A preregistered study on confidence calibration in large language models has quantified a troubling pattern: current LLMs consistently report higher confidence in their answers than their actual accuracy warrants. Researchers investigating this phenomenon across diverse tasks found that the gap between stated confidence and correct performance mirrors human overconfidence bias—but without the evolutionary context that might explain it in people. In concrete terms, models frequently claim 85-90% confidence on tasks where they achieve only 65-75% accuracy, a systematic miscalibration that could prove dangerous in medical, financial, or legal applications where stakeholders rely on model certainty estimates to make decisions. This work, published on arXiv this week, suggests that end users currently have no reliable way to trust a model's own assessment of whether it should be trusted.

Compounding this reliability problem, a separate analysis titled 'How Much Thinking is Enough?' reveals that reasoning-capable LLMs engage in extensive computational waste during their problem-solving processes. Researchers examining the internal reasoning traces of models using chain-of-thought prompting discovered that 30-40% of the computational work consists of redundant reformulation, self-verification loops, and circular self-reflection that doesn't materially improve final answers. A model might spend tokens reformulating the same insight three times or verifying conclusions it had already reached, incurring significant latency, GPU time, and energy costs without proportional gains in accuracy. This inefficiency becomes critical as organizations deploy these systems at scale, where unnecessary token consumption directly translates to infrastructure costs and slower user-facing responses.

Together, these findings expose a fundamental gap between how LLMs are marketed—as powerful reasoning engines worthy of deep trust—and how they actually operate: overconfident systems that waste resources on redundant thinking. The implications extend beyond academic concern. For developers building LLM-powered applications, these papers suggest that simply scaling up model size or compute may not address core reliability issues. Instead, systems may require explicit calibration layers to realistically estimate confidence, and inference optimization to eliminate wasteful reasoning patterns. As enterprise adoption accelerates, these technical debt items increasingly matter. Organizations deploying LLMs for consequential decisions should treat confidence outputs with skepticism and actively test whether models' self-reported certainty correlates with real-world performance.