Millimeter-wave radar offers a privacy-preserving and environment-robust alternative to vision-based sensing, enabling human motion analysis in challenging conditions such as low light, occlusions, rain, or smoke. However, its sparse point clouds pose significant challenges for semantic understanding. We present RadarLLM, the first framework that leverages large language models (LLMs) for human motion understanding from radar signals. RadarLLM introduces two key innovations: (1) a motion-guided radar tokenizer based on our Aggregate VQ-VAE architecture, integrating deformable body templates and masked trajectory modeling to convert spatial-temporal radar sequences into compact semantic tokens; and (2) a radar-aware language model that establishes cross-modal alignment between radar and text in a shared embedding space. To overcome the scarcity of paired radar-text data, we generate a realistic radar-text dataset from motion-text datasets with a physics-aware synthesis pipeline. Extensive experiments on both synthetic and real-world benchmarks show that RadarLLM achieves state-of-the-art performance, enabling robust and interpretable motion understanding under privacy and visibility constraints, even in adverse environments. This paper has been accepted for presentation at AAAI 2026. This is an extended version with supplementary materials.

Large Language Models (LLMs) have demonstrated the ability to tackle increasingly complex tasks through advanced reasoning, long-form content generation, and tool use. Solving these tasks often involves long inference-time computations. In human problem solving, a common strategy to expedite work is collaboration: by dividing the problem into sub-tasks, exploring different strategies concurrently, etc. Recent research has shown that LLMs can also operate in parallel by implementing explicit cooperation frameworks, such as voting mechanisms or the explicit creation of independent sub-tasks that can be executed in parallel. However, each of these frameworks may not be suitable for all types of tasks, which can hinder their applicability. In this work, we propose a different design approach: we run LLM "workers" in parallel , allowing them to synchronize via a concurrently-updated attention cache and prompt these workers to decide how best to collaborate. Our approach allows the LLM instances to come up with their own collaboration strategy for the problem at hand, all the while "seeing" each other's memory in the concurrent KV cache. We implement this approach via Hogwild! Inference: a parallel LLM inference engine where multiple instances of the same LLM run in parallel with the same attention cache, with "instant" access to each other's memory. Hogwild! Inference takes advantage of Rotary Position Embeddings (RoPE) to avoid recomputation while improving parallel hardware utilization. We find that modern reasoning-capable LLMs can perform inference with shared Key-Value cache out of the box, without additional fine-tuning.

In the landscape of Fact-based Judgment Prediction and Explanation (FJPE), reliance on factual data is essential for developing robust and realistic AI-driven decision-making tools. This paper introduces TathyaNyaya, the largest annotated dataset for FJPE tailored to the Indian legal context, encompassing judgments from the Supreme Court of India and various High Courts. Derived from the Hindi terms "Tathya" (fact) and "Nyaya" (justice), the TathyaNyaya dataset is uniquely designed to focus on factual statements rather than complete legal texts, reflecting real-world judicial processes where factual data drives outcomes. Complementing this dataset, we present FactLegalLlama, an instruction-tuned variant of the LLaMa-3-8B Large Language Model (LLM), optimized for generating high-quality explanations in FJPE tasks. Finetuned on the factual data in TathyaNyaya, FactLegalLlama integrates predictive accuracy with coherent, contextually relevant explanations, addressing the critical need for transparency and interpretability in AI-assisted legal systems. Our methodology combines transformers for binary judgment prediction with FactLegalLlama for explanation generation, creating a robust framework for advancing FJPE in the Indian legal domain. TathyaNyaya not only surpasses existing datasets in scale and diversity but also establishes a benchmark for building explainable AI systems in legal analysis. The findings underscore the importance of factual precision and domain-specific tuning in enhancing predictive performance and interpretability, positioning TathyaNyaya and FactLegalLlama as foundational resources for AI-assisted legal decision-making.

