arXiv subjects

Sung-Ju Lee

Publications and source records attributed to Sung-Ju Lee.

At least 19 recordsLinked to original sources

arXiv2026

MedMix: Specialization-Consistent Federated Sparse MoEs under Modality Heterogeneity

Federated multimodal medical AI faces modality heterogeneity at both the client and sample levels: clients may systematically lack access to specific modality types, while individual records within the same client may contain different partial modality subsets. Sparse Mixture-of-Experts (MoE) architectures are a promising remedy for modality-adaptive computation, but their use in federated learning is fragile under cross-client modality heterogeneity, where locally learned routing policies can diverge across clients and drive experts toward incompatible specializations. Different clients may assign the same observed modality configuration to different experts, or train similarly indexed experts on different missing-modality configurations, causing standard aggregation to misalign or overwrite the expert specialization that sparse MoEs are intended to learn. To address this challenge, we propose MedMix, a semantic-alignment framework for federated multimodal sparse MoEs that coordinates cross-client routing and expert specialization using modality context. At the client side, MedMix uses modality-context-aware routing to guide expert selection using each token's modality identity, position, and incompleteness context. Across clients, it uses consensus-guided routing alignment to construct server-side consensus anchors for shared modality patterns and align local routing distributions across clients. Complementing these routing mechanisms, client-adaptive expert aggregation leverages client-specific modality-pattern prototypes to match and aggregate functionally similar experts across clients. Experiments on real-world multimodal medical datasets show that MedMix achieves the best average F1 across diverse modality heterogeneity and modality incompleteness settings, with especially clear gains under severe heterogeneity.

cs.LG↗

arXiv2026

Multimodal Federated Learning under Dual-Axis Modality Missingness

Multimodal federated learning (FL) supports collaborative modeling in privacy-sensitive health-sensing and medical settings, but realistic deployments often exhibit dual-axis modality missingness: clients have different modality sets, and individual samples may contain only subsets of the modalities available locally. Existing methods typically address these two axes separately. We propose Flux, a multimodal federated learning framework built around two complementary components. First, modality-aware confidence tempering learns sample-specific confidence for each modality through mask-aware unimodal supervision and fuses the confidence estimates from observed modalities into a sample-adaptive temperature that adjusts predictive sharpness according to evidence quality and completeness. Second, gradient-decoupled private adaptation applies this temperature only to a client-private prediction pathway, while training the shared federated model with a standard, untempered objective. This enables sample-specific, client-local confidence adaptation without allowing confidence-dependent gradients to perturb shared representation learning. Across four multimodal datasets, Flux achieves the highest average macro-F1 on every dataset, outperforming the strongest dataset-specific baseline by 0.8~2.2 points and by 1.6 points on average. Additional analyses demonstrate favorable calibration, temperature sensitivity to both modality missingness and input corruption, and more stable shared optimization under private-only tempering. Our code is available at https://github.com/AdibaOrz/Flux.

cs.LG↗

arXiv2026

Beyond Hearing: Learning Task-Agnostic ExG Representations from Earphones via Physiology-Informed Tokenization

Electrophysiological (ExG) signals offer valuable insights into human physiology, yet building foundation models that generalize across everyday tasks remains challenging due to two key limitations: (i)~insufficient data diversity, as most ExG recordings are collected in controlled labs with bulky, expensive devices; and (ii)~task-specific model designs that require tailored processing (i.e., targeted frequency filters) and architectures, which limit generalization across tasks. To address these challenges, we introduce an approach for scalable, task-agnostic ExG monitoring in the wild. We collected 50 hours of unobtrusive free-living ExG data with an earphone-based hardware prototype to narrow the data diversity gap. At the core of our approach is Physiology-informed Multi-band Tokenization (PiMT), which decomposes ExG signals into 12 physiology-informed tokens, followed by a reconstruction task to learn robust representations. This enables adaptive feature recognition across the full frequency spectrum while capturing task-relevant information. Experiments on our new DailySense dataset, the first to enable ExG-based analysis across five human senses, together with four public ExG benchmarks, demonstrate that PiMT consistently outperforms state-of-the-art methods across diverse tasks.

