Figure 2: distributed calculation algorithmic Momentum Itiswell improv to Adam, where 1isthemomentumµin in Adaminit.xavier_normalin V methods, and AUR andissecond-best significantly and Adam Wealsosho7] in Figures itscalibration ImageNet, required Wealso different protocol 16,31,8,32] tocompare Wealsoborro16,30], sho reporting Ideally, we data.

A still from Humans in the Loop, a 2024 documentary that follows female data workers in Jharkhand state, India, whose labour underpins global AI systems. A still from Humans in the Loop, a 2024 documentary that follows female data workers in Jharkhand state, India, whose labour underpins global AI systems. 'In the end, you feel blank': India's female workers watching hours of abusive content to train AI Thu 5 Feb 2026 03.00 ESTLast modified on Thu 5 Feb 2026 03.03 EST On the veranda of her family's home, with her laptop balanced on a mud slab built into the wall, Monsumi Murmu works from one of the few places where the mobile signal holds. The familiar sounds of domestic life come from inside the house: clinking utensils, footsteps, voices. On her screen a very different scene plays: a woman is pinned down by a group of men, the camera shakes, there is shouting and the sound of breathing.

Keeping pace with the rapid growth of academia literature presents a significant challenge for researchers, funding bodies, and academic societies. To address the time-consuming manual effort required for scholarly discovery, we present a novel, fully automated system that transitions from data discovery to direct action. Our pipeline demonstrates how a specialized AI agent, 'Agent-E', can be tasked with identifying papers from specific geographic regions within conference proceedings and then executing a Robotic Process Automation (RPA) to complete a predefined action, such as submitting a nomination form. We validated our system on 586 papers from five different conferences, where it successfully identified every target paper with a recall of 100% and a near perfect accuracy of 99.4%. This demonstration highlights the potential of task-oriented AI agents to not only filter information but also to actively participate in and accelerate the workflows of the academic community.

Neural Information Processing Systems http://nips.cc/

Existing studies on fairness are largely Western-focused, making them inadequate for culturally diverse countries such as India. To address this gap, we introduce INDIC-BIAS, a comprehensive India-centric benchmark designed to evaluate fairness of LLMs across 85 identity groups encompassing diverse castes, religions, regions, and tribes. We first consult domain experts to curate over 1,800 socio-cultural topics spanning behaviors and situations, where biases and stereotypes are likely to emerge. Grounded in these topics, we generate and manually validate 20,000 real-world scenario templates to probe LLMs for fairness. We structure these templates into three evaluation tasks: plausibility, judgment, and generation. Our evaluation of 14 popular LLMs on these tasks reveals strong negative biases against marginalized identities, with models frequently reinforcing common stereotypes. Additionally, we find that models struggle to mitigate bias even when explicitly asked to rationalize their decision. Our evaluation provides evidence of both allocative and representational harms that current LLMs could cause towards Indian identities, calling for a more cautious usage in practical applications. We release INDIC-BIAS as an open-source benchmark to advance research on benchmarking and mitigating biases and stereotypes in the Indian context.

In recent times, we have seen a rapid development of large Vision-Language Models (VLMs). They have shown impressive results on academic benchmarks, primarily in widely spoken languages but lack performance on low-resource languages and varied cultural contexts. T o address these limitations, we introduce Maya, an open-source Multilingual VLM. Our contributions are: 1) a multilingual image-text pretraining dataset in eight languages, based on the LLaVA pretraining dataset; and 2) a multilingual image-text model supporting these languages, enhancing cultural and linguistic comprehension in vision-language tasks.

