The field of large language models is moving towards more efficient and effective training methods. Recent developments have focused on adaptive fine-tuning strategies, such as skill-targeted adaptive training, which improve model performance on specific tasks by identifying and addressing skill gaps. Additionally, there is a growing interest in calibration data curation, which aims to preserve model capabilities after compression. Other notable trends include the use of critical token fine-tuning, dynamic nested depth, and hierarchical alignment to enhance model reasoning and performance. These innovative approaches have shown significant improvements over traditional methods, such as supervised fine-tuning, and have the potential to advance the field of large language models. Noteworthy papers include: Skill-Targeted Adaptive Training, which introduces a new fine-tuning strategy that improves model performance on mathematical reasoning tasks. Preserving LLM Capabilities through Calibration Data Curation, which proposes a framework for calibration data curation to preserve model capabilities after compression. Enhancing Large Language Model Reasoning via Selective Critical Token Fine-Tuning, which presents a simple yet effective approach to fine-tune models on critical tokens. Hierarchical Alignment, which introduces a novel method for surgical fine-tuning via functional layer specialization. Dr.LLM, which proposes a retrofittable framework for dynamic layer routing in large language models. Informed Routing in LLMs, which introduces a new paradigm for smarter token-level computation. What Layers When, which presents a simple residual-stream gating mechanism for token-wise layer skipping.