eess.AS↗

arXiv2026

ConSensus: Multi-Agent Collaboration for Multimodal Sensing

Large language models (LLMs) are increasingly grounded in sensor data to perceive and reason about human physiology and the physical world. However, accurately interpreting heterogeneous multimodal sensor data remains a fundamental challenge. We show that a single monolithic LLM often fails to reason coherently across modalities, leading to incomplete interpretations and prior-knowledge bias. We introduce ConSensus, a training-free multi-agent collaboration framework that decomposes multimodal sensing tasks into specialized, modality-aware agents. To aggregate agent-level interpretations, we propose a hybrid fusion mechanism that balances semantic aggregation, which enables cross-modal reasoning and contextual understanding, with statistical consensus, which provides robustness through agreement across modalities. While each approach has complementary failure modes, their combination enables reliable inference under sensor noise and missing data. We evaluate ConSensus on five diverse multimodal sensing benchmarks, demonstrating an average accuracy improvement of 7.1% over the single-agent baseline. Furthermore, ConSensus matches or exceeds the performance of iterative multi-agent debate methods while achieving a 12.7 times reduction in average fusion token cost through a single-round hybrid fusion protocol, yielding a robust and efficient solution for real-world multimodal sensing tasks. The source code is available at https://github.com/nokia/multi-agent-collaboration-for-multimodal-sensing.

cs.AI↗

arXiv2026

Wearable Foundation Models Should Go Beyond Static Encoders

Wearable foundation models (WFMs), trained on large volumes of data collected by affordable, always-on devices, have demonstrated strong performance on short-term, well-defined health monitoring tasks, including activity recognition, fitness tracking, and cardiovascular signal assessment. However, most existing WFMs primarily map short temporal windows to predefined labels via static encoders, emphasizing retrospective prediction rather than reasoning over evolving personal history, context, and future risk trajectories. As a result, they are poorly suited for modeling chronic, progressive, or episodic health conditions that unfold over weeks, months or years. Hence, we argue that WFMs must move beyond static encoders and be explicitly designed for longitudinal, anticipatory health reasoning. We identify three foundational shifts required to enable this transition: (1) Structurally rich data, which goes beyond isolated datasets or outcome-conditioned collection to integrated multimodal, long-term personal trajectories, and contextual metadata, ideally supported by open and interoperable data ecosystems; (2) Longitudinal-aware multimodal modeling, which prioritizes long-context inference, temporal abstraction, and personalization over cross-sectional or population-level prediction; and (3) Agentic inference systems, which move beyond static prediction to support planning, decision-making, and clinically grounded intervention under uncertainty. Together, these shifts reframe wearable health monitoring from retrospective signal interpretation toward continuous, anticipatory, and human-aligned health support.

cs.LG↗

arXiv2026

Evaluating Visual Prompts with Eye-Tracking Data for MLLM-Based Human Activity Recognition

Large Language Models (LLMs) have emerged as foundation models for IoT applications such as human activity recognition (HAR). However, directly applying high-frequency and multi-dimensional sensor data, such as eye-tracking data, leads to information loss and high token costs. To mitigate this, we investigate a visual prompting strategy that transforms sensor signals into data visualization images as an input to multimodal LLMs (MLLMs) using eye-tracking data. We conducted a systematic evaluation of MLLM-based HAR across three public eye-tracking datasets using three visualization types of timeline, heatmap, and scanpath, under varying temporal window sizes. Our findings suggest that visual prompting provides a token-efficient and scalable representation for eye-tracking data, highlighting its potential to enable MLLMs to effectively reason over high-frequency sensor signals in IoT contexts.

cs.HC↗

arXiv2026

A Design Space for Live Music Agents

Live music provides a uniquely rich setting for studying creativity and interaction due to its spontaneous nature. The pursuit of live music agents--intelligent systems supporting real-time music performance and interaction--has captivated researchers across HCI, AI, and computer music for decades, and recent advancements in AI suggest unprecedented opportunities to evolve their design. However, the interdisciplinary nature of music has led to fragmented development across research communities, hindering effective communication and collaborative progress. In this work, we bring together perspectives from these diverse fields to map the current landscape of live music agents. Based on our analysis of 184 systems across both academic literature and video, we develop a comprehensive design space that categorizes dimensions spanning usage contexts, interactions, technologies, and ecosystems. By highlighting trends and gaps in live music agents, our design space offers researchers, designers, and musicians a structured lens to understand existing systems and shape future directions in real-time human-AI music co-creation. We release our annotated systems as a living artifact at https://live-music-agents.github.io.

