arXiv subjects

Sungeun Hong

Publications and source records attributed to Sungeun Hong.

At least 19 recordsLinked to original sources

arXiv2026

Deferred Audio Pruning with Local Audio-Visual Dynamics for Omni-LLMs

Omni-modal LLMs jointly process audio, video, and text, but long multimodal sequences incur substantial prefill and KV-cache costs. Existing omni-modal compression methods primarily focus on pre-LLM token reduction, leaving modality-specific compression across the LLM boundary underexplored. We propose A-PACK, a two-stage framework that defers audio pruning until query-conditioned multimodal interactions emerge. Our analysis shows that audio exhibits higher task-relevant information density and representational diversity per token than video. We further find that local audio-visual dynamics provide a more effective cue for visual selection than token-wise matching. We therefore preserve audio and compress video with local dynamics before the LLM, then progressively prune low-relevance audio and visual tokens and their KV-cache entries inside the LLM. Across four benchmarks on Qwen2.5-Omni-7B/3B, A-PACK achieves the strongest average performance among the evaluated prior methods while reducing prefill FLOPs by up to 78% and improving decoding throughput by up to 2.21x.

cs.AI↗

arXiv2026

CASHG: Context-Aware Stylized Online Handwriting Generation

Online handwriting represents strokes as time-ordered trajectories, which makes handwritten content easier to transform and reuse in a wide range of applications. However, generating natural sentence-level online handwriting that faithfully reflects a writer's style remains challenging, since sentence synthesis demands context-dependent characters with stroke continuity and spacing. Prior methods treat these boundary properties as implicit outcomes of sequence modeling, which becomes unreliable at the sentence scale and under limited compositional diversity. We propose CASHG, a context-aware stylized online handwriting generator that explicitly models inter-character connectivity for style-consistent sentence-level trajectory synthesis. CASHG uses a Character Context Encoder to obtain character identity and sentence-dependent context memory and fuses them in a bigram-aware sliding-window Transformer decoder that emphasizes local predecessor--current transitions, complemented by gated context fusion for sentence-level context.Training proceeds through a three-stage curriculum from isolated glyphs to full sentences, improving robustness under sparse transition coverage. We further introduce Connectivity and Spacing Metrics (CSM), a boundary-aware evaluation suite that quantifies cursive connectivity and spacing similarity. Under benchmark-matched evaluation protocols, CASHG consistently improves CSM over comparison methods while remaining competitive in DTW-based trajectory similarity, with gains corroborated by a human evaluation.

cs.CV↗

arXiv2025

Instance-Aware Test-Time Segmentation for Continual Domain Shifts

Continual Test-Time Adaptation (CTTA) enables pre-trained models to adapt to continuously evolving domains. Existing methods have improved robustness but typically rely on fixed or batch-level thresholds, which cannot account for varying difficulty across classes and instances. This limitation is especially problematic in semantic segmentation, where each image requires dense, multi-class predictions. We propose an approach that adaptively adjusts pseudo labels to reflect the confidence distribution within each image and dynamically balances learning toward classes most affected by domain shifts. This fine-grained, class- and instance-aware adaptation produces more reliable supervision and mitigates error accumulation throughout continual adaptation. Extensive experiments across eight CTTA and TTA scenarios, including synthetic-to-real and long-term shifts, show that our method consistently outperforms state-of-the-art techniques, setting a new standard for semantic segmentation under evolving conditions.

