论文检索

Roman Sultimov, Mikhail Mozikov, Dmitrii Abramov, Mariia Kovalchuk, Maksim Malykh, Aleksandr Volkov, Ilya Makarov, Andrei Osiptsov, Yury Maximov

Climate change is driving more frequent and severe disasters, putting people and infrastructure at risk. Protecting communities requires models that capture both natural disasters dynamics and how people behave under extreme conditions. This demo presents RESPOND, a multi-agent LLM-enhanced platform that jointly simulates natural hazards and human response. RESPOND couples high-fidelity flood AI forecasting an agent-based model of human behavior. LLM modules improve each agent decision-making, enabling context-aware reasoning over alerts, road closures, social signals, and changing water levels. The system simulates evacuation flows, resource seeking, and communication patterns producing actionable outputs for emergency management, urban planning, and policy. In the live demo one can run what-if or predicted scenarios, adjust assumptions, and observe emergent population behavior and risk hot spots in real time. By tightly coupling dynamic hazards with LLM-driven multi-agent behavior, RESPOND moves beyond fragmented tools and offers a practical, integrated platform for disaster preparedness and response.

Pirzada Suhail, Rehna Afroz Shaik, Gouranga Bala, Amit Sethi

Out-of-distribution (OOD) detection and uncertainty estimation (UE) are critical components for building safe machine learning systems. In this work, we propose a novel framework that combines network inversion with classifier training to simultaneously address both OOD detection and uncertainty estimation. We extend a standard n-class classifier by adding an (n+1)-th "garbage" class to capture outliers, initially populated with random Gaussian noise. After each training epoch, we use network inversion to reconstruct inputs for all classes; incoherent reconstructions are assigned to the garbage class for retraining. This iterative cycle of training, inversion, and exclusion continues until inverted samples resemble in-distribution data and uncertainty drops, indicating learned decision boundaries and cleaner class manifolds. At inference, the model detects OOD inputs by classifying them as garbage, with confidence scores estimating uncertainty. Unlike prior methods, this approach requires no external OOD data or post-hoc calibration, providing a simple, unified solution for robust classification, OOD detection, and uncertainty estimation.

Pirzada Suhail, Vrinda Goel, Amit Sethi

Shortcut learning, where machine learning models exploit spurious correlations in data instead of capturing meaningful features, poses a significant challenge to building generalizable models. Vision classifiers based on Convolutional Neural Networks (CNNs), Multi-Layer Perceptrons (MLPs), and Vision Transformers (ViTs) leverage distinct architectural principles to process spatial and structural information, making them differently susceptible to shortcut learning. In this study, we systematically evaluate these architectures by introducing deliberate shortcuts into the dataset that are correlated with class labels both positionally and via intensity, creating a controlled setup to assess whether models rely on these artificial cues or learn actual distinguishing features. We perform both quantitative evaluation by training on the shortcut-modified dataset and testing on two different test sets—one containing the same shortcuts and another without them—to determine the extent of reliance on shortcuts. Additionally, qualitative evaluation is performed using network inversion-based reconstruction techniques to analyze what the models internalize in their weights, aiming to reconstruct the training data as perceived by the classifiers. Further, we evaluate susceptibility to shortcut learning across different learning rates. Our analysis reveals that CNNs at lower learning rates tend to be more reserved against entirely picking up shortcut features, while ViTs, particularly those without positional encodings, almost entirely ignore the distinctive image features in the presence of shortcuts.

Minkyung Song, Soyoung Park, Sungsu Lim

Traditional recommenders often fail to disentangle the motivations behind user choices. To address this, we propose MV-LLMRec, a framework that models interactions through three views: Structural, Intent, and Conformity. MV-LLMRec leverages LLMs to generate rich semantic representations for intent and conformity, which are refined through graph propagation and dynamically fused via an attention mechanism. We evaluate MV-LLMRec on the Amazon-Movie and Amazon-Book datasets and show that it significantly outperforms state-of-the-art baselines, validating our approach.

