The recent developments in the research area of large language models (LLMs) and large multimodal models (LMMs) have shown a significant shift towards more efficient, unsupervised, and preference-guided methodologies. A notable trend is the exploration of unsupervised routing and ranking techniques, which aim to bypass the need for labeled data, thereby reducing resource intensity and improving generalizability. These methods leverage uncertainty signals and graphical models to estimate and optimize model performance, offering practical solutions for real-world applications. Additionally, there is a growing emphasis on active preference learning, where frameworks are being developed to integrate natural language feedback with conventional preference learning, enhancing both computational efficiency and human interpretability. The integration of Bayesian optimization with structural-aware models is also emerging as a powerful strategy for prompt selection, addressing the limitations of traditional black-box approaches. Overall, the field is progressing towards more adaptive, efficient, and human-centric model optimization and selection strategies.

Noteworthy papers include 'Smoothie: Label Free Language Model Routing,' which introduces an unsupervised routing approach that outperforms baselines, and 'Ranked from Within: Ranking Large Multimodal Models for Visual Question Answering Without Labels,' which demonstrates the effectiveness of uncertainty-based metrics for unsupervised model ranking.