Memory-augmented neural networks (MANNs) can perform algorithmic tasks such as sorting. However, they often fail to generalise to input sequence lengths not encountered during training. We introduce two approaches that constrain the state space of the MANN's controller network: state compression and state regularisation. We empirically demonstrated that both approaches can improve generalisation to input sequences of out-of-distribution lengths for a specific type of MANN: the differentiable neural computer (DNC). The constrained DNC could process input sequences that were up to 2.3 times longer than those processed by an unconstrained baseline controller network. Notably, the applied constraints enabled the extension of the DNC's memory matrix without the need for retraining and thus allowed the processing of input sequences that were 10.4 times longer. However, the improvements were not consistent across all tested algorithmic tasks. Interestingly, solutions that performed better often had a highly structured state space, characterised by state trajectories exhibiting increased curvature and loop-like patterns. Our experimental work demonstrates that state-space constraints can enable the training of a DNC using shorter input sequences, thereby saving computational resources and facilitating training when acquiring long sequences is costly.

Abstract--High-quality personalized question banks are crucial for supporting adaptive learning and individualized assessment. Manually designing questions is time-consuming and often fails to meet diverse learning needs, making automated question generation a crucial approach to reduce teachers' workload and improve the scalability of educational resources. However, most existing question generation methods rely on single-agent or rule-based pipelines, which still produce questions with unstable quality, limited diversity, and insufficient alignment with educational goals. T o address these challenges, we propose EduAgentQG, a multi-agent collaborative framework for generating high-quality and diverse personalized questions. The framework consists of five specialized agents and operates through an iterative feedback loop: the Planner generates structured design plans and multiple question directions to enhance diversity; the Writer produces candidate questions based on the plan and optimizes their quality and diversity using feedback from the Solver and Educator; the Solver and Educator perform binary scoring across multiple evaluation dimensions and feed the evaluation results back to the Writer; the Checker conducts final verification, including answer correctness and clarity, ensuring alignment with educational goals. Through this multi-agent collaboration and iterative feedback loop, EduAgentQG generates questions that are both high-quality and diverse, while maintaining consistency with educational objectives. Experiments on two mathematics question datasets demonstrate that EduAgentQG outperforms existing single-agent and multi-agent methods in terms of question diversity, goal consistency, and overall quality. High-quality personalized question banks are crucial for supporting adaptive learning and individualized assessment [1], [2], [3]. In practical teaching, experienced educators can often determine the specific educational goals a student needs to achieve based on observation and prior knowledge [4], [5], [6]. Teachers typically engage in iterative cycles of planning, drafting, validation, and optimization to design questions that are both diagnostically effective and pedagogically meaningful, balancing knowledge coverage, cognitive skill development, and difficulty levels [7], [8]. Existing question banks may not always contain suitable questions, and even when relevant questions are available, they may have been previously attempted by students [9], [10], [11].

Cloud-based large language models (LLMs) and their variants have significantly influenced real-world applications. Deploying smaller models (i.e., small language models (SLMs)) on edge devices offers additional advantages, such as reduced latency and independence from network connectivity. However, edge devices' limited computing resources and constrained energy budgets challenge efficient deployment. This study evaluates the power efficiency of five representative SLMs - Llama 3.2, Phi-3 Mini, TinyLlama, and Gemma 2 on Raspberry Pi 5, Jetson Nano, and Jetson Orin Nano (CPU and GPU configurations). Results show that Jetson Orin Nano with GPU acceleration achieves the highest energy-to-performance ratio, significantly outperforming CPU-based setups. Llama 3.2 provides the best balance of accuracy and power efficiency, while TinyLlama is well-suited for low-power environments at the cost of reduced accuracy. In contrast, Phi-3 Mini consumes the most energy despite its high accuracy. In addition, GPU acceleration, memory bandwidth, and model architecture are key in optimizing inference energy efficiency. Our empirical analysis offers practical insights for AI, smart systems, and mobile ad-hoc platforms to leverage tradeoffs from accuracy, inference latency, and power efficiency in energy-constrained environments.

The rise of Large Language Models (LLM) has increased the need for scalable, high-performance inference systems, yet most existing frameworks assume homogeneous, resource-rich hardware, often unrealistic in academic, or resource-constrained settings. We introduce AIvailable, a low-cost, highly available LLM-as-a-Service (LLMaaS) platform, that uses a software-defined approach for running LLMs across heterogeneous and legacy GPU nodes, including NVIDIA and AMD devices, with a focus on fully utilizing each node's VRAM. AIvailable operates as a fully GPU-accelerated inference without CPU fallbacks, featuring a unified client interface that allows seamless interaction with all deployed LLMs through a single logical unit. The architecture comprises four main components: the Client Interface for user access, the Service Frontend for secure request routing and load balancing, the SDAI Controller for orchestration, deployment, and monitoring, and the Service Backend of heterogeneous GPU nodes executing workloads. By abstracting GPU-specific details and providing dynamic, VRAM-aware allocation and reallocation of models, AIvailable ensures efficient use of resources and resilience against failures or workload fluctuations. Targeting academic labs, private companies, and other constrained organizations, it supports diverse open LLMs helping democratize generative AI through the repurposing of legacy GPUs.

