arXiv:2605.16824v1 Announce Type: new 
Abstract: Large language models (LLMs) generate not only reasoning text, but also token-level confidence trajectories that record how uncertainty evolves during inference. Whether these trajectories are relevant to reasoning correctness remains unclear. Here we show that confidence trajectories encode a content-agnostic confidence geometry associated with trace-level final-answer correctness. Using only token-level confidence values, without access to the input question, reasoning text, hidden states, or external verifiers, we find that low-dimensional representations of confidence trajectories separate correct from incorrect reasoning traces. Across GSM8K, MATH, and MMLU, this geometric separation is quantitatively linked to downstream predictability: stronger clustering of correct and incorrect traces, measured by the Davies--Bouldin index, consistently corresponds to higher correctness-discrimination AUC. We further show that correctness-related information is enriched in the tail of reasoning, suggesting that late-stage confidence dynamics carry key correctness signals. We propose NeuralConf, a lightweight estimator that learns from confidence trajectories for correctness evaluation. Under a fixed trace budget, NeuralConf-derived scores improve confidence-weighted answer aggregation over majority voting, tail confidence, and other static baselines. These results reveal that LLMs expose trace-intrinsic statistical signals of correctness through their own confidence dynamics, offering a route to improve inference using information already present within generation.

كشفت الأبحاث الأخيرة أن نماذج اللغة الكبيرة (LLMs) تولد مسارات ثقة على مستوى الرموز يمكن أن تشير إلى صحة استدلالاتها. تُظهر هذه الدراسة أن هذه المسارات يمكن أن تفصل بفعالية بين آثار الاستدلال الصحيحة وغير الصحيحة دون الحاجة للوصول إلى السؤال المدخل أو إلى مدققين خارجيين.

Investigaciones recientes han revelado que los modelos de lenguaje grandes (LLMs) generan trayectorias de confianza a nivel de tokens que pueden indicar la corrección de su razonamiento. Este estudio demuestra que estas trayectorias de confianza pueden separar efectivamente las trazas de razonamiento correctas de las incorrectas sin necesidad de acceder a la pregunta de entrada o a verificadores externos.

Des recherches récentes ont révélé que les grands modèles de langage (LLMs) génèrent des trajectoires de confiance au niveau des tokens qui peuvent indiquer la justesse de leur raisonnement. Cette étude démontre que ces trajectoires de confiance peuvent efficacement séparer les traces de raisonnement correctes des incorrectes sans avoir besoin d'accéder à la question d'entrée ou à des vérificateurs externes.

Recent research has revealed that large language models (LLMs) generate token-level confidence trajectories that can indicate the correctness of their reasoning. This study demonstrates that these confidence trajectories can effectively separate correct from incorrect reasoning traces without needing access to the input question or external verifiers.