Efficient Inference for Large Reasoning Models: A Survey.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Large Reasoning Models (LRMs) enhance Large Language Models (LLMs) for complex task reasoning.
• LRMs' deliberative processes cause inefficiencies in token usage, memory, and inference time.

Purpose of the Study:

• To review and categorize efficient inference methods for LRMs.
• To address token inefficiency while maintaining reasoning quality.
• To identify challenges and future research directions in efficient LRM inference.

Main Methods:

• Categorization of methods into explicit compact Chain-of-Thought (CoT) and implicit latent CoT.
• Empirical analysis of methods based on reasoning scenarios, objectives, and performance.
• Discussion of strengths, weaknesses, and open challenges.

Main Results:

• Two main categories of efficient inference methods for LRMs are identified: explicit compact CoT and implicit latent CoT.
• Analysis covers reasoning scenarios, objective functions, and performance-efficiency trade-offs.
• Key insights for enhancing LRM inference efficiency are highlighted.

Conclusions:

• Efficient inference is crucial for practical LRM deployment.
• Addressing challenges like controllability, interpretability, safety, and broader applications is essential.
• Techniques like model merging and new architectures offer promising avenues for improvement.