arXiv subjects

Dongman Lee

Publications and source records attributed to Dongman Lee.

At least 19 recordsLinked to original sources

arXiv2026

SCoPE: Training-Free Audio-Visual Event Perception via Sparse Cross-Modal Prior Exchange

Audio-visual event perception (AVEP) determines which events occur in a video, when they occur, and whether they are audible, visible, or both. Training-free methods query new event vocabularies by matching frozen audio and visual features with text-encoded event names. However, related labels share evidence. An incorrect label can then score at least as high as a correct one. We call this a false co-activation (FCA). No scalar cutoff can reject the incorrect label while keeping every correct one. Class-specific thresholds may prevent that label from becoming a final prediction, but the FCA remains in the underlying score vector. We introduce SCoPE, a training-free framework in which all queried labels compete for shared evidence and each modality guides event selection in the other. We derive an exact condition for when this competition removes an FCA in a two-label fit. With identical frozen CLIP+CLAP backbones on LLP, SCoPE improves Type@seg by 7.45 points and Event@seg by 5.04 points compared with the reported AV$^2$A values. The same fixed configuration transfers unchanged to OV-AVEBench and VGGSound-AVEL100k.

cs.CV↗

arXiv2026

Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe

Silicon sampling uses language models as proxies for human survey respondents, treating each model call as an independent draw from the persona's response distribution. We show this draw does not exist: instruction-tuned models do not sample from distributions, they collapse to a single output. The same persona on the same question returns the same answer on more than half of items in a public-opinion benchmark, and the model's internal probabilities concentrate on a single option. The failure is associated with, and amplified by, instruction-targeted post-training: instruction-tuned models are worse than their own bases in every family we can compare, the gap widens at each successive post-training stage and with the size of the tuning update, and continued pretraining on non-instruction tokens leaves it unchanged. Yet the knowledge survives: the same model that cannot sample from a distribution can describe it accurately in a single call. We call this gap the KNOWS/DOES split. Exploiting the split, a single call that asks the model to describe the response distribution more than halves the error against human survey data compared to persona aggregation. When per-persona outputs are required, we propose Prompt-Perturbed Argyle (PPA), which reduces the same error by 21\%, spreading each persona's answers to mirror real population differences at no added cost.

cs.AI↗

arXiv2026

Last But Not Least: Boundary Attention CalibratiON for Multimodal KV Cache Compression

Multimodal Large Language Models (MLLMs) achieve strong vision-language reasoning but incur large KV caches and high decoding latency with long visual contexts. Existing compression methods rely on observation window attention for stable token importance estimation, yet this aggregation can dilute sparse critical evidence and discard answer-relevant tokens under aggressive compression. We identify last query attention as a complementary signal for recovering such evidence, though its irrelevant signals may introduce additional noise. We propose BACON, a plug-and-play method that calibrates observation window attention with last query evidence while suppressing noise through intra-layer coherence and inter-layer persistence. Across diverse benchmarks, models, budgets, and compression methods, BACON improves multimodal KV-cache compression by 7.5% on average under the most aggressive budget, with gains up to 30.9%.

cs.CV↗

arXiv2026

LACO: Adaptive Latent Communication for Collaborative Driving

Collaborative driving aims to improve safety and efficiency by enabling connected vehicles to coordinate under partial observability. Recent approaches have evolved from sharing visual features for perception to exchanging language-based reasoning through foundation models for behavioral coordination. Though communicating in language provides intuitive information, it introduces two challenges: high latency caused by autoregressive decoding and information loss caused by compressing rich internal representations into discrete tokens. To address these challenges, we analyze latent communication in collaborative driving under inherent limitations of multi-agent settings. Our analysis reveals agent identity confusion, where direct fusion of latent states entangles decision representations across vehicles. Motivated by this, we propose LACO, a training-free \textbf{LA}tent \textbf{CO}mmunication paradigm that seamlessly adapts pretrained driving models to collaborative settings. LACO introduces Iterative Latent Deliberation (ILD) for latent reasoning, Cross-Horizon Saliency Attribution (CHSA) for communication-efficient information selection, and Structured Semantic Knowledge Distillation (SSKD) to stabilize ego-centric decision making. Closed-loop experiments in CARLA show that LACO notably reduces communication and inference latency while maintaining strong collaborative driving performance.

