arXiv:2606.11211v1 Announce Type: cross 
Abstract: The ability of large language models (LLMs) to express calibrated uncertainty is important for safe deployment. Chain-of-thought (CoT) reasoning is widely used to improve accuracy and reliability, but its effect on calibration is not fully understood. We show that this picture is incomplete: in some settings, increasing the reasoning budget beyond a task-specific threshold can cause models to become systematically overconfident, assigning high confidence to incorrect answers. We call this phenomenon Calibration Drift Under Reasoning (CDUR) and study it both theoretically and empirically.
  We define reasoning budget B and analyze conditions under which Expected Calibration Error ECE(B) follows a non-monotonic pattern: it first decreases as reasoning corrects errors, then increases as longer reasoning produces internally consistent but incorrect explanations. We propose a Hypothesis Lock-In model based on autoregressive generation to explain this behavior.
  We evaluate Llama-3.1-8B and Llama-3.3-70B on 47 reasoning-trap questions across four reasoning budgets and three seeds (1,368 API calls; 574 valid responses). The 8B model shows non-monotonic calibration behavior, while results for the 70B model are limited to baseline evaluation and are inconclusive for budget-dependent effects.
  We introduce CABStop, a calibration-aware stopping rule that halts reasoning when confidence diverges from an auxiliary accuracy estimate. These results suggest that increasing reasoning depth does not always improve reliability and should be monitored carefully.

أظهرت الأبحاث الأخيرة وجود ظاهرة تُسمى انحراف المعايرة تحت التفكير (CDUR) في نماذج اللغة الكبيرة (LLMs)، حيث يمكن أن تؤدي الميزانيات المفرطة للتفكير إلى ثقة مفرطة في إجابات غير صحيحة. تسلط هذه الدراسة الضوء على أنه على الرغم من أن التفكير المتسلسل يمكن أن يعزز الدقة، إلا أنه قد يقدم أيضًا أخطاء منهجية عند تجاوزه العتبات المحددة للمهام.

Investigaciones recientes han identificado un fenómeno denominado Calibration Drift Under Reasoning (CDUR) en los grandes modelos de lenguaje (LLMs), donde presupuestos de razonamiento excesivos pueden llevar a una sobreconfianza en respuestas incorrectas. Este estudio destaca que, aunque el razonamiento en cadena puede mejorar la precisión, también puede introducir errores sistemáticos cuando se supera un umbral específico de la tarea.

Des recherches récentes ont identifié un phénomène appelé Calibration Drift Under Reasoning (CDUR) dans les grands modèles de langage (LLMs), où des budgets de raisonnement excessifs peuvent entraîner une confiance excessive dans des réponses incorrectes. Cette étude souligne que, bien que le raisonnement en chaîne puisse améliorer l'exactitude, il peut également introduire des erreurs systématiques lorsqu'il est poussé au-delà de seuils spécifiques à la tâche.

Recent research has identified a phenomenon termed Calibration Drift Under Reasoning (CDUR) in large language models (LLMs), where excessive reasoning budgets can lead to overconfidence in incorrect answers. This study highlights that while chain-of-thought reasoning can enhance accuracy, it may also introduce systematic errors when pushed beyond task-specific thresholds.