cs.HC↗

arXiv2025

SNAP: Low-Latency Test-Time Adaptation with Sparse Updates

Test-Time Adaptation (TTA) adjusts models using unlabeled test data to handle dynamic distribution shifts. However, existing methods rely on frequent adaptation and high computational cost, making them unsuitable for resource-constrained edge environments. To address this, we propose SNAP, a sparse TTA framework that reduces adaptation frequency and data usage while preserving accuracy. SNAP maintains competitive accuracy even when adapting based on only 1% of the incoming data stream, demonstrating its robustness under infrequent updates. Our method introduces two key components: (i) Class and Domain Representative Memory (CnDRM), which identifies and stores a small set of samples that are representative of both class and domain characteristics to support efficient adaptation with limited data; and (ii) Inference-only Batch-aware Memory Normalization (IoBMN), which dynamically adjusts normalization statistics at inference time by leveraging these representative samples, enabling efficient alignment to shifting target domains. Integrated with five state-of-the-art TTA algorithms, SNAP reduces latency by up to 93.12%, while keeping the accuracy drop below 3.3%, even across adaptation rates ranging from 1% to 50%. This demonstrates its strong potential for practical use on edge devices serving latency-sensitive applications. The source code is available at https://github.com/chahh9808/SNAP.

cs.LG↗

arXiv2025

Collective Voice: Recovered-Peer Support Mediated by An LLM-Based Chatbot for Eating Disorder Recovery

Peer recovery narratives provide unique benefits beyond professional or lay mentoring by fostering hope and sustained recovery in eating disorder (ED) contexts. Yet, such support is limited by the scarcity of peer-involved programs and potential drawbacks on recovered peers, including relapse risk. To address this, we designed RecoveryTeller, a chatbot adopting a recovered-peer persona that portrays itself as someone recovered from an ED. We examined whether such a persona can reproduce the support affordances of peer recovery narratives. We compared RecoveryTeller with a lay-mentor persona chatbot offering similar guidance but without a recovery background. We conducted a 20-day cross-over deployment study with 26 ED participants, each using both chatbots for 10 days. RecoveryTeller elicited stronger emotional resonance than a lay-mentor chatbot, yet tensions between emotional and epistemic trust led participants to view the two personas as complementary rather than substitutes. We provide design implications for mental health chatbot persona design.

cs.HC↗

arXiv2025

Impact Ambivalence: How People with Eating Disorders Get Trapped in the Perpetual Cycle of Digital Food Content Engagement

Digital food content could impact viewers' dietary health, with individuals with eating disorders being particularly sensitive to it. However, a comprehensive understanding of why and how these individuals interact with such content is lacking. To fill this void, we conducted exploratory (N=23) and in-depth studies (N=22) with individuals with eating disorders to understand their motivations and practices of consuming digital food content. We reveal that participants engaged with digital food content for both disorder-driven and recovery-supporting motivations, leading to conflicting outcomes. This impact ambivalence, the coexistence of recovery-supporting benefits and disorder-exacerbating risks, sustained a cycle of quitting, prompted by awareness of harm, and returning, motivated by anticipated benefits. We interpret these dynamics within dual systems theory and highlight how recognizing such ambivalence can inform the design of interventions that foster healthier digital food content engagement and mitigate post-engagement harmful effects.

cs.HC↗

arXiv2025

HateBuffer: Safeguarding Content Moderators' Mental Well-Being through Hate Speech Content Modification

