Abstract

Curating high-quality, domain-specific datasets is a major bottleneck for deploying robust vision systems, requiring complex trade-offs between data quality, diversity, and cost when researching vast, unlabeled data lakes. We introduce Labeling Copilot, the first data curation deep research agent for computer vision. A central orchestrator agent, powered by a large multimodal language model, uses multi-step reasoning to execute specialized tools across three core capabilities: (1) Calibrated Discovery sources relevant, in-distribution data from large repositories; (2) Controllable Synthesis generates novel data for rare scenarios with robust filtering; and (3) Consensus Annotation produces accurate labels by orchestrating multiple foundation models via a novel consensus mechanism incorporating non-maximum suppression and voting. Our large-scale validation proves the effectiveness of Labeling Copilot's components. The Consensus Annotation module excels at object discovery: on the dense COCO dataset, it averages 14.2 candidate proposals per image-nearly double the 7.4 ground-truth objects-achieving a final annotation mAP of 37.1%. On the web-scale Open Images dataset, it navigated extreme class imbalance to discover 903 new bounding box categories, expanding its capability to over 1500 total. Concurrently, our Calibrated Discovery tool, tested at a 10-million sample scale, features an active learning strategy that is up to 40x more computationally efficient than alternatives with equivalent sample efficiency. These experiments validate that an agentic workflow with optimized, scalable tools provides a robust foundation for curating industrial-scale datasets.

Abstract

Excessive groundwater extraction has led to a significant decline in water tables, highlighting the need for continuous monitoring. This study presents an Internet of Things (IoT)-based approach for real-time estimation of groundwater levels in borewells. In addition to water level measurements, parameters such as motor current and rainfall are tracked to evaluate their influence on groundwater dynamics. To assess practical feasibility, five monitoring nodes were deployed across a small educational campus in Hyderabad, India. Data collected over a two-month period indicate that the proposed low-cost system is both reliable and effective in real-world conditions, while also revealing interesting correlations between groundwater level variations and the monitored parameters.

Abstract

Pretrained models are ubiquitous in the current deep learning landscape, offering strong results on a broad range of tasks. Recent works have shown that models differing in various design choices exhibit categorically diverse generalization behavior, resulting in one model grasping distinct data-specific insights unavailable to the other. In this paper, we propose to leverage large publicly available model repositories as an auxiliary source of model improvements. We introduce a data partitioning strategy where pretrained models autonomously adopt either the role of a student, seeking knowledge, or that of a teacher, imparting knowledge. Experiments across various tasks demonstrate the effectiveness of our proposed approach. In image classification, we improved the performance of ViT-B by approximately 1.4% through bidirectional knowledge transfer with ViT-T. For semantic segmentation, our method boosted all evaluation metrics by enabling knowledge transfer both within and across backbone architectures. In video saliency prediction, our approach achieved a new state-of-the-art. We further extend our approach to knowledge transfer between multiple models, leading to considerable performance improvements for all model participants.

Abstract

Residential electricity consumption datasets are essential for applications such as smart grid management, home automation, renewable energy integration, infrastructure planning, and policy-making. However, obtaining high-resolution residential datasets remains challenging due to the high costs and complexities of sensor installation, monitoring, and maintenance, obtaining approvals and related human factors, among other issues. To address this issue with a focus on the Indian residential context, where such data is quite limited, we propose a Residential Electricity Usage Simulator (REUS), to generate synthetic residential electricity usage data. Our approach models electricity usage for 7 different categories of homes using data collected over one year at an hourly interval, from 65 residences. In addition to energy data, we also collected 18 different features for each home to improve our modeling. The data and features are preprocessed using feature selection with Probabilistic Finite State Machines (validated through Multiple Correspondence Analysis) and further refined through systematic data cleaning and imputation. We built simulation models using the popular Machine learning techniques such as Long Short-Term Memory networks, including Vanilla, Stacked, BiDirectional, and Encoder-Decoder LSTMs and Transformer model, including Vanilla Transformer and Temporal Fusion Transformer. In addition, statistical techniques such as Markov Chains of orders (0, 1, 2, 3) and ARIMA were used as benchmarks to evaluate the models’ ability to generate a synthetic residential electricity dataset that is close to real data. Via extensive experimentation and analysis, our results show that (Bi-di) LSTMs capture the trends in electricity consumption more effectively (with the lowest RMSE) than the other models. Simulation and analysis of this nature enables broader, regionspecific energy research, reducing the need for costly or intrusive data collection.

Abstract

This study explored the cognitive status of community-dwelling Indian older adults. Our objective was to observe the association of age-related cognitive change with other physiological health parameters like, glycated hemoglobin (HbA1c), and vitamin B12 in older adults in India. Urban community dwelling, consenting older adults (55–85years, n = 123), with no clinical history of cognitive or neurological problems participated in the study. The participants underwent a detailed demographic documentation and cognitive assessment comprising of tests from different cognitive domains and blood-based assessment of glycated hemoglobin (HbA1c) and vitamin B12. As expected, performance in all cognitive domains declined with increasing age. HbA1c levels correlated inversely with processing speed and executive function. Vitamin B12 levels did not correlate with performance on any cognitive test. Interestingly, geriatric depression correlated inversely with visuospatial abilities. A stepwise multiple regression revealed that HbA1c and geriatric depression contributed to 28 % variance on Montreal Cognitive Assessment while age did not qualify as a significant contributor. Using Petersen’s criteria, Mild Cognitive Impairment (MCI) was observed in 17 % of participants. Participants classified as MCI had higher levels of HbA1c and geriatric depression, and lower performance in all cognitive domains compared to non-MCI participants. In conclusion, although cognitive performance declined with age, HbA1c and geriatric depression had a greater role in cognitive decline than age. With a high incidence of diabetes in India, this study highlights the prevalence of metabolism-linked changes in cognition, which are often ignored in community dwelling older adults in India.

