The field of natural language processing is witnessing significant advancements in tokenization and confidence estimation, leading to improved performance and efficiency in large language models. Researchers are exploring innovative approaches to tokenization, such as semantic-aware tokenization and cross-boundary pattern learning, which have shown promising results in reducing token redundancy and improving computation efficiency. Additionally, confidence estimation methods are being developed to provide fine-grained, continuous confidence estimates throughout the generation process, enhancing the trustworthiness and reliability of language model outputs. These developments have the potential to complement architectural innovations and pave the way for further improvements in language model performance. Noteworthy papers include: SupraTok, which achieves a 31% improvement in English tokenization efficiency, and FineCE, which delivers accurate, fine-grained confidence scores during text generation. QuickMerge and SemToken also demonstrate significant improvements in compute-accuracy tradeoffs and tokenization efficiency, respectively. Confidence-Modulated Speculative Decoding is another notable approach, offering a principled method for efficient and robust decoding in large language models.