Agentic Reasoning: Oxford’s New Framework for Enhanced Reasoning

Published on: February 6, 2025

Agent Application

Agentic Reasoning: Reasoning LLMs with Tools for the Deep Research

We propose a new framework called Agentic Reasoning, which enhances the reasoning capabilities of large language models (LLMs) by integrating external tool agents. Unlike traditional LLM reasoning methods that rely solely on internal reasoning, the Agentic Reasoning framework combines web search, code execution, and structured reasoning context memory, effectively solving complex problems that require deep research and multi-step logical reasoning. Our framework introduces a Mind Map agent, which tracks logical relationships by constructing structured knowledge graphs, thereby enhancing deductive reasoning capabilities. Additionally, integrating web search and code execution agents enables real-time information retrieval and computational analysis, further improving reasoning accuracy and decision-making capabilities. Evaluations conducted on PhD-level scientific reasoning tasks (GPQA) and domain-specific deep research tasks show that our approach significantly outperforms existing models, including leading Retrieval-Augmented Generation (RAG) systems and closed-source LLMs. Moreover, our experimental results indicate that the Agentic Reasoning framework achieves significant improvements in expert-level knowledge integration, reasoning scalability, and structured problem-solving. Code repository: https://github.com/theworldofagents/Agentic-Reasoning.

https://arxiv.org/abs/2502.04644

Add Note Agent

If unable to add, please + vx: iamxxn886

Why is Agentic Reasoning Needed?

The reasoning capabilities of large language models (LLMs) have always faced a “ceiling of thought.” LLMs are like students who can only take closed-book exams, relying solely on the knowledge they memorized during pre-training for internal reasoning. When faced with problems that require real-time data verification, complex calculations, or interdisciplinary knowledge integration, this single reasoning mode exposes three fatal flaws:

1. 1. Information Island Dilemma: The model cannot actively retrieve the latest information and can only provide outdated answers when confronted with time-sensitive questions (e.g., “What are the research breakthroughs of the 2024 Nobel Prize winner in Chemistry?”).

2. 2. Logical Gap Crisis: When dealing with multi-step reasoning, the model often loses the relevance of intermediate conclusions. This is akin to forgetting previous derivations when solving a math problem, leading to deviations in the final conclusion.

3. 3. Tool Usage Barriers: Human experts naturally call upon calculators, search engines, mind maps, and other tools when solving problems, but traditional LLMs lack the capability to dynamically invoke external tools.

The Oxford team found that even advanced models like DeepSeek-R1, which achieved a 90% accuracy in mathematical problem-solving, saw the completeness of logical explanation drop to 40% when faced with open-ended social science questions. This phenomenon of “strong performance in science but poor performance in humanities” exposes structural flaws in existing reasoning paradigms.

What is Agentic Reasoning?

The Agentic Reasoning framework creatively integrates three functional AI agents into the reasoning process, constructing a human-like tool-using thought chain:

The Core Trio

1. 1. Mind Map Agent: Acts as the model’s “thinking whiteboard,” constructing knowledge graphs in real-time. When the model analyzes “the impact of global climate change on the coffee industry,” this agent will automatically create a node network containing [Climate Model]-[Precipitation Changes]->[Coffee Bean Yield]->[Futures Prices], ensuring the logical relationships are visualized.

2. 2. Web Search Agent: Serves as the model’s “smart browser,” equipped with precise information retrieval capabilities. Unlike simply invoking a search engine, it can dynamically adjust search keywords based on the reasoning progress. For example, when verifying “the latest breakthroughs in CRISPR gene editing,” it will first search for top journal papers, then look for clinical trial data.

3. 3. Code Agent: Functions as the model’s “calculator + laboratory,” supporting real-time execution of Python code. When calculating “the impact of improved photovoltaic cell conversion efficiency on carbon reduction,” it can automatically write calculation scripts and call NumPy for quantitative analysis.

The three agents work collaboratively through a central control, forming an enhanced loop of “retrieve-calculate-organize-reason.” Tests show that this architecture improves the model’s accuracy by 35% on the GPQA PhD-level scientific question database, especially on interdisciplinary biology questions, where the score jumped from 62% to 79%.