Abstract

Amid the surge of Non-Fungible Tokens (NFTs) in blockchain, this study introduces a meticulous methodology focusing on transaction behaviors to unveil rug pulls — a critical issue impacting financial security and trust in the NFT landscape. Using a Graph Isomorphism Network (GIN) model with 6 behavioral patterns obtained from transaction sequences, we create a “Rug Pull Pattern Matcher” model. We provide a comprehensive analysis by applying the model on two datasets — creator’s transactions from 50 reputable NFT projects and 32 reported rug pulls. Our work utilizes automated labeling to categorize addresses and our analysis reveals several interconnected NFT creator activities. We present an in-depth mapping of fund flows and creator interactions exposing suspicious behaviors like artificial inflation and intricate network collaborations among creators. The results of our proposed model demonstrate the efficacy of our methodology with 75.4% accuracy and 85.9% precision on the dataset of reported rug pulls. This work provides comparative analyses of genuine and malicious creator networks to elucidate their structural differences, helping to identify genuine and potentially fraudulent NFT activities.

Abstract

Generating realistic full-body motion interacting with objects is critical for applications in robotics, virtual reality, and human-computer interaction. While existing methods can generate full-body motion within 3D scenes, they often lack the fidelity for fine-grained tasks like object grasping. Conversely, methods that generate precise grasping motions typically ignore the surrounding 3D scene. This gap, generating full-body grasping motions that are physically plausible within a 3D scene, remains a significant challenge. To address this, we introduce textbf{MOGRAS} (Human textbf{MO}tion with textbf{GRA}sping in 3D textbf{S}cenes), a large-scale dataset that bridges this gap. MOGRAS provides pre-grasping full-body walking motions and final grasping poses within richly annotated 3D indoor scenes. We leverage MOGRAS to benchmark existing full-body grasping methods and demonstrate their limitations in scene-aware generation. Furthermore, we propose a simple yet effective method to adapt existing approaches to work seamlessly within 3D scenes. Through extensive quantitative and qualitative experiments, we validate the effectiveness of our dataset and highlight the significant improvements our proposed method achieves, paving the way for more realistic human-scene interactions.

Abstract

The land use, land-use changes, and forestry (LULUCF) sector plays a significant role in increasing atmospheric carbon dioxide (CO2) in the current climate change scenario. The present study assessed the longterm CO2 emissions within the LULUCF sectors across India’s climatic zones as described by Köppen. The continuous combustion of fossil fuels has led to a rise in CO2 levels by 53 parts per million (ppm), increasing from 372 to 425 ppm in the Indian region between 2002 to 2022. The investigation reveals a consistent upward trend in CO2 levels across Köppen climatic zones. The semi-arid (steppe) climate zone shows an increasing trend of 2.18 ppm year−1, whereas Tundra exhibits a trend of 2.32 ppm year−1. The rate of change of atmospheric CO2 for India is 2.20 ppm year−1. A regional correlation exists between CO2 concentration and elevated anthropogenic emissions activities in the Indo-Gangetic Plains (IGP). Settlements growth has occurred across various climatic zones, with the Subtropical Highland zone showing the most significant rise at 54.84%. The findings underscore the urgent requirement for sustainable land use practices. This study utilized multivariate linear regression analysis to investigate different CO2 sources across LULUCF classes, using these as independent variables within Köppen climatic zones. The evaluation focused on identifying the classes with the highest and lowest contributions to CO2 emissions throughout the study period. Further training was carried out on different models for India as a whole and by climatic zone, revealing that the settlement class is a significant source of CO2 emission. In contrast, the forest class acts as a substantial sink for CO2.

Abstract

We show that the average of the maximum teleportation fidelities between all pairs of nodes in a large quantum repeater network is a measure of the resourcefulness of the network as a whole. It can be used to rank networks of arbitrary topologies, with and without loops. For demonstration purposes, we use simple Werner state-based models to characterize some fundamental topologies (the star, the chain, some trees, and also the ring and the complete graph) with this measure in three (semi)realistic scenarios. Most of our results are analytic and apply to arbitrary network sizes. We identify the parameter ranges where these networks can achieve quantum advantages and show the large-N behaviors.

Abstract

With the growing adoption of machine learning models in critical domains, techniques for explaining differences between models have become essential for trust, debugging, and informed deployment. Previous approaches address this by identifying input transformations that cause divergent predictions or by learning joint surrogate models to align and contrast behaviors. These methods often require access to training data and do not produce natural language explanations. In this paper, we introduce SLED, a framework that generates faithful natural language explanations of when and how two ML models converge or diverge in their predictions. SLED first uses gradient-based optimization to synthesize input samples that highlight divergence and convergence patterns, and then leverages a large language model (LLM) to generate explanations grounded in these synthetic samples. Across both text-based (3 tasks, 7 models) and structured (10 tasks, 4 models) classification tasks, we show that SLED explanations are 18--24% more faithful than the strongest baselines. User studies also indicate that SLED explanations achieve a real-world simulatability of 63.5%. Importantly, SLED requires minimal access to training data and generalizes well to real-world samples, enabling transparent and data-efficient model comparison.