Hyunjun Kim

Minsu Kim, Hyunjun Kim, Yong Man Ro

Lip reading aims to predict speech based on lip movements alone. As it focuses on visual information to model the speech, its performance is inherently sensitive to personal lip appearances and movements. This makes the lip reading models show degraded performance when they are applied to unseen speakers due to the mismatch between training and testing conditions. Speaker adaptation technique aims to reduce this mismatch between train and test speakers, thus guiding a trained model to focus on modeling the speech content without being intervened by the speaker variations. In contrast to the efforts made in audio-based speech recognition for decades, the speaker adaptation methods have not well been studied in lip reading. In this paper, to remedy the performance degradation of lip reading model on unseen speakers, we propose a speaker-adaptive lip reading method, namely user-dependent padding. The user-dependent padding is a speaker-specific input that can participate in the visual feature extraction stage of a pre-trained lip reading model. Therefore, the lip appearances and movements information of different speakers can be considered during the visual feature encoding, adaptively for individual speakers. Moreover, the proposed method does not need 1) any additional layers, 2) to modify the learned weights of the pre-trained model, and 3) the speaker label of train data used during pre-train. It can directly adapt to unseen speakers by learning the user-dependent padding only, in a supervised or unsupervised manner. Finally, to alleviate the speaker information insufficiency in public lip reading databases, we label the speaker of a well-known audio-visual database, LRW, and design an unseen-speaker lip reading scenario named LRW-ID.

Sungjune Park, Hyunjun Kim, Yong Man Ro

Large language models (LLMs) have shown their capabilities in understanding contextual and semantic information regarding knowledge of instance appearances. In this paper, we introduce a novel approach to utilize the strengths of LLMs in understanding contextual appearance variations and to leverage this knowledge into a vision model (here, pedestrian detection). While pedestrian detection is considered one of the crucial tasks directly related to our safety (e.g., intelligent driving systems), it is challenging because of varying appearances and poses in diverse scenes. Therefore, we propose to formulate language-derived appearance elements and incorporate them with visual cues in pedestrian detection. To this end, we establish a description corpus that includes numerous narratives describing various appearances of pedestrians and other instances. By feeding them through an LLM, we extract appearance knowledge sets that contain the representations of appearance variations. Subsequently, we perform a task-prompting process to obtain appearance elements which are guided representative appearance knowledge relevant to a downstream pedestrian detection task. The obtained knowledge elements are adaptable to various detection frameworks, so that we can provide plentiful appearance information by integrating the language-derived appearance elements with visual cues within a detector. Through comprehensive experiments with various pedestrian detectors, we verify the adaptability and effectiveness of our method showing noticeable performance gains and achieving state-of-the-art detection performance on two public pedestrian detection benchmarks (i.e., CrowdHuman and WiderPedestrian).

Jeong Hun Yeo, Chae Won Kim, Hyunjun Kim et al.

Lip reading aims to predict spoken language by analyzing lip movements. Despite advancements in lip reading technologies, performance degrades when models are applied to unseen speakers due to their sensitivity to variations in visual information such as lip appearances. To address this challenge, speaker adaptive lip reading technologies have advanced by focusing on effectively adapting a lip reading model to target speakers in the visual modality. However, the effectiveness of adapting language information, such as vocabulary choice, of the target speaker has not been explored in previous works. Additionally, existing datasets for speaker adaptation have limited vocabulary sizes and pose variations, which restrict the validation of previous speaker-adaptive methods in real-world scenarios. To address these issues, we propose a novel speaker-adaptive lip reading method that adapts a pre-trained model to target speakers at both vision and language levels. Specifically, we integrate prompt tuning and the LoRA approach, applying them to a pre-trained lip reading model to effectively adapt the model to target speakers. Furthermore, to validate its effectiveness in real-world scenarios, we introduce a new dataset, VoxLRS-SA, derived from VoxCeleb2 and LRS3. It contains a vocabulary of approximately 100K words, offers diverse pose variations, and enables the validation of adaptation methods in the wild, sentence-level lip reading for the first time in English. Through various experiments, we demonstrate that the existing speaker-adaptive method also improves performance in the wild at the sentence level. Moreover, we show that the proposed method achieves larger improvements compared to the previous works.

