Reproduce “A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems” on OpenResearch

About this paper

Zixuan Ke, Fangkai Jiao, Yifei Ming, Xuan-Phi Nguyen, Austin Xu, Do Xuan Long, Minzhi Li, Chengwei Qin and 4 more

The survey organizes approaches to language-model reasoning across when computation occurs and which system performs it. It distinguishes inference-time scaling from training on reasoning trajectories, and standalone models from tool-using and multi-agent systems. It applies input and output perspectives across these settings. The framework supports comparison, not ranking: it establishes no evidence that one approach is more effective or efficient.

Key insights

• The field of LLM reasoning is evolving from relying primarily on inference-time techniques (e.g., Chain-of-Thought prompting) towards dedicated training methods, termed 'learning to reason', for more profound improvements.
• There is a clear shift from standalone LLMs to interactive agentic systems that leverage external tools, knowledge bases, and multi-agent collaboration for enhanced reasoning capabilities.
• The process of evaluating LLM reasoning is increasingly becoming a reasoning task itself, with growing reliance on automated feedback mechanisms like tree search and LLM-as-judge models to address annotation costs.

Summary by alphaXiv.