arXiv:2602.15829v2 Announce Type: replace 
Abstract: The superficial alignment hypothesis (SAH) posits that large language models learn most of their knowledge during pre-training, and that post-training merely surfaces this knowledge. The SAH, however, lacks a precise definition, which has led to (i) different and seemingly orthogonal arguments supporting it, and (ii) important critiques to it. We propose a new metric called task complexity: the length of the shortest program that achieves a target performance on a task. In this framework, the SAH simply claims that pre-trained models drastically reduce the complexity of achieving high performance on many tasks. Our definition unifies prior arguments supporting the SAH, interpreting them as different strategies to find such short programs. Experimentally, we estimate the task complexity of mathematical reasoning, machine translation, and instruction following; we then show that these complexities can be remarkably low when conditioned on a pre-trained model. Further, we find that pre-training enables access to strong performances on our tasks, but it can require programs of gigabytes of length to access them. Post-training, on the other hand, collapses the complexity of reaching this same performance by several orders of magnitude. Overall, our results highlight that task adaptation often requires surprisingly little information -- often just a few kilobytes.

الدراسة الحديثة تقوم بتفعيل فرضية المحاذاة السطحية (SAH) من خلال تقديم مقياس جديد يسمى تعقيد المهمة، والذي يقيس أقصر طول للبرنامج المطلوب لتحقيق أداء مستهدف في مهام متنوعة. يقترح هذا الإطار أن النماذج اللغوية الكبيرة المدربة مسبقًا تبسط بشكل كبير عملية تحقيق أداء عالٍ في مجالات مثل التفكير الرياضي والترجمة الآلية واتباع التعليمات.

El estudio reciente operacionaliza la hipótesis de alineación superficial (SAH) al introducir una nueva métrica llamada complejidad de tarea, que mide la longitud del programa más corto necesario para alcanzar un rendimiento objetivo en diversas tareas. Este marco sugiere que los grandes modelos de lenguaje preentrenados simplifican drásticamente el proceso de lograr un alto rendimiento en áreas como el razonamiento matemático, la traducción automática y el seguimiento de instrucciones.

La récente étude opérationnalise l'hypothèse d'alignement superficiel (SAH) en introduisant une nouvelle métrique appelée complexité de tâche, qui mesure la longueur du programme le plus court nécessaire pour atteindre une performance cible sur diverses tâches. Ce cadre suggère que les grands modèles de langage pré-entraînés simplifient considérablement le processus d'atteinte de performances élevées dans des domaines tels que le raisonnement mathématique, la traduction automatique et le suivi d'instructions.

The recent study operationalizes the superficial alignment hypothesis (SAH) by introducing a new metric called task complexity, which measures the shortest program length required to achieve target performance on various tasks. This framework suggests that pre-trained large language models significantly simplify the process of attaining high performance in areas such as mathematical reasoning, machine translation, and instruction following.