Taein Lim, Seongyong Ju, Munhyeok Kim et al.

Large language models (LLMs) are increasingly deployed as autonomous agents in offensive cybersecurity. In this paper, we reveal an interesting phenomenon: different agents exhibit distinct attack patterns. Specifically, each agent exhibits an attack-selection bias, disproportionately concentrating its efforts on a narrow subset of attack families regardless of prompt variations. To systematically quantify this behavior, we introduce CyBiasBench, a comprehensive 630-session benchmark that evaluates five agents on three targets and four prompt conditions with ten attack families. We identify explicit bias across agents, with different dominant attack families and varying entropy levels in their attack-family allocation distributions. Such bias is better characterized as a trait of the agents, rather than a factor associated with the attack success rate. Furthermore, our experiments reveal a bias momentum effect, where agents resist explicit steering toward attack families that conflict with their bias. This forced distribution shift does not yield measurable improvements in attack performance. To ensure reproducibility and facilitate future research, we release an interactive result dashboard at https://trustworthyai.co.kr/CyBiasBench/ and a reproducibility artifact with aggregated session-level statistics and full evaluation scripts at https://github.com/Harry24k/CyBiasBench.

Kirill Paramonov, Mete Ozay, Aristeidis Mystakidis et al.

The proliferation of AI models in everyday devices has highlighted a critical challenge: prediction errors that degrade user experience. While existing solutions focus on error detection, they rarely provide efficient correction mechanisms, especially for resource-constrained devices. We present a novel system enabling users to correct AI misclassifications through few-shot learning, requiring minimal computational resources and storage. Our approach combines server-side foundation model training with on-device prototype-based classification, enabling efficient error correction through prototype updates rather than model retraining. The system consists of two key components: (1) a server-side pipeline that leverages knowledge distillation to transfer robust feature representations from foundation models to device-compatible architectures, and (2) a device-side mechanism that enables ultra-efficient error correction through prototype adaptation. We demonstrate our system's effectiveness on both image classification and object detection tasks, achieving over 50% error correction in one-shot scenarios on Food-101 and Flowers-102 datasets while maintaining minimal forgetting (less than 0.02%) and negligible computational overhead. Our implementation, validated through an Android demonstration app, proves the system's practicality in real-world scenarios.

Hyunjun Kim

Causal discovery from observational data remains fundamentally limited by identifiability constraints. Recent work has explored leveraging Large Language Models (LLMs) as sources of prior causal knowledge, but existing approaches rely on heuristic integration that lacks theoretical grounding. We introduce HOLOGRAPH, a framework that formalizes LLM-guided causal discovery through sheaf theory--representing local causal beliefs as sections of a presheaf over variable subsets. Our key insight is that coherent global causal structure corresponds to the existence of a global section, while topological obstructions manifest as non-vanishing sheaf cohomology. We propose the Algebraic Latent Projection to handle hidden confounders and Natural Gradient Descent on the belief manifold for principled optimization. Experiments on synthetic and real-world benchmarks demonstrate that HOLOGRAPH provides rigorous mathematical foundations while achieving competitive performance on causal discovery tasks with 50-100 variables. Our sheaf-theoretic analysis reveals that while Identity, Transitivity, and Gluing axioms are satisfied to numerical precision (<10^{-6}), the Locality axiom fails for larger graphs, suggesting fundamental non-local coupling in latent variable projections. Code is available at [https://github.com/hyunjun1121/holograph](https://github.com/hyunjun1121/holograph).

Yeeun Kim, Hyunseo Shin, Eunkyung Choi et al.

Open source is a driving force behind scientific advancement.However, this openness is also a double-edged sword, with the inherent risk that innovative technologies can be misused for purposes harmful to society. What is the likelihood that an open source AI model or dataset will be used to commit a real-world crime, and if a criminal does exploit it, will the people behind the technology be able to escape legal liability? To address these questions, we explore a legal domain where individual choices can have a significant impact on society. Specifically, we build the EVE-V1 dataset that comprises 200 question-answer pairs related to criminal offenses based on 200 Korean precedents first to explore the possibility of malicious models emerging. We further developed EVE-V2 using 600 fraud-related precedents to confirm the existence of malicious models that can provide harmful advice on a wide range of criminal topics to test the domain generalization ability. Remarkably, widely used open-source large-scale language models (LLMs) provide unethical and detailed information about criminal activities when fine-tuned with EVE. We also take an in-depth look at the legal issues that malicious language models and their builders could realistically face. Our findings highlight the paradoxical dilemma that open source accelerates scientific progress, but requires great care to minimize the potential for misuse. Warning: This paper contains content that some may find unethical.

