The recent developments in the field of active learning and preference learning highlight a significant shift towards more efficient, adaptive, and generalized methods that can operate across varying conditions and requirements. A common theme across the research is the exploration of innovative strategies to minimize the reliance on extensive labeled datasets, which are often costly and time-consuming to produce. Techniques leveraging deep reinforcement learning, randomized algorithms, and advanced uncertainty quantification methods are at the forefront, aiming to dynamically adapt to the learning environment and optimize the selection of data points for labeling. These approaches not only promise to enhance the performance of machine learning models in scenarios with limited labeled data but also aim to bridge the gap between low and high label budget regimes, offering more versatile solutions to active learning challenges.

Noteworthy papers include:

• A study on active human preference learning that introduces a randomized Frank-Wolfe algorithm for efficiently learning from limited comparison feedback, significantly advancing the field's capability to handle large-scale preference learning tasks.
• Research on image classification with deep reinforcement active learning, which combines deep reinforcement learning with active learning to dynamically adapt sample selection strategies, showcasing superior performance on image classification benchmarks.
• The proposal of uncertainty herding, a method that generalizes across low- and high-budget active learning scenarios, demonstrating state-of-the-art performance across various tasks.
• A novel approach to learning general halfspaces that distinguishes between the capabilities of label queries and membership queries, providing a strong theoretical foundation for understanding the limitations and potentials of different query models in active learning.
• The introduction of BALSA, an adaptation of the BALD algorithm for regression with normalizing flows, which sets new standards for uncertainty quantification and active learning in regression tasks.