Abstract--Deep Neural Networks (DNNs) are increasingly deployed across diverse industries, driving a growing demand to enable their capabilities on mobile devices. However, existing mobile inference frameworks are often rely on a single processor to handle each model's inference, limiting hardware utilization and leading to suboptimal performance and energy efficiency . Expanding DNNs accessibility on mobile platforms requires more adaptive and resource-efficient solutions to meet increasing computational demands without compromising device functionality . Nevertheless, parallel inference of multiple DNNs on heterogeneous processors remains a significant challenge. Several works have explored partitioning DNN operations into subgraphs to enable parallel execution across heterogeneous processors. However, these approaches typically generate excessive subgraphs based solely on hardware compatibility, increasing scheduling complexity and memory management overhead. T o address these limitations, we propose an Advanced Multi-DNN Model Scheduling (ADMS) strategy that optimizes multi-DNN inference across heterogeneous processors on mobile devices. ADMS constructs an optimal subgraph partitioning strategy offline, considering both hardware support of operations and scheduling granularity, while employing a processor-state-aware scheduling algorithm that dynamically balances workloads based on real-time operational conditions. This ensures efficient workload distribution and maximizes the utilization of available processors. Experimental results show that, compared to vanilla inference frameworks, ADMS reduced multi-DNN inference latency by 4.04 T o reduce interaction latency and lower server-side computing costs, an increasing number of applications are shifting inference tasks to mobile devices. In many real-world scenarios, multiple independent or related DNN models run concurrently on mobile devices. For instance, in the smart agriculture scenario, farmers capture video frames using smartphone camera and perform real-time parallel inference with multiple DNN models. These models include crop identification [5], pest and disease detection [6], plant health assessment [7], and soil quality analysis [8]. Gao, X. Wang are with School of Computer Science and T echnology, Anhui Engineering Research Center for Intelligent Applications and Security of Industrial Internet, Anhui University of T echnology, Ma'anshan, Anhui, 243032, China.

LoRaWAN is a low-power long-range protocol that enables reliable and robust communication. This paper addresses the challenge of predicting the spreading factor (SF) in LoRaWAN networks using machine learning (ML) techniques. Optimal SF allocation is crucial for optimizing data transmission in IoT-enabled mobile devices, yet it remains a challenging task due to the fluctuation in environment and network conditions. We evaluated ML model performance across a large publicly available dataset to explore the best feature across key LoRaWAN features such as RSSI, SNR, frequency, distance between end devices and gateways, and antenna height of the end device, further, we also experimented with 31 different combinations possible for 5 features. We trained and evaluated the model using k-nearest neighbors (k-NN), Decision Tree Classifier (DTC), Random Forest (RF), and Multinomial Logistic Regression (MLR) algorithms. The combination of RSSI and SNR was identified as the best feature set. The finding of this paper provides valuable information for reducing the overall cost of dataset collection for ML model training and extending the battery life of LoRaWAN devices. This work contributes to a more reliable LoRaWAN system by understanding the importance of specific feature sets for optimized SF allocation.

Deep learning, when integrated with a large amount of training data, has the potential to outperform machine learning in terms of high accuracy. Recently, privacy-preserving deep learning has drawn significant attention of the research community. Different privacy notions in deep learning include privacy of data provided by data-owners and privacy of parameters and/or hyperparameters of the underlying neural network. Federated learning is a popular privacy-preserving execution environment where data-owners participate in learning the parameters collectively without leaking their respective data to other participants. However, federated learning suffers from certain security/privacy issues. In this paper, we propose Split-n-Chain, a variant of split learning where the layers of the network are split among several distributed nodes. Split-n-Chain achieves several privacy properties: data-owners need not share their training data with other nodes, and no nodes have access to the parameters and hyperparameters of the neural network (except that of the respective layers they hold). Moreover, Split-n-Chain uses blockchain to audit the computation done by different nodes. Our experimental results show that: Split-n-Chain is efficient, in terms of time required to execute different phases, and the training loss trend is similar to that for the same neural network when implemented in a monolithic fashion.

In modern large language models (LLMs), LLM alignment is of crucial importance and is typically achieved through methods such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO). However, in most existing methods for LLM alignment, all tokens in the response are optimized using a sparse, response-level reward or preference annotation. The ignorance of token-level rewards may erroneously punish high-quality tokens or encourage low-quality tokens, resulting in suboptimal performance and slow convergence speed. To address this issue, we propose AlignDistil, an RLHF-equivalent distillation method for token-level reward optimization. Specifically, we introduce the reward learned by DPO into the RLHF objective and theoretically prove the equivalence between this objective and a token-level distillation process, where the teacher distribution linearly combines the logits from the DPO model and a reference model. On this basis, we further bridge the accuracy gap between the reward from the DPO model and the pure reward model, by building a contrastive DPO reward with a normal and a reverse DPO model. Moreover, to avoid under- and over-optimization on different tokens, we design a token adaptive logit extrapolation mechanism to construct an appropriate teacher distribution for each token. Experimental results demonstrate the superiority of our AlignDistil over existing methods and showcase fast convergence due to its token-level distributional reward optimization.