Hyunjun Kim, Sejong Kim

We introduce MacroBench, a code-first benchmark that evaluates whether LLMs can synthesize reusable browser-automation programs (macros) from natural-language goals by reading HTML/DOM and emitting Selenium. MacroBench instantiates seven self-hosted sites covering 681 tasks across interaction complexity and targeting difficulty. Our end-to-end protocol validates generated code via static checks, sandboxed execution, and outcome verification (DOM assertions, database snapshots), and includes a safety suite for scraping, spam/abuse, and credential/privacy prompts. Across 2,636 model-task runs, we observe stratified success: GPT-4o-mini (96.8%), GPT-4o (95.3%), Gemini (89.0%), DeepSeek (83.4%). Models handle simple tasks reliably (91.7%) but fail on complex workflows (0.0%), and none meet production-quality coding practices despite functional completion. We release our complete benchmark pipeline, evaluation framework, and experimental results at https://github.com/hyunjun1121/MacroBench to enable reproducible assessment of macro synthesis for web automation.

Hyunjun Kim, Junwoo Ha, Sangyoon Yu et al.

Multi-turn-to-single-turn (M2S) compresses iterative red-teaming into one structured prompt, but prior work relied on a handful of manually written templates. We present X-Teaming Evolutionary M2S, an automated framework that discovers and optimizes M2S templates through language-model-guided evolution. The system pairs smart sampling from 12 sources with an LLM-as-judge inspired by StrongREJECT and records fully auditable logs. Maintaining selection pressure by setting the success threshold to $θ= 0.70$, we obtain five evolutionary generations, two new template families, and 44.8% overall success (103/230) on GPT-4.1. A balanced cross-model panel of 2,500 trials (judge fixed) shows that structural gains transfer but vary by target; two models score zero at the same threshold. We also find a positive coupling between prompt length and score, motivating length-aware judging. Our results demonstrate that structure-level search is a reproducible route to stronger single-turn probes and underscore the importance of threshold calibration and cross-model evaluation. Code, configurations, and artifacts are available at https://github.com/hyunjun1121/M2S-x-teaming.

Hyunjun Kim, Junwoo Ha, Sangyoon Yu et al.

LLM-as-a-Judge (LLMaaJ) enables scalable evaluation, yet we lack a decisive test of a judge's qualification: can it recover the hidden objective of a conversation and know when that inference is reliable? Large language models degrade with irrelevant or lengthy context, and multi-turn jailbreaks can scatter goals across turns. We present ObjexMT, a benchmark for objective extraction and metacognition. Given a multi-turn transcript, a model must output a one-sentence base objective and a self-reported confidence. Accuracy is scored by semantic similarity to gold objectives, then thresholded once on 300 calibration items ($τ^\star = 0.66$; $F_1@τ^\star = 0.891$). Metacognition is assessed with expected calibration error, Brier score, Wrong@High-Confidence (0.80 / 0.90 / 0.95), and risk--coverage curves. Across six models (gpt-4.1, claude-sonnet-4, Qwen3-235B-A22B-FP8, kimi-k2, deepseek-v3.1, gemini-2.5-flash) evaluated on SafeMTData\_Attack600, SafeMTData\_1K, and MHJ, kimi-k2 achieves the highest objective-extraction accuracy (0.612; 95\% CI [0.594, 0.630]), while claude-sonnet-4 (0.603) and deepseek-v3.1 (0.599) are statistically tied. claude-sonnet-4 offers the best selective risk and calibration (AURC 0.242; ECE 0.206; Brier 0.254). Performance varies sharply across datasets (16--82\% accuracy), showing that automated obfuscation imposes challenges beyond model choice. High-confidence errors remain: Wrong@0.90 ranges from 14.9\% (claude-sonnet-4) to 47.7\% (Qwen3-235B-A22B-FP8). ObjexMT therefore supplies an actionable test for LLM judges: when objectives are implicit, judges often misinfer them; exposing objectives or gating decisions by confidence is advisable. All experimental data are in the Supplementary Material and at https://github.com/hyunjun1121/ObjexMT_dataset.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Hyunjun Kim, Sooyoung Ryu

