arXiv:2505.13820v5 Announce Type: replace 
Abstract: Large language models (LLMs) exhibit strong capabilities as decision-making agents by interleaving reasoning and actions, as seen in ReAct-style frameworks. Yet, their practical deployment is constrained by high inference costs and large model sizes. We propose Structured Agent Distillation, a framework that compresses large LLM-based agents into smaller student models while preserving both reasoning fidelity and action consistency. Unlike standard token-level distillation, our method segments trajectories into [REASON] and [ACT] spans, applying segment-specific losses to align each component with the teacher's behavior. This structure-aware supervision enables compact agents to better replicate the teacher's decision process. Experiments on ALFWorld, HotPotQA-ReAct, and WebShop show that our approach consistently outperforms token-level and imitation learning baselines, achieving significant compression with minimal performance drop. Scaling and ablation results further highlight the importance of span-level alignment for efficient and deployable agents.

تم تقديم إطار جديد يسمى تقطير الوكيل المنظم لضغط نماذج اللغة الكبيرة (LLMs) إلى نماذج طلابية أصغر مع الحفاظ على دقة التفكير وتناسق الأفعال. تقوم هذه الطريقة بتقسيم عمليات اتخاذ القرار إلى فترات تفكير وفعل، وتطبيق خسائر محددة لمواءمة النماذج الطلابية مع نظرائها الأكبر.

Se ha introducido un nuevo marco llamado Destilación de Agente Estructurado para comprimir grandes modelos de lenguaje (LLMs) en modelos estudiantes más pequeños, manteniendo la fidelidad del razonamiento y la consistencia de las acciones. Este método segmenta los procesos de toma de decisiones en tramos de razonamiento y acción, aplicando pérdidas específicas para alinear los modelos estudiantes con sus contrapartes más grandes.

Un nouveau cadre appelé Distillation d'Agent Structuré a été introduit pour compresser les grands modèles de langage (LLMs) en modèles étudiants plus petits tout en maintenant la fidélité du raisonnement et la cohérence des actions. Cette méthode segmente les processus de prise de décision en intervalles de raisonnement et d'action, appliquant des pertes spécifiques pour aligner les modèles étudiants avec leurs homologues plus grands.

A new framework called Structured Agent Distillation has been introduced to compress large language models (LLMs) into smaller student models while maintaining reasoning fidelity and action consistency. This method segments decision-making processes into reasoning and action spans, applying specific losses to align the student models with their larger counterparts.