arXiv:2602.03542v2 Announce Type: replace-cross 
Abstract: Large language models (LLMs) are trained and tested extensively on symbolic representations such as code and graphs, yet real-world user tasks are often specified in natural language. To what extent can LLMs generalize across these representations? Here, we approach this question by studying isomorphic tasks involving procedures represented in code, graphs, and natural language (e.g., scheduling steps in planning). We find that training LLMs with popular post-training methods on graphs or code data alone does not reliably generalize to corresponding natural language tasks, while training solely on natural language can lead to inefficient performance gains. To address this gap, we propose a two-stage reinforcement learning curriculum that first trains on symbolic, then natural language data. The curriculum substantially improves model performance across model families and tasks. Remarkably, a 1.5B Qwen model trained by our method can closely match zero-shot GPT-4o in naturalistic planning. Finally, our analysis suggests that successful cross-representation generalization can be interpreted as a form of generative analogy, which our curriculum effectively encourages. The dataset and code used in this paper can be found \href{https://github.com/fangru-lin/procedure_generalization_llm}{here}.

استكشفت الأبحاث الحديثة قدرة نماذج اللغة الكبيرة (LLMs) على تعميم الإجراءات عبر تمثيلات مختلفة، مثل الشيفرة، والرسوم البيانية، واللغة الطبيعية. تكشف الدراسة أن تدريب نماذج LLMs فقط على البيانات الرمزية لا يترجم بشكل فعال إلى مهام اللغة الطبيعية، مما أدى إلى تقديم منهجية تعلم تعزيزية من مرحلتين تعزز الأداء عبر مهام متنوعة.

Investigaciones recientes han explorado la capacidad de los grandes modelos de lenguaje (LLMs) para generalizar procedimientos a través de diferentes representaciones, como código, gráficos y lenguaje natural. El estudio revela que entrenar a los LLMs únicamente con datos simbólicos no se traduce de manera efectiva en tareas de lenguaje natural, lo que ha llevado a la introducción de un currículo de aprendizaje por refuerzo en dos etapas que mejora el rendimiento en diversas tareas.

Des recherches récentes ont exploré la capacité des grands modèles de langage (LLMs) à généraliser des procédures à travers différentes représentations, telles que le code, les graphiques et le langage naturel. L'étude révèle que former les LLMs uniquement sur des données symboliques ne se traduit pas efficacement par des tâches en langage naturel, ce qui a conduit à l'introduction d'un curriculum d'apprentissage par renforcement en deux étapes qui améliore les performances sur diverses tâches.

Recent research has explored the ability of large language models (LLMs) to generalize procedures across different representations, such as code, graphs, and natural language. The study reveals that training LLMs solely on symbolic data does not effectively translate to natural language tasks, prompting the introduction of a two-stage reinforcement learning curriculum that enhances performance across various tasks.