The predictions of Large Language Models (LLMs) on downstream tasks often improve significantly when including examples of the input--label relationship in the context. However, there is currently no consensus about how this in-context learning (ICL) ability of LLMs works. For example, while Xie et al. (2021) liken ICL to a general-purpose learning algorithm, Min et al. (2022) argue ICL does not even learn label relationships from in-context examples. In this paper, we provide novel insights into how ICL leverages label information, revealing both capabilities and limitations. To ensure we obtain a comprehensive picture of ICL behavior, we study probabilistic aspects of ICL predictions and thoroughly examine the dynamics of ICL as more examples are provided. Our experiments show that ICL predictions almost always depend on in-context labels, and that ICL can learn truly novel tasks in-context. However, we also find that ICL struggles to fully overcome prediction preferences acquired from pre-training data, and, further, that ICL does not consider all in-context information equally.

Quality-sensitive applications of machine learning (ML) require quality assurance (QA) by humans before the predictions of an ML model can be deployed. QA for ML (QA4ML) interfaces require users to view a large amount of data and perform many interactions to correct errors made by the ML model. An optimized user interface (UI) can significantly reduce interaction costs. While UI optimization can be informed by user studies evaluating design options, this approach is not scalable because there are typically numerous small variations that can affect the efficiency of a QA4ML interface. Hence, we propose using simulation to evaluate and aid the optimization of QA4ML interfaces. In particular, we focus on simulating the combined effects of human intelligence in initiating appropriate interaction commands and machine intelligence in providing algorithmic assistance for accelerating QA4ML processes. As QA4ML is usually labor-intensive, we use the simulated task completion time as the metric for UI optimization under different interface and algorithm setups. We demonstrate the usage of this UI design method in several QA4ML applications.