Language Models (Mostly) Know When to Stop Reading
Conference on Neural Information Processing Systems (NeurIPS 2025)
This paper investigates when language models can determine they have gathered sufficient information to answer a question, demonstrating their ability to adaptively stop reading.