Lost in Context: Addressing Context Anxiety in Large Language Models
06:00 · July 27, 2026 · arXiv cs.AI RSS
Conventional wisdom suggests that reasoning models fail when problems exceed their capabilities. However, we find that frontier reasoning models sometimes possess the necessary capabilities to solve problems but fail due to premature self-doubt -- a phenomenon informally known as context anxiety. We provide the first systematic study of context anxiety, demonstrating that it arises, in part, from a model's inability to accurately estimate the tokens required to complete a task. We also show that context anxiety leads to material efficiency losses when models operate under perceived constraints. Building on this analysis, we further show that models can learn alternative strategies for solving long-horizon problems without exhibiting context anxiety, suggesting that performance improvements may be achievable not through scaling model capabilities, but by improving models' ability to accurately assess and adapt to their own limitations.
Summary
A recent arXiv paper identifies context anxiety as a distinct failure mode in frontier reasoning models. Rather than abandoning tasks because they exceed model capabilities, these systems sometimes terminate early after misjudging the number of tokens their solution will require. The authors frame this premature self-doubt as a calibration problem: models overestimate token demand, abandon viable paths, and, when they do continue, generate unnecessarily long outputs.
To isolate the phenomenon, the work introduces detection protocols that separate anxiety-driven failures—explicit claims of insufficient context—from capability-driven failures in which models attempt solutions yet still err. Experiments on the Tower of Hanoi puzzle show that models expressing context anxiety overestimate required tokens by roughly 24 percent. This miscalibration correlates with a 15 percent drop in accuracy and, on successful runs, a 54 percent rise in tokens consumed. Comparable patterns appear on a shortest-path grid task, indicating the issue is not task-specific.
The authors further demonstrate that the behavior is mutable. Lightweight supervised fine-tuning on reasoning traces that exhibit no context anxiety reduces the frequency of anxious terminations by more than half. The same intervention improves both final accuracy and token efficiency, suggesting that gains can come from better internal estimation of task length rather than from increases in model scale or training compute alone. The findings are relevant to researchers working on long-horizon planning and self-calibration in large language models.
Why it matters
Actionable for Dutch AI teams via fine-tuning protocols that improve LLM reliability without scaling; aligns with EU emphasis on transparent, robust AI systems for SMEs and regulated deployments.