cs.CV↗

arXiv2025

Logic-Guided Socially-aware Robot Navigation World Model

Social robot navigation increasingly relies on large language models for reasoning, path planning, and enabling movement in dynamic human spaces. However, relying solely on LLMs for planning often leads to unpredictable and unsafe behaviors, especially in dynamic human spaces, due to limited physical grounding and weak logical consistency. In this work, we introduce NaviWM, a socially-aware robot Navigation World Model that augments LLM reasoning with a structured world model and a logic-driven chain-of-thought process. NaviWM consists of two main components: (1) a spatial-temporal world model that captures the positions, velocities, and activities of agents in the environment, and (2) a deductive reasoning module that guides LLMs through a multi-step, logic-based inference process. This integration enables the robot to generate navigation decisions that are both socially compliant and physically safe, under well-defined constraints such as personal space, collision avoidance, and timing. Unlike previous methods based on prompting or fine-tuning, NaviWM encodes social norms as first-order logic, enabling interpretable and verifiable reasoning. Experiments show that NaviWM improves success rates and reduces social violations, particularly in crowded environments. These results demonstrate the benefit of combining formal reasoning with LLMs for robust social navigation. Additional experimental details and demo videos for this work can be found at: https://sites.google.com/view/NaviWM.

cs.RO↗

arXiv2025

ZOO-Prune: Training-Free Token Pruning via Zeroth-Order Gradient Estimation in Vision-Language Models

Large Vision-Language Models (VLMs) enable strong multimodal reasoning but incur heavy inference costs from redundant visual tokens. Token pruning alleviates this issue, yet existing approaches face limitations. Attention-based methods rely on raw attention scores, which are often unstable across layers and heads and can lead to redundant selections. Diversity-based methods improve robustness by selecting tokens far apart in feature space, but risk dropping regions needed for accurate prediction. We propose ZOO-Prune, a training-free framework built on the intuition that highly sensitive tokens have a stronger influence on the model's output and capture complementary visual cues rather than redundant ones. To achieve this, we estimate token sensitivity using zeroth-order perturbations at the lightweight projection layer. This measures how small random perturbations affect the projected features and enables efficient approximation of each token's influence without backpropagation. Extensive experiments across multiple VLMs and benchmarks show that ZOO-Prune consistently outperforms prior methods while pruning up to 94.4% of tokens without sacrificing accuracy. Our method also improves efficiency, reaching up to 2.30x faster end-to-end inference compared to the baseline.

cs.CV↗

arXiv2025

PRIMT: Preference-based Reinforcement Learning with Multimodal Feedback and Trajectory Synthesis from Foundation Models

Preference-based reinforcement learning (PbRL) has emerged as a promising paradigm for teaching robots complex behaviors without reward engineering. However, its effectiveness is often limited by two critical challenges: the reliance on extensive human input and the inherent difficulties in resolving query ambiguity and credit assignment during reward learning. In this paper, we introduce PRIMT, a PbRL framework designed to overcome these challenges by leveraging foundation models (FMs) for multimodal synthetic feedback and trajectory synthesis. Unlike prior approaches that rely on single-modality FM evaluations, PRIMT employs a hierarchical neuro-symbolic fusion strategy, integrating the complementary strengths of large language models and vision-language models in evaluating robot behaviors for more reliable and comprehensive feedback. PRIMT also incorporates foresight trajectory generation, which reduces early-stage query ambiguity by warm-starting the trajectory buffer with bootstrapped samples, and hindsight trajectory augmentation, which enables counterfactual reasoning with a causal auxiliary loss to improve credit assignment. We evaluate PRIMT on 2 locomotion and 6 manipulation tasks on various benchmarks, demonstrating superior performance over FM-based and scripted baselines.

cs.RO↗

arXiv2025

Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment

Test-time adaptation (TTA) enhances the zero-shot robustness under distribution shifts by leveraging unlabeled test data during inference. Despite notable advances, several challenges still limit its broader applicability. First, most methods rely on backpropagation or iterative optimization, which limits scalability and hinders real-time deployment. Second, they lack explicit modeling of class-conditional feature distributions. This modeling is crucial for producing reliable decision boundaries and calibrated predictions, but it remains underexplored due to the lack of both source data and supervision at test time. In this paper, we propose ADAPT, an Advanced Distribution-Aware and backPropagation-free Test-time adaptation method. We reframe TTA as a Gaussian probabilistic inference task by modeling class-conditional likelihoods using gradually updated class means and a shared covariance matrix. This enables closed-form, training-free inference. To correct potential likelihood bias, we introduce lightweight regularization guided by CLIP priors and a historical knowledge bank. ADAPT requires no source data, no gradient updates, and no full access to target data, supporting both online and transductive settings. Extensive experiments across diverse benchmarks demonstrate that our method achieves state-of-the-art performance under a wide range of distribution shifts with superior scalability and robustness.

