project
Agentic Reasoning - Oxford University Introduces Framework to Enhance LLM Reasoning Skills
Agentic Reasoning is a framework developed by Oxford University to enhance the reasoning capabilities of Large Language Models (LLMs). It solves complex, multi-step reasoning problems by integrating external tools such as web search, code execution, and structured memory...
What is Agentic Reasoning?
Agentic Reasoning is a framework developed by Oxford University to enhance the reasoning capabilities of Large Language Models (LLMs). It solves complex, multi-step reasoning problems by integrating external tools such as web search, code execution, and structured memory. The core idea of Agentic Reasoning is to allow the LLM to dynamically invoke external agents (such as Mind Map agents, web search agents, and code agents) during the reasoning process, retrieving information, performing computational analysis, and organizing complex logical relationships in real time. The framework performs exceptionally well in doctoral-level scientific reasoning (such as the GPQA dataset) and domain-specific deep research tasks, outperforming existing Retrieval Augmentation (RAG) systems and closed-source LLMs.
The main functions of Agentic Reasoning
- Enhance multi-step reasoning abilityLLMs enable more efficient handling of complex problems requiring in-depth research and multi-step logical derivation by leveraging external tools such as web search, code execution, and structured memory.
- Real-time information retrieval and updatingUse web search agents to obtain the latest information in real time, ensuring the accuracy and timeliness of knowledge during the reasoning process.
- Organization of complex logical relationshipsKnowledge graphs are built based on Mind Map agents to help LLMs clearly organize and track logical relationships in the reasoning process, thereby improving their deductive reasoning capabilities.
- Computational analysis supportIt leverages code agents to perform programming tasks, providing accurate calculation results for problems requiring quantitative analysis.
- Improve reasoning efficiency and accuracyBased on task allocation and tool invocation, the burden on the main inference model is reduced, and the inference chain is avoided from being interrupted due to the processing of auxiliary tasks.
- Expert-level knowledge synthesisIn in-depth research tasks, it generates high-quality analysis reports that reach or even surpass the level of human experts.
The technical principles of Agentic Reasoning
- Dynamic tool invocation mechanismDuring inference, the LLM decides in real time whether to invoke external tools (such as web search or code execution) based on the current inference needs. When external information is needed, the LLM generates a specific query request and embeds a special tag (such as a "search" or "code" tag), pauses inference, and sends the request to the appropriate agent.
- Collaborative work of external agents:
- Mind Map AgentThis approach transforms the reasoning chain into a structured knowledge graph, providing logical support for reasoning based on entity recognition and semantic relationship extraction. The knowledge graph can be queried, helping LLMs quickly obtain relevant information during the reasoning process.
- Web search agentThe system retrieves relevant documents from the internet and uses the LLM framework to extract key information, generating concise summaries that ensure relevance and logical coherence.
- Code ProxyReceives code requests from the LLM, writes and executes the code, and returns the results. This avoids the complexity of the LLM directly generating and executing code, improving inference efficiency.
- Iterative Reasoning and Knowledge UpdateThe reasoning process is an iterative loop. The LLM updates the reasoning chain based on the results returned by the external agent, gradually improving the logical deduction until the final answer is obtained.
- Probability-based generative modelsThe generation of reasoning chains and final answers is based on a joint probability model, which combines task instructions, queries, tool outputs, and knowledge graphs to dynamically generate coherent reasoning processes and accurate answers.
- Inference optimization and verificationOptimize the reasoning process based on indicators such as tool call frequency, select the best reasoning path, and improve the accuracy and efficiency of reasoning.
The project address for Agentic Reasoning
- GitHub repository:https://github.com/theworldofagents/Agentic-Reasoning
- arXiv technical paper:https://arxiv.org/pdf/2502.04644
Application scenarios of Agentic Reasoning
- Academic research and solutions to complex problemsBased on multi-step reasoning and tool invocation, it solves doctoral-level scientific problems and provides high-precision answers.
- Medical decision supportIt combines web search and code execution to provide precise diagnostic and treatment solutions for medical scenarios.
- Financial and Legal StudiesQuickly retrieve regulations and data, generate high-quality research reports, and assist in professional decision-making.
- Complex Logic Games and Strategy OptimizationIn games like Werewolf, strategies with high win rates are achieved based on logical reasoning and relationship tracking.
- In-depth cross-disciplinary researchIt integrates information from multiple fields to generate comprehensive reports, facilitating the knowledge synthesis and analysis of complex problems.