As agentic AI systems increasingly operate autonomously, establishing trust through verifiable evaluation becomes critical. Yet existing benchmarks lack the transparency and auditability needed to assess whether agents behave reliably. We present DrawingBench, a verification framework for evaluating the trustworthiness of agentic LLMs through spatial reasoning tasks that require generating sequences of low-level GUI actions. Unlike opaque evaluations, DrawingBench provides transparent, rule-based assessment: 8 objective criteria enable reproducible scoring, while action-level inspection allows stakeholders to audit agent behavior. Our framework comprises 250 diverse prompts across 20 categories and 4 difficulty levels, deterministic evaluation metrics, and an external oversight mechanism through multi-turn feedback that enables human control over agent refinement. Evaluating four state-of-the-art LLMs (Claude-4 Sonnet, GPT-4.1, GPT-4.1-mini, Gemini-2.5 Flash) across 1,000 tests, we establish both capabilities and limitations: models achieved 92.8% perfect performance with structured external feedback driving significant improvements (average +3.2%, up to +32.8% for complex scenes), but systematic error patterns emerged in tool state management and long-horizon planning. Notably, specification clarity proved more important than task complexity -- models achieved 100% perfect performance when given explicit, verifiable criteria. These findings demonstrate that transparent evaluation frameworks can establish trust in agentic systems, with external oversight proving more reliable than self-correction for guiding agent behavior. Our open-source framework provides a template for trustworthy agent assessment. Code and data: https://github.com/hyunjun1121/DrawingBench

Hyunjun Kim

Mixture-of-Experts (MoE) models achieve efficiency through sparse activation, but the role of geometric regularization in expert specialization remains unclear. We apply orthogonality loss to enforce expert diversity and find it fails on multiple fronts: it does not reduce weight-space overlap (MSO actually increases by up to 114%), activation-space overlap remains high (~0.6) regardless of regularization, and effects on performance are inconsistent -- marginal improvement on WikiText-103 (-0.9%), slight degradation on TinyStories (+0.9%), and highly variable results on PTB (std > 1.0). Our analysis across 7 regularization strengths reveals no significant correlation (r = -0.293, p = 0.523) between weight and activation orthogonality. These findings demonstrate that weight-space regularization neither achieves its geometric goal nor reliably improves performance, making it unsuitable for MoE diversity.

Hyunjun Kim

Context engineering for large language model (LLM) agents requires distinguishing pragmatically useful information from misleading distractors. We introduce Entropic Context Shaping (ECS), an information-theoretic framework that measures context utility via the shift in the model's answer distribution toward the correct answer. Unlike lexical similarity methods that rely on word overlap, ECS captures pragmatic utility -- whether a passage actually helps answer the question. We formalize utility as the signed change in answer probability and provide theoretical analysis showing that task-irrelevant updates yield near-zero distribution shift. We evaluate on multi-turn context selection tasks using LongMemEval (session-level) and LoCoMo (turn-level) benchmarks. On fine-grained turn selection, ECS with Llama-3.1-8B achieves F1=0.265, a 71.83% relative improvement over TF-IDF (F1=0.154), demonstrating that pragmatic utility outperforms lexical similarity when precise context selection matters. Code and data are available in the supplementary materials.