Hate speech remains a persistent and unresolved challenge in online platforms. Content moderators, working on the front lines to review user-generated content and shield viewers from hate speech, often find themselves unprotected from the mental burden as they continuously engage with offensive language. To safeguard moderators' mental well-being, we designed HateBuffer, which anonymizes targets of hate speech, paraphrases offensive expressions into less offensive forms, and shows the original expressions when moderators opt to see them. Our user study with 80 participants consisted of a simulated hate speech moderation task set on a fictional news platform, followed by semi-structured interviews. Although participants rated the hate severity of comments lower while using HateBuffer, contrary to our expectations, they did not experience improved emotion or reduced fatigue compared with the control group. In interviews, however, participants described HateBuffer as an effective buffer against emotional contagion and the normalization of biased opinions in hate speech. Notably, HateBuffer did not compromise moderation accuracy and even contributed to a slight increase in recall. We explore possible explanations for the discrepancy between the perceived benefits of HateBuffer and its measured impact on mental well-being. We also underscore the promise of text-based content modification techniques as tools for a healthier content moderation environment.

cs.HC↗

arXiv2025

QuRe: Query-Relevant Retrieval through Hard Negative Sampling in Composed Image Retrieval

Composed Image Retrieval (CIR) retrieves relevant images based on a reference image and accompanying text describing desired modifications. However, existing CIR methods only focus on retrieving the target image and disregard the relevance of other images. This limitation arises because most methods employing contrastive learning-which treats the target image as positive and all other images in the batch as negatives-can inadvertently include false negatives. This may result in retrieving irrelevant images, reducing user satisfaction even when the target image is retrieved. To address this issue, we propose Query-Relevant Retrieval through Hard Negative Sampling (QuRe), which optimizes a reward model objective to reduce false negatives. Additionally, we introduce a hard negative sampling strategy that selects images positioned between two steep drops in relevance scores following the target image, to effectively filter false negatives. In order to evaluate CIR models on their alignment with human satisfaction, we create Human-Preference FashionIQ (HP-FashionIQ), a new dataset that explicitly captures user preferences beyond target retrieval. Extensive experiments demonstrate that QuRe achieves state-of-the-art performance on FashionIQ and CIRR datasets while exhibiting the strongest alignment with human preferences on the HP-FashionIQ dataset. The source code is available at https://github.com/jackwaky/QuRe.

cs.CV↗

arXiv2025

Test-Time Adaptation with Binary Feedback

Deep learning models perform poorly when domain shifts exist between training and test data. Test-time adaptation (TTA) is a paradigm to mitigate this issue by adapting pre-trained models using only unlabeled test samples. However, existing TTA methods can fail under severe domain shifts, while recent active TTA approaches requiring full-class labels are impractical due to high labeling costs. To address this issue, we introduce a new setting of TTA with binary feedback. This setting uses a few binary feedback inputs from annotators to indicate whether model predictions are correct, thereby significantly reducing the labeling burden of annotators. Under the setting, we propose BiTTA, a novel dual-path optimization framework that leverages reinforcement learning to balance binary feedback-guided adaptation on uncertain samples with agreement-based self-adaptation on confident predictions. Experiments show BiTTA achieves 13.3%p accuracy improvements over state-of-the-art baselines, demonstrating its effectiveness in handling severe distribution shifts with minimal labeling effort. The source code is available at https://github.com/taeckyung/BiTTA.

cs.LG↗

arXiv2025

Design Opportunities for Explainable AI Paraphrasing Tools: A User Study with Non-native English Speakers

We investigate how non-native English speakers (NNESs) interact with diverse information aids to assess and select AI-generated paraphrases. We develop ParaScope, an AI paraphrasing assistant that integrates diverse information aids, such as back-translation, explanations, and usage examples, and logs user interaction data. Our in-lab study with 22 NNESs reveals that user preferences for information aids vary by language proficiency, with workflows progressing from global to more detailed information. While back-translation was the most frequently used aid, it was not a decisive factor in suggestion acceptance; users combined multiple information aids to make informed decisions. Our findings demonstrate the potential of explainable AI paraphrasing tools to enhance NNESs' confidence, autonomy, and writing efficiency, while also emphasizing the importance of thoughtful design to prevent information overload. Based on these findings, we offer design implications for explainable AI paraphrasing tools that support NNESs in making informed decisions when using AI writing systems.

