The field of large language models (LLMs) is undergoing significant transformations, driven by the need for efficient deployment on edge devices. Recent research has focused on reducing computational overhead and memory demands, enabling the deployment of LLMs on resource-constrained devices. One of the key areas of research is model compression, where techniques such as quantization and sparsity are being explored to minimize computational overhead. Notable papers include D$^2$MoE, which proposes a dynamic scheduling algorithm and matryoshka weight quantization to improve inference throughput and reduce memory footprint. Another area of innovation is hardware acceleration, where researchers are developing novel architectures and accelerators to support efficient LLM deployment. TeLLMe presents a ternary LLM accelerator for edge FPGAs, achieving significant energy-efficiency advances and establishing a new benchmark for generative AI. COBRA introduces a binary Transformer accelerator with real 1-bit binary multiplication, surpassing ternary methods and delivering high energy efficiency and throughput improvements. Furthermore, research has demonstrated the importance of optimizing inference methods, including quantization techniques and novel sparsity paradigms. Gradual Binary Search and Dimension Expansion demonstrates a 40% increase in accuracy on common benchmarks compared to state-of-the-art methods. FGMP: Fine-Grained Mixed-Precision Weight and Activation Quantization achieves less than 1% perplexity degradation on Wikitext-103 for the Llama-2-7B model relative to an all-FP8 baseline design while consuming 14% less energy during inference. The development of more efficient and compressed models is also a key area of research, with importance-aware delta sparsification, saliency-aware partial retraining, and rate-constrained optimized training showing significant promise. ImPart achieves state-of-the-art delta sparsification performance, and Backslash enables flexible trade-offs between model accuracy and complexity. Finally, optimizing throughput, latency, and memory usage is crucial for real-world applications. Optimizing SLO-oriented LLM Serving with PD-Multiplexing presents a new LLM serving framework that achieves an average 5.1 times throughput improvement over state-of-the-art baselines. SlimPipe reduces accumulated activations and achieves near-zero memory overhead and minimal pipeline bubbles. L3 integrates DIMM-PIM and GPU devices to achieve up to 6.1 times speedup over state-of-the-art HBM-PIM solutions. Overall, the field of LLMs is rapidly evolving, with a focus on efficient deployment, model compression, and optimized inference methods. As researchers continue to innovate and push the boundaries of what is possible, we can expect to see significant advancements in the deployment of LLMs on edge devices.