Superficial Beliefs in LLM Decision-Making

The paper presents findings on the decision-making processes of large language models (LLMs), highlighting a phenomenon termed "superficial belief," where models exhibit systematic behavior based on visible attributes but demonstrate limited verbal access to the underlying decision drivers. Through synthetic binary decision tasks, the authors show that while models can predict choices based on behavioral models, their self-reported reasons only partially align with these drivers. This research is significant for practitioners as it underscores the need to critically evaluate LLM outputs and understand the limitations in the models' ability to articulate their decision-making rationales.