cs.CV↗

arXiv2025

Dynamic Rank Adjustment for Accurate and Efficient Neural Network Training

Low-rank training methods reduce the number of trainable parameters by re-parameterizing the weights with matrix decompositions (e.g., singular value decomposition). However, enforcing a fixed low-rank structure caps the rank of the weight matrices and can hinder the model's ability to learn complex patterns. Furthermore, the effective rank of the model's weights tends to decline during training, and this drop is accelerated when the model is reparameterized into a low-rank structure. In this study, we argue that strategically interleaving full-rank training epochs within low-rank training epochs can effectively restore the rank of the model's weights. Based on our findings, we propose a general dynamic-rank training framework that is readily applicable to a wide range of neural-network tasks. We first describe how to adjust the rank of weight matrix to alleviate the inevitable rank collapse that arises during training, and then present extensive empirical results that validate our claims and demonstrate the efficacy of the proposed framework. Our empirical study shows that the proposed method achieves almost the same computational cost as SVD-based low-rank training while achieving a comparable accuracy to full-rank training across various benchmarks.

cs.LG↗

arXiv2025

RA-Touch: Retrieval-Augmented Touch Understanding with Enriched Visual Data

Visuo-tactile perception aims to understand an object's tactile properties, such as texture, softness, and rigidity. However, the field remains underexplored because collecting tactile data is costly and labor-intensive. We observe that visually distinct objects can exhibit similar surface textures or material properties. For example, a leather sofa and a leather jacket have different appearances but share similar tactile properties. This implies that tactile understanding can be guided by material cues in visual data, even without direct tactile supervision. In this paper, we introduce RA-Touch, a retrieval-augmented framework that improves visuo-tactile perception by leveraging visual data enriched with tactile semantics. We carefully recaption a large-scale visual dataset with tactile-focused descriptions, enabling the model to access tactile semantics typically absent from conventional visual datasets. A key challenge remains in effectively utilizing these tactile-aware external descriptions. RA-Touch addresses this by retrieving visual-textual representations aligned with tactile inputs and integrating them to focus on relevant textural and material properties. By outperforming prior methods on the TVL benchmark, our method demonstrates the potential of retrieval-based visual reuse for tactile understanding. Code is available at https://aim-skku.github.io/RA-Touch

cs.CV↗

arXiv2025

Grasping Deformable Objects via Reinforcement Learning with Cross-Modal Attention to Visuo-Tactile Inputs

We consider the problem of grasping deformable objects with soft shells using a robotic gripper. Such objects have a center-of-mass that changes dynamically and are fragile so prone to burst. Thus, it is difficult for robots to generate appropriate control inputs not to drop or break the object while performing manipulation tasks. Multi-modal sensing data could help understand the grasping state through global information (e.g., shapes, pose) from visual data and local information around the contact (e.g., pressure) from tactile data. Although they have complementary information that can be beneficial to use together, fusing them is difficult owing to their different properties. We propose a method based on deep reinforcement learning (DRL) that generates control inputs of a simple gripper from visuo-tactile sensing information. Our method employs a cross-modal attention module in the encoder network and trains it in a self-supervised manner using the loss function of the RL agent. With the multi-modal fusion, the proposed method can learn the representation for the DRL agent from the visuo-tactile sensory data. The experimental result shows that cross-modal attention is effective to outperform other early and late data fusion methods across different environments including unseen robot motions and objects.