cs.AI↗

arXiv2026

Delta Forcing: Trust Region Steering for Interactive Autoregressive Video Generation

Interactive real-time autoregressive video generation is essential for applications such as content creation and world modeling, where visual content must adapt to dynamically evolving event conditions. A fundamental challenge lies in balancing reactivity and stability: models must respond promptly to new events while maintaining temporal coherence over long horizons. Existing approaches distill bidirectional models into autoregressive generators and further adapt them via streaming long tuning, yet often exhibit persistent drift after condition changes. We identify the cause as conditional bias, where the teacher may provide condition-aligned but trajectory-agnostic guidance, biasing generation toward locally valid yet globally inconsistent modes. Inspired by Trust Region Policy Optimization, we propose Delta Forcing, a simple yet effective framework that constrains unreliable teacher supervision within an adaptive trust region. Specifically, Delta Forcing estimates transition consistency from the latent delta between teacher and generator trajectories, and uses it to balance teacher supervision with a monotonic continuity objective. This suppress unreliable teacher-induced shifts while preserving responsiveness to new events. Extensive experiments demonstrate that Delta Forcing significantly improves consistency while maintaining event reactivity.

cs.CV↗

arXiv2026

In-Context Examples Suppress Scientific Knowledge Recall in LLMs

Scientific reasoning rarely stops at what is directly observable; it often requires uncovering hidden structure from data. From estimating reaction constants in chemistry to inferring demand elasticities in economics, this latent structure recovery is what distinguishes scientific reasoning from curve fitting. Large language models (LLMs) can often recall and apply relevant scientific formulas, but we show that this ability is surprisingly easy to suppress. We show that adding in-context examples makes models rely less on pretrained domain knowledge, even when those examples are generated by the very same formula. Rather than reinforcing knowledge-driven derivation, examples shift computation toward empirical pattern fitting. We document this knowledge displacement on 60 latent structure recovery tasks across five scientific domains, 6,000 trials, and four models. This displacement is consistent across domains, but its accuracy consequences depend on how the displaced strategy compares to the one that replaces it: the same shift can lower accuracy, leave it unchanged, or appear to improve it. In all cases, however, the model shifts away from knowledge-driven reasoning. For practitioners deploying LLMs on scientific tasks, the message is cautionary: in-context examples may displace, rather than reinforce, the knowledge they are intended to support.

cs.AI↗

arXiv2026

Urban Vibrancy Embedding and Application on Traffic Prediction

Urban vibrancy reflects the dynamic human activity within urban spaces and is often measured using mobile data that captures floating population trends. This study proposes a novel approach to derive Urban Vibrancy embeddings from real-time floating population data to enhance traffic prediction models. Specifically, we utilize variational autoencoders (VAE) to compress this data into actionable embeddings, which are then integrated with long short-term memory (LSTM) networks to predict future embeddings. These are subsequently applied in a sequence-to-sequence framework for traffic forecasting. Our contributions are threefold: (1) We use principal component analysis (PCA) to interpret the embeddings, revealing temporal patterns such as weekday versus weekend distinctions and seasonal patterns; (2) We propose a method that combines VAE and LSTM, enabling forecasting dynamic urban knowledge embedding; and (3) Our approach improves accuracy and responsiveness in traffic prediction models, including RNN, DCRNN, GTS, and GMAN. This study demonstrates the potential of Urban Vibrancy embeddings to advance traffic prediction and offer a more nuanced analysis of urban mobility.

cs.LG↗

arXiv2025

DETACH : Decomposed Spatio-Temporal Alignment for Exocentric Video and Ambient Sensors with Staged Learning