cs.HC↗

arXiv2025

Automated Attack Synthesis for Constant Product Market Makers

Decentralized Finance (DeFi) enables many novel applications that were impossible in traditional finances. However, it also introduces new types of vulnerabilities. An example of such vulnerabilities is a composability bug between token contracts and Decentralized Exchange (DEX) that follows the Constant Product Market Maker (CPMM) model. This type of bug, which we refer to as CPMM composability bug, originates from issues in token contracts that make them incompatible with CPMMs, thereby endangering other tokens within the CPMM ecosystem. Since 2022, 23 exploits of such kind have resulted in a total loss of 2.2M USD. BlockSec, a smart contract auditing company, reported that 138 exploits of such kind occurred just in February 2023. In this paper, we propose CPMMX , a tool that automatically detects CPMM composability bugs across entire blockchains. To achieve such scalability, we first formalized CPMM composability bugs and found that these bugs can be induced by breaking two safety invariants. Based on this finding, we designed CPMMX equipped with a two-step approach, called shallow-then-deep search. In more detail, it first uses shallow search to find transactions that break the invariants. Then, it uses deep search to refine these transactions, making them profitable for the attacker. We evaluated CPMMX against five baselines on two public datasets and one synthetic dataset. In our evaluation, CPMMX detected 2.5x to 1.5x more vulnerabilities compared to baseline methods. It also analyzed contracts significantly faster, achieving higher F1 scores than the baselines. Additionally, we applied CPMMX to all contracts on the latest blocks of the Ethereum and Binance networks and discovered 26 new exploits that can result in 15.7K USD profit in total.

cs.CR↗

arXiv2025

SelfReplay: Adapting Self-Supervised Sensory Models via Adaptive Meta-Task Replay

Self-supervised learning has emerged as a method for utilizing massive unlabeled data for pre-training models, providing an effective feature extractor for various mobile sensing applications. However, when deployed to end-users, these models encounter significant domain shifts attributed to user diversity. We investigate the performance degradation that occurs when self-supervised models are fine-tuned in heterogeneous domains. To address the issue, we propose SelfReplay, a few-shot domain adaptation framework for personalizing self-supervised models. SelfReplay proposes self-supervised meta-learning for initial model pre-training, followed by a user-side model adaptation by replaying the self-supervision with user-specific data. This allows models to adjust their pre-trained representations to the user with only a few samples. Evaluation with four benchmarks demonstrates that SelfReplay outperforms existing baselines by an average F1-score of 8.8%p. Our on-device computational overhead analysis on a commodity off-the-shelf (COTS) smartphone shows that SelfReplay completes adaptation within an unobtrusive latency (in three minutes) with only a 9.54% memory consumption, demonstrating the computational efficiency of the proposed method.

eess.SP↗

arXiv2025

Amuse: Human-AI Collaborative Songwriting with Multimodal Inspirations

Songwriting is often driven by multimodal inspirations, such as imagery, narratives, or existing music, yet songwriters remain unsupported by current music AI systems in incorporating these multimodal inputs into their creative processes. We introduce Amuse, a songwriting assistant that transforms multimodal (image, text, or audio) inputs into chord progressions that can be seamlessly incorporated into songwriters' creative processes. A key feature of Amuse is its novel method for generating coherent chords that are relevant to music keywords in the absence of datasets with paired examples of multimodal inputs and chords. Specifically, we propose a method that leverages multimodal large language models (LLMs) to convert multimodal inputs into noisy chord suggestions and uses a unimodal chord model to filter the suggestions. A user study with songwriters shows that Amuse effectively supports transforming multimodal ideas into coherent musical suggestions, enhancing users' agency and creativity throughout the songwriting process.

cs.HC↗

arXiv2025

SoundCollage: Automated Discovery of New Classes in Audio Datasets

Developing new machine learning applications often requires the collection of new datasets. However, existing datasets may already contain relevant information to train models for new purposes. We propose SoundCollage: a framework to discover new classes within audio datasets by incorporating (1) an audio pre-processing pipeline to decompose different sounds in audio samples, and (2) an automated model-based annotation mechanism to identify the discovered classes. Furthermore, we introduce the clarity measure to assess the coherence of the discovered classes for better training new downstream applications. Our evaluations show that the accuracy of downstream audio classifiers within discovered class samples and a held-out dataset improves over the baseline by up to 34.7% and 4.5%, respectively. These results highlight the potential of SoundCollage in making datasets reusable by labeling with newly discovered classes. To encourage further research in this area, we open-source our code at https://github.com/nokia-bell-labs/audio-class-discovery.

cs.SD↗