ToolRAG-T1-GTE-Qwen2-1.5B model selects suitable candidates from ToolUniverse based on descriptions provided by TxAgent. ToolRAG-T1-GTE-Qwen2-1.5B utilizes gte-Qwen2-1.5B-instruct as the base model, and is fine-tuned on pairs of requirements and tool descriptions using the multiple negatives ranking loss.

Precision therapeutics require multimodal adaptive models that generate personalized treatment recommendations. We introduce TxAgent, an AI agent that leverages multi-step reasoning and real-time biomedical knowledge retrieval across a toolbox of 211 tools to analyze drug interactions, contraindications, and patient-specific treatment strategies.

TxAgent outperforms leading LLMs, tool-use models, and reasoning agents across five new benchmarks: DrugPC, BrandPC, GenericPC, TreatmentPC, and DescriptionPC, covering 3,168 drug reasoning tasks and 456 personalized treatment scenarios.

By integrating multi-step inference, real-time knowledge grounding, and tool- assisted decision-making, TxAgent ensures that treatment recommendations align with established clinical guidelines and real-world evidence, reducing the risk of adverse events and improving therapeutic decision-making.

ToolRAG model is used in both TxAgent inference process and the training data collection phase. It utilizes gte-Qwen2-1.5B-instruct as the base model, which is fine-tuned on pairs of requirements and tool descriptions using the multiple negatives ranking loss. In the TOOL PROVIDER module, we use the ToolRAG model to identify tools beyond the initial list retrieved from the reference information of the question. While the ToolRAG model requires training data from reasoning traces, we propose an iterative training process for ToolRAG model, where it is trained on the generated reasoning traces, which in turn helps improve the generation of future reasoning traces. In the first stage, since the ToolRAG model model is not yet available, we rely solely on the initial set of tools that are obtained from the reference information of the question, to generate the reasoning trace. From this trace, we extract pairs of tool requirements and tool descriptions, which are then used to train the ToolRAG model. In the second stage, after the initial training of ToolRAG model, we use it to select tools instead of relying exclusively on the initial set of tools. This approach allows the reasoning trace to better reflect real-world use cases, as tools are now retrieved directly by the ToolRAG model agent. Using the data collected from this stage, we continue to gather new pairs for further training of the ToolRAG model. This process is repeated iteratively, continually refining both the ToolRAG model and the quality of reasoning trace generation.

Please refer to our project page for more details: Project Page.

ToolRAG model is used by the TxAgent.

Install ToolUniverse:

# Install from source code:
git clone https://github.com/mims-harvard/ToolUniverse.git
cd ToolUniverse
python -m pip install . --no-cache-dir

# Install from pypi:
pip install tooluniverse

Install TxAgent:

# Install from source code:
git clone https://github.com/mims-harvard/TxAgent.git

python -m pip install . --no-cache-dir
# Install from pypi:
pip install txagent

Run the example with the run_example.py.

Run the gradio demo with the run_txagent_app.py.

@misc{gao2025txagent,
      title={TxAgent: An AI Agent for Therapeutic Reasoning Across a Universe of Tools}, 
      author={Shanghua Gao and Richard Zhu and Zhenglun Kong and Ayush Noori and Xiaorui Su and Curtis Ginder and Theodoros Tsiligkaridis and Marinka Zitnik},
      year={2025},
      eprint={2503.10970},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2503.10970}, 
}

If you have any questions or suggestions, please email Shanghua Gao and Marinka Zitnik.