Aligning egocentric video with wearable sensors have shown promise for human action recognition, but face practical limitations in user discomfort, privacy concerns, and scalability. We explore exocentric video with ambient sensors as a non-intrusive, scalable alternative. While prior egocentric-wearable works predominantly adopt Global Alignment by encoding entire sequences into unified representations, this approach fails in exocentric-ambient settings due to two problems: (P1) inability to capture local details such as subtle motions, and (P2) over-reliance on modality-invariant temporal patterns, causing misalignment between actions sharing similar temporal patterns with different spatio-semantic contexts. To resolve these problems, we propose DETACH, a decomposed spatio-temporal framework. This explicit decomposition preserves local details, while our novel sensor-spatial features discovered via online clustering provide semantic grounding for context-aware alignment. To align the decomposed features, our two-stage approach establishes spatial correspondence through mutual supervision, then performs temporal alignment via a spatial-temporal weighted contrastive loss that adaptively handles easy negatives, hard negatives, and false negatives. Comprehensive experiments with downstream tasks on Opportunity++ and HWU-USP datasets demonstrate substantial improvements over adapted egocentric-wearable baselines.

cs.CV↗

arXiv2025

Enhancing Regional Airbnb Trend Forecasting Using LLM-Based Embeddings of Accessibility and Human Mobility

The expansion of short-term rental platforms, such as Airbnb, has significantly disrupted local housing markets, often leading to increased rental prices and housing affordability issues. Accurately forecasting regional Airbnb market trends can thus offer critical insights for policymakers and urban planners aiming to mitigate these impacts. This study proposes a novel time-series forecasting framework to predict three key Airbnb indicators -- Revenue, Reservation Days, and Number of Reservations -- at the regional level. Using a sliding-window approach, the model forecasts trends 1 to 3 months ahead. Unlike prior studies that focus on individual listings at fixed time points, our approach constructs regional representations by integrating listing features with external contextual factors such as urban accessibility and human mobility. We convert structured tabular data into prompt-based inputs for a Large Language Model (LLM), producing comprehensive regional embeddings. These embeddings are then fed into advanced time-series models (RNN, LSTM, Transformer) to better capture complex spatio-temporal dynamics. Experiments on Seoul's Airbnb dataset show that our method reduces both average RMSE and MAE by approximately 48% compared to conventional baselines, including traditional statistical and machine learning models. Our framework not only improves forecasting accuracy but also offers practical insights for detecting oversupplied regions and supporting data-driven urban policy decisions.

cs.AI↗

arXiv2025

Background Fades, Foreground Leads: Curriculum-Guided Background Pruning for Efficient Foreground-Centric Collaborative Perception

Collaborative perception enhances the reliability and spatial coverage of autonomous vehicles by sharing complementary information across vehicles, offering a promising solution to long-tail scenarios that challenge single-vehicle perception. However, the bandwidth constraints of vehicular networks make transmitting the entire feature map impractical. Recent methods, therefore, adopt a foreground-centric paradigm, transmitting only predicted foreground-region features while discarding the background, which encodes essential context. We propose FadeLead, a foreground-centric framework that overcomes this limitation by learning to encapsulate background context into compact foreground features during training. At the core of our design is a curricular learning strategy that leverages background cues early on but progressively prunes them away, forcing the model to internalize context into foreground representations without transmitting background itself. Extensive experiments on both simulated and real-world benchmarks show that FadeLead outperforms prior methods under different bandwidth settings, underscoring the effectiveness of context-enriched foreground sharing.

cs.CV↗

arXiv2025

How2Compress: Scalable and Efficient Edge Video Analytics via Adaptive Granular Video Compression

With the rapid proliferation of the Internet of Things, video analytics has become a cornerstone application in wireless multimedia sensor networks. To support such applications under bandwidth constraints, learning-based adaptive quantization for video compression have demonstrated strong potential in reducing bitrate while maintaining analytical accuracy. However, existing frameworks often fail to fully exploit the fine-grained quality control enabled by modern blockbased video codecs, leaving significant compression efficiency untapped. In this paper, we present How2Compress, a simple yet effective framework designed to enhance video compression efficiency through precise, fine-grained quality control at the macroblock level. How2Compress is a plug-and-play module and can be seamlessly integrated into any existing edge video analytics pipelines. We implement How2Compress on the H.264 codec and evaluate its performance across diverse real-world scenarios. Experimental results show that How2Compress achieves up to $50.4\%$ bitrate savings and outperforms baselines by up to $3.01\times$ without compromising accuracy, demonstrating its practical effectiveness and efficiency. Code is available at https://github.com/wyhallenwu/how2compress and a reproducible docker image at https://hub.docker.com/r/wuyuheng/how2compress.

