Agentic Reasoning: Reasoning LLMs with Tools for the Deep Research

Agentic Reasoning augments a reasoning language model with three external capabilities: web search, code execution, and a structured Mind-Map memory. The system lets the model invoke these tools during reasoning rather than relying only on internal knowledge. The paper evaluates the approach on expert-level scientific question answering and deep-research tasks, reporting gains over direct reasoning and retrieval-based baselines.

Deep research tasks require a model to gather changing information, perform computation, maintain intermediate context, and combine evidence over long reasoning chains. The paper frames tool use as part of the reasoning process itself: the reasoning model can decide when external information or computation is needed and then incorporate the resulting observation back into the chain.

A reasoning model can be limited by static knowledge, insufficient computation, and loss of coherence across long multi-step tasks. Retrieval added only outside the reasoning loop can also fail to provide information at the point where it is needed. The paper therefore asks how a reasoning model can selectively access tools and structured memory while preserving the main reasoning process.

The work sits at the intersection of reasoning models, retrieval-augmented generation, tool use, and agentic systems. Unlike conventional RAG, where retrieval is commonly a separate preprocessing or generation step, Agentic Reasoning allows search and computation to occur during reasoning. Its structured Mind-Map memory retains entities and relationships from the reasoning process so earlier information can be revisited without reproducing the entire chain.

The framework integrates three specialized agents into the reasoning loop. The Web-Search agent retrieves and summarizes external information, the Coding agent handles executable computation, and the Mind-Map agent converts relevant reasoning context into a structured graph of entities and relationships. The main reasoning model remains responsible for deciding when these capabilities are needed and for integrating their outputs into the solution.

The execution path is a closed loop: the reasoning model generates a tool invocation, the selected agent performs the operation, the result is returned to the reasoning context, and the model continues from the new evidence. The Mind-Map provides a structured representation of useful entities and relations. This separates retrieval, computation, and memory operations from the core reasoning model while keeping them available at test time.

The primary evaluation includes GPQA, a PhD-level science benchmark spanning physics, chemistry, and biology. The paper reports results on GPQA Diamond and the broader GPQA Extended Set, alongside additional evaluations for GAIA and deep-research tasks. The deep-research evaluation includes FreshWiki-based article-generation tasks and human assessment of generated articles.

The framework is primarily a test-time augmentation of a reasoning model rather than a new end-to-end training objective. The current paper version deploys the system with DeepSeek-R1 and delegates search, coding, and structured memory to specialized agents. The paper also uses dedicated models inside those agents, including DeepSeek-V3 for parts of search and Mind-Map processing and Claude-3.5-Sonnet for coding, with execution returning results to the main reasoning loop.

Baselines span direct reasoning, retrieval-augmented approaches, search-oriented systems, larger general-purpose toolboxes, and proprietary research systems where the comparison is available. The paper specifically studies alternatives including RAG and search variants and compares tool selection against Hugging Face’s seven-tool agent toolbox and LangChain’s much larger toolbox.

The ablations test both the number of tools and the composition of the proposed toolbox. The three-tool combination—Web Search, Coding, and Mind-Map—performs best in the reported comparisons. Pairing tools produces complementary gains, while larger generic toolboxes can degrade performance because extra capabilities increase inappropriate tool-selection risk or duplicate capabilities already available in the reasoning model. Within the Web-Search agent, query breakdown, reranking, and Mind-Map reasoning context each improve the search pipeline; the paper also reports that Knowledge Refinement becomes redundant when combined with those components.

The reported gains depend on external retrieval and execution, so failures in search quality or tool outputs can propagate into the reasoning chain. The system also adds inference cost and latency because a single task may trigger multiple model and tool calls. Benchmark results do not by themselves establish reliability in unrestricted production settings, and the paper leaves open questions around tool-invocation reliability, reproducibility of the orchestration protocol, and robustness to misinformation in retrieved sources.

Agentic Reasoning shows a practical pattern for extending reasoning models: keep the core reasoner focused on inference while giving it selective access to search, computation, and structured memory. The main systems lesson is that capability comes from the interaction between the model and its tools, not from the model in isolation.

Primary paper: Junde Wu, Jiayuan Zhu, Yuyuan Liu, “Agentic Reasoning: Reasoning LLMs with Tools for the Deep Research,” arXiv:2502.04644. The work was subsequently published at ACL 2025 as “Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning with Agentic Tools.”