Ravada Satyadev, Aditya Ganesh Kumar, Avinash Anand, Rajiv Ratn Shah, Zhengkui Wang, Mukesh Prasad

Bias in Large Language Models (LLMs) is increasingly addressed through fairness-oriented techniques. However, in some cases, these approaches may inadvertently remove genuine cultural differences between groups, leading to “over-normalization” or models losing important socio-cultural distinctions. In this work, we introduce OverNormEval, a benchmark designed to detect when an LLM exhibits such over-normalization. We further explore the use of Direct Preference Optimization (DPO) to mitigate over-normalization.

Kaede Saito, Ryota Higa, Hiromu Imura, Masaaki Kondo

The Pickup-and-Delivery Problem with Time Windows (PDPTW) is a time-constrained variant of the vehicle-routing problem (VRP). Complex time constraints make it difficult to solve using existing NCO methods. In this paper, we present the Feasibility-Aware Masked Transformer (FAM-Trans) specialized for PDPTW. FAM-Trans integrates a lightweight side encoder with a context-aware embedding scheme that effectively captures temporal dependencies. A dynamic key-value module continuously updates node embeddings as the route progresses. During inference, a feasibility-guided post-inference filtering strategy suppresses constraint violations without post-hoc repair. Experiments on standard PDPTW benchmarks show that FAM-Trans outperforms NCO baselines by 20~35% in solution quality and constraint satisfaction.

Rohan Saha, Tushar Shinde

Building temperature prediction is crucial for energy optimization and control in smart cities. We present a physics-enhanced XGBoost framework in a multi-stage sequential scaling approach. Starting from single-zone, single-day predictions, we progressively scale to multi-zone, multi-year forecasts using real-world data from Google's Smart Building Simulator. Our method incorporates physics-enhanced features, temporal encodings, and inter-zone interactions, achieving mean absolute errors (MAE) as low as 0.169°F for weekly multi-zone predictions. For longer horizons, we employ ensemble strategies, demonstrating robust performance up to 2.5 years. Compared to baseline models, our framework consistently improves long-term prediction fidelity. This work advances urban AI by enabling accurate long-term building dynamics modeling for downstream control tasks and bridges machine learning with physics-based modeling approaches.

Nilanjana Saha, Animesh Dutta

In social networks, revealing the structure of communities can expose sensitive groups to detection. Traditional approaches, such as DICE, attempt to hide these communities by randomly rewiring links, but this strategy is often inefficient and insecure. We propose an efficient heuristic method called CRIME (Community Rewiring for Influence and Masking Entities) to address this challenge. CRIME removes the most influential internal links, measured by edge-betweenness centrality, and adds external links with the least betweenness centrality. Experiments on real-world networks demonstrate that CRIME hides targeted communities more effectively than DICE, and also achieves faster execution and improves hiding effectiveness by up to 99.8%.

Alexander Sachuk, Vyacheslav Chukanov, Ekaterina Pchitskaya

The management and annotation of complex, multi-modal scientific data remains a major obstacle for AI-driven research due to poor reusability and scalability of current solutions. We propose SciDataMAS, a novel LLM-powered multi-agent system (MAS), which automate scientific data management through a structured data lake with provenance-based organization and an adaptive metadata taxonomy. The system uses specialized workflows for automated dataset creation, data insertion and retrieval. Experiments show the system's proficiency, with modern LLMs like GPT-5 successfully generating rich metadata schemas and filling them with high accuracy. This work provides a foundational step towards fully automated, reusable, and scalable scientific data organization which may lead to generation and accumulation by scientific community well annotated AI-ready datasets.

Jeong Hwan Ryu, Azimjon Akhtamov, Md Azher Uddin, Aziz Nasridinov

Ensuring proper use of personal protective equipment (PPE), especially helmets, is essential for workplace safety. Conventional object detectors often fail to distinguish whether a helmet is worn correctly, and existing approaches relying on single-model pipelines are prone to localization errors and false alarms. Moreover, most prior studies do not guarantee real-time performance. To resolve these challenges, we propose a lightweight multimodal approach that integrates a YOLO11-based object detector with a pose estimation model, achieving higher F1 scores and lower false alarm rates while maintaining real-time performance.