Four Breakthrough Advantages of the Framework

1. Dynamic Knowledge Enhancement

Traditional RAG technology is like a fixed menu, only able to retrieve information according to a predetermined process. The search agent of Agentic Reasoning can achieve “intelligent ordering,” dynamically adjusting retrieval strategies based on the reasoning phase. When handling “the impact of quantum computing on cryptography,” it initially retrieves basic concepts, then locates the latest papers, and finally searches for industry application cases, forming a progressive knowledge acquisition.

2. Logical Visualization Tracking

The knowledge graph constructed by the Mind Map agent allows the model’s “thinking process” to have interpretability for the first time. When solving complex physics problems, the system will generate a logical network containing [Conservation of Energy]-[Momentum Theorem]->[Collision Coefficient], not only improving reasoning accuracy but also providing developers with a “thinking X-ray” for debugging the model.

3. Real-time Calculation Verification

The integration of the code agent breaks the limitations of “talking without action.” When the model deduces “a certain economic model’s prediction results,” it will immediately write Python code for Monte Carlo simulations, using the computational results to validate theoretical deductions. Tests show that this real-time verification mechanism reduces the error rate of quantitative analysis problems by 58%.

4. Memory Reinforcement Learning

The framework’s “structured memory” function enables the model to have continuous learning capabilities. When continuously processing related questions, the Mind Map will automatically merge knowledge nodes. For instance, after completing research on “the mutation mechanism of the COVID-19 virus,” related knowledge points will be integrated into a larger “epidemiology” knowledge graph, providing a cognitive foundation for subsequent tasks.

Three Innovations

1. Standardized Tool Invocation Interface

The framework abstracts tool invocation into a unified API, allowing developers to easily integrate new tools. For example, when adding a chemical equation solver, it only requires packaging according to the “input parameters-output structure” standard, and it can immediately be incorporated into the existing reasoning process.

2. Dynamic Allocation of Cognitive Load

The intelligent task allocation mechanism ensures the rational use of computational resources. Simple retrieval tasks are handled by lightweight agents, while complex calculations are assigned to dedicated code agents. In testing scenarios for medical diagnosis, this mechanism improved response speed by 40%, while reducing computational power consumption by 72%.

3. Multimodal Memory Fusion

The Mind Map agent supports multimodal storage of text, formulas, and charts. When dealing with problems like “the black hole information paradox,” it can save the mathematical derivation of Hawking radiation while also relating to the latest observational data charts, forming a three-dimensional knowledge system.

Code Implementation

The Oxford team has open-sourced the framework implementation on GitHub, providing developers with a scalable infrastructure.

GitHub repository: https://github.com/theworldofagents/Agentic-Reasoning

The core code adopts a modular design, mainly consisting of:

class AgenticReasoner:
    def __init__(self):
        self.mind_map = KnowledgeGraph()
        self.search_agent = SmartSearch()
        self.code_agent = PythonExecutor()

    def reason(self, query):
        # Multi-agent collaborative workflow
        search_results = self.search_agent.retrieve(query)
        self.mind_map.update(search_results)
        calculations = self.code_analyze(search_results)
        return self.synthesize(calculations)

Developers can quickly create new agents by inheriting the BaseAgent class. For example, when adding a PDF parsing agent for papers, it only needs to implement the document parsing logic, and the framework will automatically handle task scheduling and memory storage.

Evolution Directions

The team is exploring a “federated learning” mechanism for agents, allowing different specialized agents to share cognitive results. In the future, a “biology agent” may synchronize its knowledge graph with a “chemistry agent,” achieving true interdisciplinary reasoning. In testing cases for drug development, this mechanism has tripled the efficiency of compound screening.

For LLM application developers, Agentic Reasoning is not just a tool framework, but also indicates the evolution direction of the next generation of AI systems—from single models to agent ecosystems, from static knowledge to dynamic cognition; the reasoning revolution may just be beginning.

• • Original paper: https://arxiv.org/abs/2502.04644

• • Get more updates on the latest Arxiv papers: https://github.com/HuggingAGI/HuggingArxiv!

• • Join the community, +v: iamxxn886

• • Click the public account menu to join the discussion