Atom of Thoughts: A New Approach for Efficient Reasoning in Large Language Models

Large language models (LLMs) have made impressive progress in recent years. Their performance is further enhanced by scaling during training and through so-called test-time scaling. Test-time scaling allows for more effective reasoning during the inference phase. However, existing methods of test-time scaling encounter difficulties with increasing complexity of the reasoning tasks. The accumulation of historical information not only leads to a waste of computational resources but can also impair effective reasoning.

A new approach, called "Atom of Thoughts" (AoT), promises a remedy here. The core idea of AoT is based on the observation that complex thought processes are often achieved by solving a series of independent sub-questions. Each of these sub-questions is self-contained and verifiable. These sub-questions, referred to as "atomic questions," depend mainly on their current state and not on the accumulated history, similar to the memoryless transitions in a Markov process.

AoT implements a two-stage transition mechanism, consisting of decomposition and contraction. In the decomposition phase, the current question is decomposed into a dependency-based directed acyclic graph. Subsequently, in the contraction phase, the sub-questions of this graph are combined into a new atomic question. This iterative decomposition-contraction process continues until directly solvable atomic questions are reached. This realizes Markov transitions between the question states.

Advantages and Applications of Atom of Thoughts

A key advantage of AoT lies in the elimination of the need to store historical dependencies during the reasoning process. This allows models to focus their computational resources on the current question status, thus increasing efficiency. Furthermore, the atomic questions can be seamlessly integrated into existing test-time scaling methods. AoT can thus serve as a plug-in extension to improve reasoning capabilities.

Experimental results on various benchmarks demonstrate the effectiveness of AoT both as a standalone framework and as a plug-in extension for existing methods. For example, in combination with gpt-4o-mini on HotpotQA, AoT achieved an F1 score of 80.6%, which is a significant improvement over o3-mini (3.4%) and DeepSeek-R1 (10.6%).

Outlook and Significance for AI Development

AoT represents a promising approach for optimizing test-time scaling in LLMs. The ability to decompose complex reasoning into a series of atomic questions allows for more efficient use of computational resources and improves the accuracy of the results. The integration of AoT into existing and future LLMs could lead to significant advances in areas such as question-answering systems, automated text generation, and knowledge representation.

The development of efficient test-time scaling methods is crucial for the further improvement of the performance of LLMs. AoT offers an innovative approach that has the potential to expand the boundaries of what is possible in the field of artificial intelligence.

Bibliography: - https://arxiv.org/abs/2502.12018 - https://arxiv.org/html/2502.12018v1 - https://chatpaper.com/chatpaper/zh-CN/paper/108366 - https://paperreading.club/page?id=284805 - https://huggingface.co/papers - https://github.com/ThreeSR/Awesome-Inference-Time-Scaling - https://blog.ml.cmu.edu/2025/01/08/optimizing-llm-test-time-compute-involves-solving-a-meta-rl-problem/ - https://eccv.ecva.net/virtual/2024/papers.html - https://www.researchgate.net/publication/385944696_AtomThink_A_Slow_Thinking_Framework_for_Multimodal_Mathematical_Reasoning