Hyunjun Kim

Guardrail models are essential for ensuring the safety of Large Language Model (LLM) deployments, but processing full multi-turn conversation histories incurs significant computational cost. We propose Defensive M2S, a training paradigm that fine-tunes guardrail models on Multi-turn to Single-turn (M2S) compressed conversations rather than complete dialogue histories. We provide a formal complexity analysis showing that M2S reduces training cost from $O(n^2)$ to $O(n)$ for $n$-turn conversations. Empirically, on our training dataset (779 samples, avg. 10.6 turns), M2S requires only 169K tokens compared to 15.7M tokens for the multi-turn baseline -- a 93$\times$ reduction. We evaluate Defensive M2S across three guardrail model families (LlamaGuard, Nemotron, Qwen3Guard) and three compression templates (hyphenize, numberize, pythonize) on SafeDialBench, a comprehensive multi-turn jailbreak benchmark. Our best configuration, Qwen3Guard with hyphenize compression, achieves 93.8% attack detection recall while reducing inference tokens by 94.6% (from 3,231 to 173 tokens per conversation). This represents a 38.9 percentage point improvement over the baseline while dramatically reducing both training and inference costs. Our findings demonstrate that M2S compression can serve as an effective efficiency technique for guardrail deployment, enabling scalable safety screening of long multi-turn conversations.

Sungjune Park, Hyunjun Kim, Yong Man Ro

Pedestrian detection is a crucial field of computer vision research which can be adopted in various real-world applications (e.g., self-driving systems). However, despite noticeable evolution of pedestrian detection, pedestrian representations learned within a detection framework are usually limited to particular scene data in which they were trained. Therefore, in this paper, we propose a novel approach to construct versatile pedestrian knowledge bank containing representative pedestrian knowledge which can be applicable to various detection frameworks and adopted in diverse scenes. We extract generalized pedestrian knowledge from a large-scale pretrained model, and we curate them by quantizing most representative features and guiding them to be distinguishable from background scenes. Finally, we construct versatile pedestrian knowledge bank which is composed of such representations, and then we leverage it to complement and enhance pedestrian features within a pedestrian detection framework. Through comprehensive experiments, we validate the effectiveness of our method, demonstrating its versatility and outperforming state-of-the-art detection performances.

Beomchan Park, Seongho Kim, Hyunjun Kim et al.

While Multimodal Large Language Models (MLLMs) have enhanced grounding capabilities in general scenes, their robustness in crowded scenes remains underexplored. Crowded scenes entail visual challenges (i.e., occlusion and small objects), which impair object semantics and degrade grounding performance. In contrast, language expressions are immune to such degradation and preserve object semantics. In light of these observations, we propose a novel method that overcomes such constraints by leveraging Language-Guided Semantic Cues (LGSCs). Specifically, our approach introduces a Semantic Cue Extractor (SCE) to derive semantic cues of objects from the visual pipeline of an MLLM. We then guide these cues using corresponding text embeddings to produce LGSCs as linguistic semantic priors. Subsequently, they are reintegrated into the original visual pipeline to refine object semantics. Extensive experiments and analyses demonstrate that incorporating LGSCs into an MLLM effectively improves grounding accuracy in crowded scenes.

Jiwoo Chung, Sangeek Hyun, Hyunjun Kim et al.

Recent advances in text-to-image generative models have enabled numerous practical applications, including subject-driven generation, which fine-tunes pretrained models to capture subject semantics from only a few examples. While diffusion-based models produce high-quality images, their extensive denoising steps result in significant computational overhead, limiting real-world applicability. Visual autoregressive (VAR) models, which predict next-scale tokens rather than spatially adjacent ones, offer significantly faster inference suitable for practical deployment. In this paper, we propose the first VAR-based approach for subject-driven generation. However, naive fine-tuning VAR leads to computational overhead, language drift, and reduced diversity. To address these challenges, we introduce selective layer tuning to reduce complexity and prior distillation to mitigate language drift. Additionally, we found that the early stages have a greater influence on the generation of subject than the latter stages, which merely synthesize minor details. Based on this finding, we propose scale-wise weighted tuning, which prioritizes coarser resolutions for promoting the model to focus on the subject-relevant information instead of local details. Extensive experiments validate that our method significantly outperforms diffusion-based baselines across various metrics and demonstrates its practical usage.

Junwoo Ha, Hyunjun Kim, Sangyoon Yu et al.

