Context Anxiety: How Self-Doubt Hinders LLM Reasoning

A new study from arXiv identifies 'context anxiety' in large language models, where models with sufficient capability to solve problems fail due to premature self-doubt. The research provides the first systematic analysis, showing that context anxiety partly stems from the model's inability to estimate required token counts for task completion. This phenomenon causes efficiency losses under perceived constraints. The authors demonstrate that models can learn alternative strategies to solve long-horizon problems without exhibiting context anxiety, suggesting performance gains may come from improving self-assessment rather than scaling capabilities.