Tate Rowney, Xuning Ying

We identify a jailbreaking vulnerability in multiple open-source LLMs: by augmenting dangerous requests using certain "distractors" to obfuscate their intent, we elicit specific, actionable responses on a wide variety of harmful topics. We find that such an attack noticeably alters the contents of these models' chains of thought, including changed frequencies of seemingly unrelated n-grams and heightened ethical scrutiny about harmful requests even when their response is ultimately jailbroken.

Benjamin Richards, Pushpa Kumar Balan

Deep learning models are emerging as strong alternatives to numerical weather prediction, yet their internal representations remain poorly understood. We analyze the latent space of Microsoft’s Aurora model to test whether its embed- dings align with known physical processes. First, we show that land–sea distinctions are strongly captured, with errors mainly at coastlines. Second, we examine extreme surface temperatures using percentile-based thresholds, finding that embeddings reveal a gradient from moderate to severe events, though recall degrades at the rarest percentiles. These results suggest that Aurora’s encoder encodes physically consistent features but underestimates rare extremes. Our study combines deep learning forecasting, interpretable representation learning, and classical ML probing, illustrating how cross-disciplinary AI methods can yield insight into foundation models

Julia Rezvani, Alina Hyk, Thuyen Pham, Leonardo Marciaga, Chunyang Liao, Raffaele Vardavas, Konstantinos Mitsopoulos

Longitudinal surveys are a crucial component of behavioral research. Such surveys, however, face significant gaps in the data created by item and unit non-responses as well as semantic gaps resulting from questionnaires, assessed trends, and data collection methods evolving over time. Using 15 waves of vaccination surveys as a test-bed, we demonstrate how modern AI techniques can bridge both item and unit gaps, originating from non-response, and semantic gaps, originating from instrument evolution. We address these gaps through a two-component framework. We leverage LLM-generated semantic embeddings of survey questions to encode question meaning, enabling a Deep & Cross Network used for imputation to jointly model responses across item semantics, individual characteristics, and temporal dynamics. This structure directly addresses survey evolution by operating in learned semantic space. To overcome data scarcity, we use cluster-informed synthetic data generation via hierarchical prompting that produces synthetic responses preserving distributional properties and empirical cluster structure. Our approach achieves a strong improvement in semantic gap tasks and 80-90% synthetic data fidelity, providing practical solutions for evolving longitudinal studies.

Abdolazim Rezaei, Mehdi Sookhak

This study introduces PEFT-DML, a parameter-efficient deep metric learning framework for robust multi-modal 3D object detection in autonomous driving. Unlike conventional models that assume fixed sensor availability, PEFT-DML maps diverse modalities (LiDAR, radar, camera, IMU, GNSS) into a shared latent space, enabling reliable detection even under sensor dropout or unseen modality–class combinations. By integrating Low-Rank Adaptation (LoRA) and adapter layers, PEFT-DML achieves significant training efficiency while enhancing robustness to fast motion, weather variability, and domain shifts. Experiments on benchmarks nuScenes demonstrate superior accuracy.

Ye Qiao, Sitao Huang

Extending LLM context windows is key for long-range tasks. RoPE-based position interpolation (PI) scales input length without retraining, and post-training quantization (PTQ) enables efficient deployment; however, combining PI with PTQ degrades accuracy due to long-context aliasing, dynamic-range dilation, axis-grid anisotropy, and outlier shifts that induce position-dependent logit noise. We give the first systematic analysis of PI+PTQ and propose two diagnostics: Interpolation Pressure (per-band phase-scaling sensitivity) and Tail Inflation Ratio (outlier shift from short to long contexts). We then introduce Q-ROAR, a RoPE-aware, weight-only stabilization that bands RoPE dimensions and lightly searches per-band scales for W_Q,W_K, with an optional symmetric variant. Q-ROAR needs only a tiny long-context dev set and no fine-tuning or kernel changes, recovering up to 0.7% accuracy and more than 14% GovReport perplexity reduction while preserving short-context performance.