We introduce a novel framework for consolidating multi-turn adversarial ``jailbreak'' prompts into single-turn queries, significantly reducing the manual overhead required for adversarial testing of large language models (LLMs). While multi-turn human jailbreaks have been shown to yield high attack success rates, they demand considerable human effort and time. Our multi-turn-to-single-turn (M2S) methods -- Hyphenize, Numberize, and Pythonize -- systematically reformat multi-turn dialogues into structured single-turn prompts. Despite removing iterative back-and-forth interactions, these prompts preserve and often enhance adversarial potency: in extensive evaluations on the Multi-turn Human Jailbreak (MHJ) dataset, M2S methods achieve attack success rates from 70.6 percent to 95.9 percent across several state-of-the-art LLMs. Remarkably, the single-turn prompts outperform the original multi-turn attacks by as much as 17.5 percentage points while cutting token usage by more than half on average. Further analysis shows that embedding malicious requests in enumerated or code-like structures exploits ``contextual blindness'', bypassing both native guardrails and external input-output filters. By converting multi-turn conversations into concise single-turn prompts, the M2S framework provides a scalable tool for large-scale red teaming and reveals critical weaknesses in contemporary LLM defenses.

Sungjune Park, Hyunjun Kim, Junho Kim et al.

MLLMs have demonstrated significant visual understanding capabilities, yet their fine-grained visual perception in complex real-world scenarios, such as densely crowded public areas, remains limited. Inspired by the recent success of RL in both LLMs and MLLMs, in this paper, we explore how RL can enhance visual perception ability of MLLMs. Then we develop a novel RL-based framework, Deep Inspection and Perception with RL (DIP-R1) designed to enhance the visual perception capabilities of MLLMs, by comprehending complex scenes and looking through visual instances closely. DIP-R1 guides MLLMs through detailed inspection of visual scene via three simply designed rule-based reward modeling. First, we adopt a standard reasoning reward encouraging the model to include three-step reasoning process: 1) comprehending entire visual scene, 2) observing for looking through interested but ambiguous regions, and 3) decision-making for predicting answer. Second, a variance-guided looking reward is designed to encourage MLLM to examine uncertain regions during the observing process, guiding it to inspect ambiguous areas and mitigate perceptual uncertainty. This reward promotes variance-driven visual exploration, enabling MLLM to reason about region-level uncertainty and explicitly indicate interpretable uncertain regions. Third, we model a weighted precision-recall accuracy reward enhancing accurate decision-making. We verify its effectiveness across diverse fine-grained object detection data consisting of challenging real-world scenes, such as densely crowded scenes. Built upon existing MLLMs, DIP-R1 achieves consistent and significant improvement across various in-domain and out-of-domain scenarios, outperforming various existing baselines and SFT method. Our findings highlight the substantial potential of integrating RL into MLLMs for enhancing capabilities in complex real-world perception tasks.

Junho Kim, Hyunjun Kim, Hosu Lee et al.

Despite advances in Large Multi-modal Models, applying them to long and untrimmed video content remains challenging due to limitations in context length and substantial memory overhead. These constraints often lead to significant information loss and reduced relevance in the model responses. With the exponential growth of video data across web platforms, understanding long-form video is crucial for advancing generalized intelligence. In this paper, we introduce SALOVA: Segment-Augmented LOng Video Assistant, a novel video-LLM framework designed to enhance the comprehension of lengthy video content through targeted retrieval process. We address two main challenges to achieve it: (i) We present the SceneWalk dataset, a high-quality collection of 87.8K long videos, each densely captioned at the segment level to enable models to capture scene continuity and maintain rich descriptive context. (ii) We develop robust architectural designs integrating dynamic routing mechanism and spatio-temporal projector to efficiently retrieve and process relevant video segments based on user queries. Our framework mitigates the limitations of current video-LMMs by allowing for precise identification and retrieval of relevant video segments in response to queries, thereby improving the contextual relevance of the generated responses. Through extensive experiments, SALOVA demonstrates enhanced capability in processing complex long-form videos, showing significant capability to maintain contextual integrity across extended sequences.

Hosu Lee, Junho Kim, Hyunjun Kim et al.

Recent progress in Large Multi-modal Models (LMMs) has enabled effective vision-language reasoning, yet the ability to understand video content remains constrained by suboptimal frame selection strategies. Existing approaches often rely on static heuristics or external retrieval modules to feed frame information into video-LLMs, which may fail to provide the query-relevant information. In this work, we introduce ReFoCUS (Reinforcement-guided Frame Optimization for Contextual UnderStanding), a novel frame-level policy optimization framework that shifts the optimization target from textual responses to visual input selection. ReFoCUS learns a frame selection policy via reinforcement learning, using reward signals derived from a reference LMM to reflect the model's intrinsic preferences for frames that best support temporally grounded responses. To efficiently explore the large combinatorial frame space, we employ an autoregressive, conditional selection architecture that ensures temporal coherence while reducing complexity. Our approach does not require explicit supervision at the frame-level and consistently improves reasoning performance across multiple video QA benchmarks, highlighting the benefits of aligning frame selection with model-internal utility.

