Lost in Context: AI Models Doubt Themselves Before They Should
A new arXiv study reveals that frontier reasoning models fail solvable problems due to 'context anxiety'—premature self-doubt caused by an inability to estimate required tokens. This leads to significant efficiency losses and shorter, less accurate responses.