Three Chinese students from Oxford University proposed a new framework calledAgentic ReasoningToolbench, which aims to significantly enhance the reasoning capabilities of large language models (LLMs) by integrating external tools through tool-using agents.
Unlike traditional LLMs that rely entirely on internal reasoning, Agentic Reasoning dynamically invokes external tools such as web search, code execution, and structured reasoning-context memory to solve complex problems requiring in-depth research and multi-step logical reasoning.
The framework introduces an agent namedMind Map agentKnowledge Weaver, which can construct structured knowledge graphs to clearly track logical relationships, effectively improving deductive reasoning abilities. Additionally, by integrating web search and programming agents, the model can obtain information and perform computational analysis in real time, thereby enhancing reasoning accuracy and decision-making quality.
Below is a polished version of your provided content in Chinese, retaining the professional English terminology and original meaning while adjusting it for a more suitable expression style for Chinese readers and providing a clear structure:
Overall Process
Experimental Evaluation and Case Analysis
The research team conducted the following experiments and case analyses, demonstrating the advantages ofAgentic ReasoningAgentic Reasoning:
Case One: Doctoral-Level Scientific Reasoning Task (GPQA)
Evaluation was performed on GPQA, a high-difficulty scientific reasoning task. It significantly outperforms existing leading RAG (Retrieval-Augmented Generation) systems and closed-source LLMs. 
Case Two: Domain-Specific Deep Research Tasks
In domain-specific deep research questions, Agentic Reasoning significantly enhances expert-level knowledge synthesis and reasoning accuracy. 
Case Three: Misleading Question Answering
The paper presents a special misleading question that caused most existing LLM models to produce incorrect answers, butAgentic ReasoningAgentic Reasoning successfully answered correctly. 
Application of Mind Map in the "Werewolf" Game
Code
The paper's code has been open-sourced: https://github.com/theworldofagents/Agentic-Reasoning