Suraj Prasad, Navyansh Mahla, Sunny Gupta, Amit Sethi

Existing federated prompt learning methods for vision-language models like CLIP rely solely on text-based prompts and final-layer visual features, missing crucial multiscale visual details and client-specific style variations. This limits generalization across non-IID distributions and novel classes. We introduce FedCSAP (Federated Cross-Modal Style-Aware Prompt Generation), which harnesses multiscale features from CLIP's vision encoder alongside domain-aware style statistics from client data. By fusing these visual representations with textual context, FedCSAP generates adaptive, context-aware prompts that enhance robustness across seen and unseen classes. Our privacy-preserving approach operates through local training and global aggregation, effectively handling heterogeneous client distributions. Experiments on multiple image classification datasets demonstrate that FedCSAP significantly outperforms existing federated prompt learning methods in both accuracy and generalization.

Ha-Hieu Pham, Minh Le, Han Huynh, Nguyen Quoc Khanh Le, Huy-Hieu Pham

Semi-supervised semantic segmentation (SSSS) is vital in computational pathology, where dense annotations are costly and limited. Existing methods often rely on pixel-level consistency, which propagates noisy pseudo-labels and produces fragmented or topologically invalid masks. We propose Topology Graph Consistency (TGC), a framework that integrates graph-theoretic constraints by aligning Laplacian spectra, component counts, and adjacency statistics between prediction graphs and references. This enforces global topology and improves segmentation accuracy. Experiments on GlaS and CRAG demonstrate that TGC achieves state-of-the-art performance under 5–10% supervision and significantly narrows the gap to full supervision.

Manali Patel, Krupa Jariwala, Chiranjoy Chattopadhyay

The dependency of stock prices on a multitude of factors makes the task of prediction exceedingly challenging. Given the volatile nature of stock data, it is imperative to integrate multiple sources of information to accurately encompass the various factors that influence market trends. To capture these complex dynamics, several multimodal methodologies have been proposed, integrating market data, technical indicators, and textual information. However, it is claimed that these coarse-grained information sources do not offer a holistic view of the market. Furthermore, these sources are stock-specific and do not elucidate the interconnections between various stocks. To address this deficiency, we propose a multimodal approach that incorporates this relational aspect alongside fine-grained information sources. The applicability of our framework is underscored by empirical results, which demonstrate the superiority of our approach.

Iñigo Parra, Mano Bharathi M, Mayank Kumar, Pushpa Kumar Balan, Priyadarsi Mishra

The computational cost of large language models (LLMs) is a primary obstacle to sustainable deployment. Static resource allocation is inefficient, as not all inputs require the same depth of processing. We propose a framework for adaptive, compute-efficient learning via conceptual criticality, which dynamically tailors computation to the assessed difficulty of an input. A lightweight criticality prediction module es- timates conceptual complexity on a continuous scale, and this score governs the LLM’s inference pathway, selectively activating token pruning, layer skipping, and quantization. Simple inputs are processed with minimal FLOPs and la- tency, while complex inputs use the model’s full capacity to preserve accuracy. We benchmark our framework and in- troduce metrics to quantify sensitivity to input criticality and per-sample computational savings. Results demonstrate an improved accuracy-efficiency trade-off, paving the way for more resource-aware systems.

Beomdo Park, Hyeonseok Jang, Junseong Park, Minu Baek, Gihun Gil, Minsung Jung, Woohyeon Kwon, Harin Jang, Yeojin Jang, Hoon Jeong 等

Quantum batteries have emerged as a next-generation energy storage solution, leveraging quantum phenomena such as superabsorption to overcome the limitations of conventional energy technologies. However, noise arising from interactions with the external environment degrades the charging efficiency and stability of the battery by disrupting the system's quantum coherence. To address this challenge, this study proposes a robust charging framework for a single-qubit quantum battery based on the Jaynes-Cummings (JC) model. The proposed framework combines the Proximal Policy Optimization (PPO) algorithm with a multi-stage reinforcement learning structure. The agent first learns fundamental control principles in a noise-free, ideal environment and subsequently performs robust learning in progressively noisier and more complex settings. Simulation results demonstrate that the trained agent navigates a stable charging trajectory on the Bloch sphere, thereby achieving high ergotropy even in the presence of noise. These findings suggest that multi-stage reinforcement learning is an effective solution for control problems in noisy quantum systems and provides a theoretical foundation for designing charging protocols for multi-qubit systems.