Probing techniques have shown promise in revealing how LLMs encode human-interpretable concepts, particularly when applied to curated datasets. However, the factors governing a dataset's suitability for effective probe training are not well-understood. This study hypothesizes that probe performance on such datasets reflects characteristics of both the LLM's generated responses and its internal feature space. Through quantitative analysis of probe performance and LLM response uncertainty across a series of tasks, we find a strong correlation: improved probe performance consistently corresponds to a reduction in response uncertainty, and vice versa. Subsequently, we delve deeper into this correlation through the lens of feature importance analysis. Our findings indicate that high LLM response variance is associated with a larger set of important features, which poses a greater challenge for probe models and often results in diminished performance. Moreover, leveraging the insights from response uncertainty analysis, we are able to identify concrete examples where LLM representations align with human knowledge across diverse domains, offering additional evidence of interpretable reasoning in LLMs.

Large language models (LLMs) have demonstrated impressive performance across a variety of tasks (Google, 2023; OpenAI, 2023; Zhao et al., 2023). This success has led to their widespread adoption and significant involvement in various decision-making applications, such as healthcare (Karabacak and Margetis, 2023; Sallam, 2023; Yang et al., 2023), education (Xiao et al., 2023), finance (Wu et al., 2023b), and law (Zhang et al., 2023a). However, despite their rapid adoption, the reliability of LLMs in handling high-stakes tasks has yet to be demonstrated (Arkoudas, 2023; Huang et al., 2023a). The reliability is particularly critical in domains such as healthcare, where model responses can have immediate and significant impacts on human behavior and hence their well-being (Ji et al., 2023). Therefore, understanding LLMs' reasoning and decision-making processes and how they influence response uncertainty is critical for their safe and reliable deployment.