Linear recurrent neural networks enable powerful long-range sequence modeling with constant memory usage and time-per-token during inference. These architectures hold promise for streaming applications at the edge, but deployment in resource-constrained environments requires hardware-aware optimizations to minimize latency and energy consumption. Unstructured sparsity offers a compelling solution, enabling substantial reductions in compute and memory requirements–when accelerated by compatible hardware platforms. In this paper, we conduct a scaling study to investigate the Pareto front of performance and efficiency across inference compute budgets. We find that highly sparse linear RNNs consistently achieve better efficiency-performance trade-offs than dense baselines, with $2\times$ less compute and $36$% less memory at iso-accuracy. Our models achieve state-of-the-art results on a real-time streaming task for audio denoising. By quantizing our sparse models to fixed-point arithmetic and deploying them on the Intel Loihi 2 neuromorphic chip for real-time processing, we translate model compression into tangible gains of $42\times$ lower latency and $149\times$ lower energy consumption compared to a dense model on an edge GPU. Our findings showcase the transformative potential of unstructured sparsity, paving the way for highly efficient recurrent neural networks in real-world, resource-constrained environments.

Large Language Models (LLMs) have demonstrated unprecedented capabilities in language understanding and generation, yet their significant computational requirements pose substantial challenges for scalability and environmental sustainability. In this introductory review, we examine the algorithm-hardware co-design strategies necessary for developing ultra-low-power LLMs. We begin by reviewing model-level approaches– efficient architectures, quantization, sparsity, distillation–that reduce parameter count and memory movement. We discuss hardware-centric innovations, including event-driven neuromor-phic accelerators, near-memory computing paradigms, and spe-cialized number formats, illustrating how these platforms leverage compressed models for substantial gains in energy efficiency. Finally, we address emerging frontiers beyond digital systems that may further reduce power consumption. We offer a blueprint for enabling low-power LLMs, emphasizing the necessity of cross-disciplinary collaboration for efficient AI at scale.

XR devices running chat-bots powered by Large Language Models (LLMs) have the to become always-on agents that enable much better productivity scenarios. Current screen based chat-bots do not take advantage of the the full-suite of natural inputs available in XR, including inward facing sensor data, instead they over-rely on explicit voice or text prompts, sometimes paired with multi-modal data dropped as part of the query. We propose a solution that leverages an attention framework that derives context implicitly from user actions, eye-gaze, and contextual memory within the XR environment. Our work minimizes the need for engineered explicit prompts, fostering grounded and intuitive interactions that glean user insights for the chat-bot.