AtomThink: Multimodal Slow Thinking With Atomic Step Reasoning.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Multimodal reasoning in large language models (LLMs) is a complex challenge.
• Existing methods often use rigid templates or unstructured approaches, leading to inefficiencies.
• Adaptive reasoning strategies are needed to handle varying question complexities.

Purpose of the Study:

• To develop a novel framework, AtomThink, for adaptive multimodal reasoning in LLMs.
• To introduce a Self-structured Chain of Thought (SCoT) paradigm for flexible reasoning.
• To improve both accuracy and efficiency in multimodal tasks.

Main Methods:

• Proposed the Self-structured Chain of Thought (SCoT) paradigm with minimal semantic atomic steps.
• Designed the AtomThink framework with a data engine, supervised fine-tuning, policy-guided inference, and an atomic capability metric.
• Utilized serialized inference data for supervised fine-tuning.

Main Results:

• Achieved over 10% average accuracy gains on MathVista and MathVerse datasets.
• Demonstrated significant improvements compared to state-of-the-art structured Chain of Thought (CoT) approaches.
• Improved data utilization by 5x and inference efficiency by 85.3%.

Conclusions:

• AtomThink enables adaptive reasoning in multimodal large language models (MLLMs).
• The SCoT paradigm offers flexible and efficient reasoning structures.
• AtomThink significantly advances the performance and efficiency of multimodal AI systems.