tuned with video inputs and subtitles. An ablation study is performed by changing the existing DramaQA dataset to an open-ended question answering, and it shows that performance can be improved using video metadata.