A Survey of Frontiers in LLM Reasoning

Analysis

Why This Paper Matters

This survey addresses a critical need in the rapidly advancing field of large language models (LLMs): organizing the vast and fragmented body of work on reasoning. As LLMs move beyond simple chatbot interactions to complex problem-solving, understanding how reasoning is achieved becomes essential. The paper's two-dimensional taxonomy—regimes (inference-time vs. training) and architectures (standalone vs. agentic)—provides a clear lens to compare and contrast diverse approaches, from chain-of-thought prompting to reinforcement learning-based training like DeepSeek-R1.

The timing is particularly relevant as the field shifts from inference-scaling (e.g., spending more compute at test time) to learning-to-reason (e.g., training models to reason more effectively). The survey also captures the emergence of agentic workflows, where LLMs are embedded in compound systems with external tools and multi-agent collaboration, as exemplified by OpenAI Deep Research and Manus Agent. This systematic categorization helps researchers identify gaps and opportunities, making it a valuable resource for both newcomers and experts.

Technical Contributions

• Two-dimensional taxonomy: Regimes (inference-time vs. training) and architectures (standalone vs. agentic compound systems) provide a structured framework for classifying reasoning methods.
• Input-level vs. output-level analysis: Within each dimension, the survey distinguishes techniques that construct high-quality prompts (input-level) from those that refine multiple sampled candidates (output-level).
• Comprehensive coverage of learning algorithms: Includes supervised fine-tuning, reinforcement learning (PPO, GRPO), and training of reasoners and verifiers.
• Agentic workflow designs: Examines established patterns like generator-evaluator and LLM debate, as well as recent innovations.
• Trend identification: Highlights the shift from inference-scaling to learning-to-reason and the transition to agentic workflows.

Results

As a survey, the paper does not present new experimental results. Instead, it synthesizes existing literature, offering a structured overview of the state of the art. The main outcome is the taxonomy itself, which organizes methods along two orthogonal dimensions and two perspectives, enabling systematic understanding. The survey also identifies emerging trends, such as the rise of DeepSeek-R1 and agentic systems, which are shaping the future of LLM reasoning.

Significance

The survey provides a foundational reference for the AI community, helping to consolidate knowledge and guide future research. By clarifying the landscape, it enables researchers to position their work within the broader context and identify underexplored areas. For practitioners, it offers a practical guide to selecting appropriate reasoning techniques based on their needs. The emphasis on agentic workflows and learning-to-reason reflects the direction of the field, making this survey a timely and impactful contribution.