cs.RO↗

arXiv2025

Task Vector Quantization for Memory-Efficient Model Merging

Model merging enables efficient multi-task models by combining task-specific fine-tuned checkpoints. However, storing multiple task-specific checkpoints requires significant memory, limiting scalability and restricting model merging to larger models and diverse tasks. In this paper, we propose quantizing task vectors (i.e., the difference between pre-trained and fine-tuned checkpoints) instead of quantizing fine-tuned checkpoints. We observe that task vectors exhibit a narrow weight range, enabling low precision quantization (e.g., 4 bit) within existing task vector merging frameworks. To further mitigate quantization errors within ultra-low bit precision (e.g., 2 bit), we introduce Residual Task Vector Quantization, which decomposes the task vector into a base vector and offset component. We allocate bits based on quantization sensitivity, ensuring precision while minimizing error within a memory budget. Experiments on image classification and dense prediction show our method maintains or improves model merging performance while using only 8% of the memory required for full-precision checkpoints.

cs.LG↗

arXiv2025

Question-Aware Gaussian Experts for Audio-Visual Question Answering

Audio-Visual Question Answering (AVQA) requires not only question-based multimodal reasoning but also precise temporal grounding to capture subtle dynamics for accurate prediction. However, existing methods mainly use question information implicitly, limiting focus on question-specific details. Furthermore, most studies rely on uniform frame sampling, which can miss key question-relevant frames. Although recent Top-K frame selection methods aim to address this, their discrete nature still overlooks fine-grained temporal details. This paper proposes QA-TIGER, a novel framework that explicitly incorporates question information and models continuous temporal dynamics. Our key idea is to use Gaussian-based modeling to adaptively focus on both consecutive and non-consecutive frames based on the question, while explicitly injecting question information and applying progressive refinement. We leverage a Mixture of Experts (MoE) to flexibly implement multiple Gaussian models, activating temporal experts specifically tailored to the question. Extensive experiments on multiple AVQA benchmarks show that QA-TIGER consistently achieves state-of-the-art performance. Code is available at https://aim-skku.github.io/QA-TIGER/

cs.CV↗

arXiv2025

IAM: Enhancing RGB-D Instance Segmentation with New Benchmarks

Image segmentation is a vital task for providing human assistance and enhancing autonomy in our daily lives. In particular, RGB-D segmentation-leveraging both visual and depth cues-has attracted increasing attention as it promises richer scene understanding than RGB-only methods. However, most existing efforts have primarily focused on semantic segmentation and thus leave a critical gap. There is a relative scarcity of instance-level RGB-D segmentation datasets, which restricts current methods to broad category distinctions rather than fully capturing the fine-grained details required for recognizing individual objects. To bridge this gap, we introduce three RGB-D instance segmentation benchmarks, distinguished at the instance level. These datasets are versatile, supporting a wide range of applications from indoor navigation to robotic manipulation. In addition, we present an extensive evaluation of various baseline models on these benchmarks. This comprehensive analysis identifies both their strengths and shortcomings, guiding future work toward more robust, generalizable solutions. Finally, we propose a simple yet effective method for RGB-D data integration. Extensive evaluations affirm the effectiveness of our approach, offering a robust framework for advancing toward more nuanced scene understanding.

cs.CV↗

arXiv2024

SyMerge: From Non-Interference to Synergistic Merging via Single-Layer Adaptation

Model merging combines independently trained models into a single multi-task model. However, most existing approaches focus primarily on avoiding task interference. We argue that its greater potential lies in enabling task synergy, where tasks actively improve one another. We identify cross-task performance, defined by compatibility between encoders and predictors across tasks, as a key indicator of merge quality. We demonstrate that adapting only a single task-specific layer is sufficient to induce such synergy. This study proposes SyMerge, a lightweight framework that jointly optimizes merging coefficients and a single task-specific layer. We adopt an expert-guided self-labeling objective, providing stable supervision beyond entropy minimization. Intriguingly, we further show that SyMerge successfully merges models trained from different initializations, a regime where standard methods break down. Our minimalist yet principled method achieves state-of-the-art results across vision, dense prediction, and NLP benchmarks. Our code is available at https://aim-skku.github.io/SyMerge