cs.MM↗

arXiv2025

FCPO: Federated Continual Policy Optimization for Real-Time High-Throughput Edge Video Analytics

The growing complexity of Edge Video Analytics (EVA) facilitates new kind of intelligent applications, but creates challenges in real-time inference serving systems. State-of-the-art (SOTA) scheduling systems optimize global workload distributions for heterogeneous devices but often suffer from extended scheduling cycles, leading to sub-optimal processing in rapidly changing Edge environments. Local Reinforcement Learning (RL) enables quick adjustments between cycles but faces scalability, knowledge integration, and adaptability issues. Thus, we propose FCPO, which combines Continual RL (CRL) with Federated RL (FRL) to address these challenges. This integration dynamically adjusts inference batch sizes, input resolutions, and multi-threading during pre- and post-processing. CRL allows agents to learn from changing Markov Decision Processes, capturing dynamic environmental variations, while FRL improves generalization and convergence speed by integrating experiences across inference models. FCPO combines these via an agent-specific aggregation scheme and a diversity-aware experience buffer. Experiments on a real-world EVA testbed showed over 5 times improvement in effective throughput, 60% reduced latency, and 20% faster convergence with up to 10 times less memory consumption compared to SOTA RL-based approaches.

cs.DC↗

arXiv2025

OCTOPINF: Workload-Aware Inference Serving for Edge Video Analytics

Edge Video Analytics (EVA) has gained significant attention as a major application of pervasive computing, enabling real-time visual processing. EVA pipelines, composed of deep neural networks (DNNs), typically demand efficient inference serving under stringent latency requirements, which is challenging due to the dynamic Edge environments (e.g., workload variability and network instability). Moreover, EVA pipelines also face significant resource contention caused by resource (e.g., GPU) constraints at the Edge. In this paper, we introduce OCTOPINF, a novel resource-efficient and workload-aware inference serving system designed for real-time EVA. OCTOPINF tackles the unique challenges of dynamic edge environments through fine-grained resource allocation, adaptive batching, and workload balancing between edge devices and servers. Furthermore, we propose a spatiotemporal scheduling algorithm that optimizes the co-location of inference tasks on GPUs, improving performance and ensuring service-level objectives (SLOs) compliance. Extensive evaluations on a real-world testbed demonstrate the effectiveness of our approach. It achieves an effective throughput increase of up to 10x compared to the baselines and shows better robustness in challenging scenarios. OCTOPINF can be used for any DNN-based EVA inference task with minimal adaptation and is available at https://github.com/tungngreen/PipelineScheduler.

cs.DC↗

arXiv2024

Feature Augmentation based Test-Time Adaptation

Test-time adaptation (TTA) allows a model to be adapted to an unseen domain without accessing the source data. Due to the nature of practical environments, TTA has a limited amount of data for adaptation. Recent TTA methods further restrict this by filtering input data for reliability, making the effective data size even smaller and limiting adaptation potential. To address this issue, We propose Feature Augmentation based Test-time Adaptation (FATA), a simple method that fully utilizes the limited amount of input data through feature augmentation. FATA employs Normalization Perturbation to augment features and adapts the model using the FATA loss, which makes the outputs of the augmented and original features similar. FATA is model-agnostic and can be seamlessly integrated into existing models without altering the model architecture. We demonstrate the effectiveness of FATA on various models and scenarios on ImageNet-C and Office-Home, validating its superiority in diverse real-world conditions.

