From Tokens to Thought: Efficient Architectures for Machine Reasoning
Alice Chen, Liam O’Connor, Fatima Al-Mansouri, Mareike Gerhardt, Rajesh Iyer, Sofia Hernández, Daniel Weber, Min-Jae Lee, Chloe Dubois · 2025
The rapid advancements in large language models (LLMs) have revolutionized natural language processing and artificial intelligence, enabling unprecedented capabilities in understanding and generating human-like text. However, harnessing the full potential of LLMs for complex reasoning tasks poses significant challenges due to their substantial computational demands and limitations in reasoning efficiency. This survey provides a comprehensive overview of efficient reasoning models in the era of LLMs, focusing on architectural innovations, training methodologies, and integration strategies that balance reasoning accuracy with computational feasibility. We systematically categorize stateof-the-art approaches, including neurosymbolic methods, modular and compositional frameworks, and learning paradigms such as curriculum and reinforcement learning. Furthermore, we analyze key challenges related to robustness, interpretability, scalability, and evaluation, and identify promising future research directions. Finally, we highlight a wide range of practical applications across domains such as question answering, theorem proving, robotics, legal analysis, and education, demonstrating the transformative impact of efficient reasoning models.