cs.LG↗

arXiv2024

DisCoRD: Discrete Tokens to Continuous Motion via Rectified Flow Decoding

Human motion is inherently continuous and dynamic, posing significant challenges for generative models. While discrete generation methods are widely used, they suffer from limited expressiveness and frame-wise noise artifacts. In contrast, continuous approaches produce smoother, more natural motion but often struggle to adhere to conditioning signals due to high-dimensional complexity and limited training data. To resolve this 'discord' between discrete and continuous representations we introduce DisCoRD: Discrete Tokens to Continuous Motion via Rectified Flow Decoding, a novel method that leverages rectified flow to decode discrete motion tokens in the continuous, raw motion space. Our core idea is to frame token decoding as a conditional generation task, ensuring that DisCoRD captures fine-grained dynamics and achieves smoother, more natural motions. Compatible with any discrete-based framework, our method enhances naturalness without compromising faithfulness to the conditioning signals on diverse settings. Extensive evaluations demonstrate that DisCoRD achieves state-of-the-art performance, with FID of 0.032 on HumanML3D and 0.169 on KIT-ML. These results establish DisCoRD as a robust solution for bridging the divide between discrete efficiency and continuous realism. Project website: https://whwjdqls.github.io/discord-motion/

cs.CV↗

arXiv2024

Hyper-STTN: Hypergraph Augmented Spatial-Temporal Transformer Network for Trajectory Prediction

Predicting crowd intentions and trajectories is critical for a range of real-world applications, involving social robotics and autonomous driving. Accurately modeling such behavior remains challenging due to the complexity of pairwise spatial-temporal interactions and the heterogeneous influence of groupwise dynamics. To address these challenges, we propose Hyper-STTN, a Hypergraph-based Spatial-Temporal Transformer Network for crowd trajectory prediction. Hyper-STTN constructs multiscale hypergraphs of varying group sizes to model groupwise correlations, captured through spectral hypergraph convolution based on random-walk probabilities. In parallel, a spatial-temporal transformer is employed to learn pedestrians' pairwise latent interactions across spatial and temporal dimensions. These heterogeneous groupwise and pairwise features are subsequently fused and aligned via a multimodal transformer. Extensive experiments on public pedestrian motion datasets demonstrate that Hyper-STTN consistently outperforms state-of-the-art baselines and ablation models.

cs.CV↗

arXiv2022

Spatio-channel Attention Blocks for Cross-modal Crowd Counting

Crowd counting research has made significant advancements in real-world applications, but it remains a formidable challenge in cross-modal settings. Most existing methods rely solely on the optical features of RGB images, ignoring the feasibility of other modalities such as thermal and depth images. The inherently significant differences between the different modalities and the diversity of design choices for model architectures make cross-modal crowd counting more challenging. In this paper, we propose Cross-modal Spatio-Channel Attention (CSCA) blocks, which can be easily integrated into any modality-specific architecture. The CSCA blocks first spatially capture global functional correlations among multi-modality with less overhead through spatial-wise cross-modal attention. Cross-modal features with spatial attention are subsequently refined through adaptive channel-wise feature aggregation. In our experiments, the proposed block consistently shows significant performance improvement across various backbone networks, resulting in state-of-the-art results in RGB-T and RGB-D crowd counting.

cs.CV↗

arXiv2022

IFQA: Interpretable Face Quality Assessment

Existing face restoration models have relied on general assessment metrics that do not consider the characteristics of facial regions. Recent works have therefore assessed their methods using human studies, which is not scalable and involves significant effort. This paper proposes a novel face-centric metric based on an adversarial framework where a generator simulates face restoration and a discriminator assesses image quality. Specifically, our per-pixel discriminator enables interpretable evaluation that cannot be provided by traditional metrics. Moreover, our metric emphasizes facial primary regions considering that even minor changes to the eyes, nose, and mouth significantly affect human cognition. Our face-oriented metric consistently surpasses existing general or facial image quality assessment metrics by impressive margins. We demonstrate the generalizability of the proposed strategy in various architectural designs and challenging scenarios. Interestingly, we find that our IFQA can lead to performance improvement as an objective function.

cs.CV↗