cs.CV↗

arXiv2024

Spatio-Temporal Road Traffic Prediction using Real-time Regional Knowledge

For traffic prediction in transportation services such as car-sharing and ride-hailing, mid-term road traffic prediction (within a few hours) is considered essential. However, the existing road-level traffic prediction has mainly studied how significantly micro traffic events propagate to the adjacent roads in terms of short-term prediction. On the other hand, recent attempts have been made to incorporate regional knowledge such as POIs, road characteristics, and real-time social events to help traffic prediction. However, these studies lack in understandings of different modalities of road-level and region-level spatio-temporal correlations and how to combine such knowledge. This paper proposes a novel method that embeds real-time region-level knowledge using POIs, satellite images, and real-time LTE access traces via a regional spatio-temporal module that consists of dynamic convolution and temporal attention, and conducts bipartite spatial transform attention to convert into road-level knowledge. Then the model ingests this embedded knowledge into a road-level attention-based prediction model. Experimental results on real-world road traffic prediction show that our model outperforms the baselines.

cs.AI↗

arXiv2024

Multiple Areal Feature Aware Transportation Demand Prediction

A reliable short-term transportation demand prediction supports the authorities in improving the capability of systems by optimizing schedules, adjusting fleet sizes, and generating new transit networks. A handful of research efforts incorporate one or a few areal features while learning spatio-temporal correlation, to capture similar demand patterns between similar areas. However, urban characteristics are polymorphic, and they need to be understood by multiple areal features such as land use, sociodemographics, and place-of-interest (POI) distribution. In this paper, we propose a novel spatio-temporal multi-feature-aware graph convolutional recurrent network (ST-MFGCRN) that fuses multiple areal features during spatio-temproal understanding. Inside ST-MFGCRN, we devise sentinel attention to calculate the areal similarity matrix by allowing each area to take partial attention if the feature is not useful. We evaluate the proposed model on two real-world transportation datasets, one with our constructed BusDJ dataset and one with benchmark TaxiBJ. Results show that our model outperforms the state-of-the-art baselines up to 7\% on BusDJ and 8\% on TaxiBJ dataset.

cs.AI↗

arXiv2024

DOO-RE: A dataset of ambient sensors in a meeting room for activity recognition

With the advancement of IoT technology, recognizing user activities with machine learning methods is a promising way to provide various smart services to users. High-quality data with privacy protection is essential for deploying such services in the real world. Data streams from surrounding ambient sensors are well suited to the requirement. Existing ambient sensor datasets only support constrained private spaces and those for public spaces have yet to be explored despite growing interest in research on them. To meet this need, we build a dataset collected from a meeting room equipped with ambient sensors. The dataset, DOO-RE, includes data streams from various ambient sensor types such as Sound and Projector. Each sensor data stream is segmented into activity units and multiple annotators provide activity labels through a cross-validation annotation process to improve annotation quality. We finally obtain 9 types of activities. To our best knowledge, DOO-RE is the first dataset to support the recognition of both single and group activities in a real meeting room with reliable annotations.

cs.HC↗

arXiv2024

Self-supervised New Activity Detection in Sensor-based Smart Environments

With the rapid advancement of ubiquitous computing technology, human activity analysis based on time series data from a diverse range of sensors enables the delivery of more intelligent services. Despite the importance of exploring new activities in real-world scenarios, existing human activity recognition studies generally rely on predefined known activities and often overlook detecting new patterns (novelties) that have not been previously observed during training. Novelty detection in human activities becomes even more challenging due to (1) diversity of patterns within the same known activity, (2) shared patterns between known and new activities, and (3) differences in sensor properties of each activity dataset. We introduce CLAN, a two-tower model that leverages Contrastive Learning with diverse data Augmentation for New activity detection in sensor-based environments. CLAN simultaneously and explicitly utilizes multiple types of strongly shifted data as negative samples in contrastive learning, effectively learning invariant representations that adapt to various pattern variations within the same activity. To enhance the ability to distinguish between known and new activities that share common features, CLAN incorporates both time and frequency domains, enabling the learning of multi-faceted discriminative representations. Additionally, we design an automatic selection mechanism of data augmentation methods tailored to each dataset's properties, generating appropriate positive and negative pairs for contrastive learning. Comprehensive experiments on real-world datasets show that CLAN achieves a 9.24% improvement in AUROC compared to the best-performing baseline model.

cs.LG↗