The field of large language models is moving towards efficient inference-time scaling, with a focus on improving performance without requiring costly model re-training. Recent developments have introduced novel techniques for accelerating retrieval-augmented generation systems, scaling context lengths for diffusion language models, and adapting computation to input complexity. These innovations enable substantial performance gains on downstream tasks, such as long-context tasks and complex reasoning benchmarks. Notable papers in this area include CacheClip, which achieves fast time-to-first-token and high generation quality, and UltraLLaDA, which scales the context length to 128K for diffusion large language models. Additionally, EAGER and Catch Your Breath propose entropy-aware generation and adaptive computation methods for improving efficiency and performance.