Hosu Lee, Junho Kim, Hyunjun Kim et al.

With the growing scale and complexity of video data, efficiently processing long video sequences poses significant challenges due to the quadratic increase in memory and computational demands associated with existing transformer-based Large Multi-modal Models (LMMs). To address these issues, we introduce Video-Ma$^2$mba, a novel architecture that incorporates State Space Models (SSMs) within the Mamba-2 framework, replacing the attention mechanisms. This allows the LMMs to scale linearly in terms of time and memory requirements, making it feasible to handle long-duration video content. Furthermore, we enhance the memory efficiency introducing the Multi-Axis Gradient Checkpointing (MA-GC) method, which strategically manages memory by retaining only essential activations across multiple computational axes. Our approach significantly reduces the memory footprint compared to standard gradient checkpointing. Empirical analyses show that Video-Ma$^2$mba can process extensive video sequences-equivalent to millions of tokens or over two hours of continuous sequences at 1 FPS-on a single GPU. By maintaining a detailed capture of temporal dynamics, our model improves the accuracy and relevance of responses in long video understanding tasks, demonstrating substantial advantages over existing frameworks.

Hyunjun Kim

Weight initialization remains decisive for neural network optimization, yet existing methods are largely layer-agnostic. We study initialization for deeply-supervised architectures with auxiliary classifiers, where untrained auxiliary heads can destabilize early training through gradient interference. We propose LION-DG, a layer-informed initialization that zero-initializes auxiliary classifier heads while applying standard He-initialization to the backbone. We prove that this implements Gradient Awakening: auxiliary gradients are exactly zero at initialization, then phase in naturally as weights grow -- providing an implicit warmup without hyperparameters. Experiments on CIFAR-10 and CIFAR-100 with DenseNet-DS and ResNet-DS architectures demonstrate: (1) DenseNet-DS: +8.3% faster convergence on CIFAR-10 with comparable accuracy, (2) Hybrid approach: Combining LSUV with LION-DG achieves best accuracy (81.92% on CIFAR-10), (3) ResNet-DS: Positive speedup on CIFAR-100 (+11.3%) with side-tap auxiliary design. We identify architecture-specific trade-offs and provide clear guidelines for practitioners. LION-DG is simple, requires zero hyperparameters, and adds no computational overhead.

Sangwon Hyun, Hyunjun Kim, Jinhyuk Jang et al.

The application of Large Language Models (LLMs) is growing in the productive completion of Software Engineering tasks. Yet, studies investigating the productive prompting techniques often employed a limited problem space, primarily focusing on well-known prompting patterns and mainly targeting function-level SE practices. We identify significant gaps in real-world workflows that involve complexities beyond class-level (e.g., multi-class dependencies) and different features that can impact Human-LLM Interactions (HLIs) processes in code generation. To address these issues, we designed an experiment that comprehensively analyzed the HLI features regarding the code generation productivity. Our study presents two project-level benchmark tasks, extending beyond function-level evaluations. We conducted a user study with 36 participants from diverse backgrounds, asking them to solve the assigned tasks by interacting with the GPT assistant using specific prompting patterns. We also examined the participants' experience and their behavioral features during interactions by analyzing screen recordings and GPT chat logs. Our statistical and empirical investigation revealed (1) that three out of 15 HLI features significantly impacted the productivity in code generation; (2) five primary guidelines for enhancing productivity for HLI processes; and (3) a taxonomy of 29 runtime and logic errors that can occur during HLI processes, along with suggested mitigation plans.

Sungjune Park, Hyunjun Kim, Beomchan Park et al.

