The field of in-context learning is moving towards more efficient and effective methods for improving language model performance. Research is focusing on reducing the computational costs associated with fine-tuning large language models, while also exploring new approaches for incorporating contextual information into the learning process. Notably, the development of novel training objectives and frameworks is enabling better control over specific module sequences within language models, leading to improved performance and interpretability. Additionally, the application of in-context learning to multimodal tasks, such as visual emotion understanding and sentiment analysis, is showing promising results. Overall, the field is advancing towards more adaptable and generalizable language models that can effectively leverage contextual information to improve their performance. Noteworthy papers include: MAPLE, which proposes a many-shot adaptive pseudo-labeling framework for in-context learning, enabling the use of pseudo-labeled samples to compensate for the lack of labeled data. EmoGist, which introduces a training-free, in-context learning method for visual emotion classification, achieving up to 13 points improvement in micro F1 scores on the multi-label Memotion dataset.