Large language models (LLMs) are increasingly explored for their reasoning capabilities, yet their ability to perform structured, constraint-based optimization from natural language remains insufficiently understood. This study evaluates twenty-one publicly available LLMs on a representative heterogeneous high-performance computing (HPC) workload mapping and scheduling problem. Each model received the same textual description of system nodes, task requirements, and scheduling constraints, and was required to assign tasks to nodes, compute the total makespan, and explain its reasoning. A manually derived analytical optimum of nine hours and twenty seconds served as the ground truth reference. Three models exactly reproduced the analytical optimum while satisfying all constraints, twelve achieved near-optimal results within two minutes of the reference, and six produced suboptimal schedules with arithmetic or dependency errors. All models generated feasible task-to-node mappings, though only about half maintained strict constraint adherence. Nineteen models produced partially executable verification code, and eighteen provided coherent step-by-step reasoning, demonstrating strong interpretability even when logical errors occurred. Overall, the results define the current capability boundary of LLM reasoning in combinatorial optimization: leading models can reconstruct optimal schedules directly from natural language, but most still struggle with precise timing, data transfer arithmetic, and dependency enforcement. These findings highlight the potential of LLMs as explainable co-pilots for optimization and decision-support tasks rather than autonomous solvers.

Modern DNNs often rely on edge-cloud model partitioning (MP), but widely used schemes fix shallow, static split points that underutilize edge compute and concentrate latency and energy on the server. The problem is exacerbated in autoregressive (AR) LLM inference, where per-token forward passes repeatedly generate bulky intermediate features (IFs). We introduce SLICER, a retraining-free, architecture-agnostic framework that compresses IFs to reduce both communication and server load in split computing. SLICER combines (i) asymmetric top-K filtering (ATKF) to sparsify low-magnitude activations, (ii) magnitude-splitting (MS) to group the remaining non-zeros into equal-cardinality blocks, and (iii) adaptive bit quantization (ABQ) that selects per-block bitwidths under a distortion budget. Across standard vision and LLM workloads (e.g., ImageNet/COCO; HellaSwag, PIQA, ARC-E/C, GSM8K, HumanEval), SLICER reduces uplink volume by up to 10x and server GPU time by up to 4.4x, while keeping task quality within ~0-3 pp of baseline. In multi-device settings and AR LLMs, SLICER scales by shifting meaningful compute to the edge and lowering bits-per-token and server time per token, stabilizing per-step traffic. The codec attaches to off-the-shelf models without retraining or architectural changes, offering a plug-and-play path to scalable, low-latency distributed inference. Code is provided in the supplementary material.

While large language models have transformed how we interact with AI systems, they have a critical weakness: they confidently state false information that sounds entirely plausible. This "hallucination" problem has become a major barrier to using these models where accuracy matters most. Existing solutions either require retraining the entire model, add significant computational costs, or miss the root causes of why these hallucinations occur in the first place. We present CausalGuard, a new approach that combines causal reasoning with symbolic logic to catch and prevent hallucinations as they happen. Unlike previous methods that only check outputs after generation, our system understands the causal chain that leads to false statements and intervenes early in the process. CausalGuard works through two complementary paths: one that traces causal relationships between what the model knows and what it generates, and another that checks logical consistency using automated reasoning. Testing across twelve different benchmarks, we found that CausalGuard correctly identifies hallucinations 89.3\% of the time while missing only 8.3\% of actual hallucinations. More importantly, it reduces false claims by nearly 80\% while keeping responses natural and helpful. The system performs especially well on complex reasoning tasks where multiple steps of logic are required. Because CausalGuard shows its reasoning process, it works well in sensitive areas like medical diagnosis or financial analysis where understanding why a decision was made matters as much as the decision itself.