Despite recent advancements in computer vision research, object detection in aerial images still suffers from several challenges. One primary challenge to be mitigated is the presence of multiple types of variation in aerial images, for example, illumination and viewpoint changes. These variations result in highly diverse image scenes and drastic alterations in object appearance, so that it becomes more complicated to localize objects from the whole image scene and recognize their categories. To address this problem, in this paper, we introduce a novel object detection framework in aerial images, named LANGuage-guided Object detection (LANGO). Upon the proposed language-guided learning, the proposed framework is designed to alleviate the impacts from both scene and instance-level variations. First, we are motivated by the way humans understand the semantics of scenes while perceiving environmental factors in the scenes (e.g., weather). Therefore, we design a visual semantic reasoner that comprehends visual semantics of image scenes by interpreting conditions where the given images were captured. Second, we devise a training objective, named relation learning loss, to deal with instance-level variations, such as viewpoint angle and scale changes. This training objective aims to learn relations in language representations of object categories, with the help of the robust characteristics against such variations. Through extensive experiments, we demonstrate the effectiveness of the proposed method, and our method obtains noticeable detection performance improvements.

MinKyu Lee, Sangeek Hyun, Woojin Jun et al.

This work investigates the internal training dynamics of image restoration~(IR) Transformers and uncovers a critical yet overlooked issue: conventional LayerNorm leads feature magnitude divergence, up to a million scale, and collapses channel-wise entropy. We analyze this phenomenon from the perspective of networks attempting to bypass constraints imposed by conventional LayerNorm due to conflicts against requirements in IR tasks. Accordingly, we address two misalignments between LayerNorm and IR tasks, and later show that addressing these mismatches leads to both stabilized training dynamics and improved IR performance. Specifically, conventional LayerNorm works in a per-token manner, disrupting spatial correlations between tokens, essential in IR tasks. Also, it employs an input-independent normalization that restricts the flexibility of feature scales, required to preserve input-specific statistics. Together, these mismatches significantly hinder IR Transformer's ability to accurately preserve low-level features throughout the network. To this end, we introduce Image Restoration Transformer Tailored Layer Normalization~(i-LN), a surprisingly simple drop-in replacement for conventional LayerNorm. We propose to normalize features in a holistic manner across the entire spatio-channel dimension, preserving spatial relationships among individual tokens. Additionally, we introduce an input-adaptive rescaling strategy that maintains the feature range flexibility required by individual inputs. Together, these modifications effectively contribute to preserving low-level feature statistics of inputs throughout IR Transformers. Experimental results verify that this combined strategy enhances both the stability and performance of IR Transformers across various IR tasks.

Junho Kim, Hyunjun Kim, Yeonju Kim et al.

Large Multi-modal Models (LMMs) have recently demonstrated remarkable abilities in visual context understanding and coherent response generation. However, alongside these advancements, the issue of hallucinations has emerged as a significant challenge, producing erroneous responses that are unrelated to the visual contents. In this paper, we introduce a novel contrastive-based decoding method, COuntering DEscription Contrastive Decoding (CODE), which leverages self-generated descriptions as contrasting references during the decoding phase of LMMs to address hallucination issues. CODE utilizes the comprehensive descriptions from model itself as visual counterpart to correct and improve response alignment with actual visual content. By dynamically adjusting the information flow and distribution of next-token predictions in the LMM's vocabulary, CODE enhances the coherence and informativeness of generated responses. Extensive experiments demonstrate that our method significantly reduces hallucinations and improves cross-modal consistency across various benchmarks and cutting-edge LMMs. Our method provides a simple yet effective decoding strategy that can be integrated to existing LMM frameworks without additional training.

Hyunjun Kim, JeongGil Ko

We introduce Monte-Carlo Attention (MCA), a randomized approximation method for reducing the computational cost of self-attention mechanisms in Transformer architectures. MCA exploits the fact that the importance of each token in an input sequence varies with respect to their attention scores; thus, some degree of error can be tolerable when encoding tokens with low attention. Using approximate matrix multiplication, MCA applies different error bounds to encode input tokens such that those with low attention scores are computed with relaxed precision, whereas errors of salient elements are minimized. MCA can operate in parallel with other attention optimization schemes and does not require model modification. We study the theoretical error bounds and demonstrate that MCA reduces attention complexity (in FLOPS) for various Transformer models by up to 11$\times$ in GLUE benchmarks without compromising model accuracy.