ATHENA-R1-Qwen3-8B/README.md

---
license: mit
language:
- en
base_model: Qwen/Qwen3-8B
library_name: transformers
pipeline_tag: text-generation
tags:
- agent
- biomedical
- treatment-reasoning
- tool-use
- reinforcement-learning
- qwen3
---

# ATHENA-R1-Qwen3-8B

**Project page:** [athena.openscientist.ai](https://athena.openscientist.ai/) · **Code:** [mims-harvard/ATHENA](https://github.com/mims-harvard/ATHENA)

**ATHENA-R1** is an AI agent for treatment reasoning, trained through
reinforcement learning over a universe of 212 biomedical tools. It performs
*multi-step* reasoning — identifying what evidence is needed, selecting tools,
and incorporating retrieved evidence into subsequent steps — with tool calls
served through the
[ToolUniverse](https://github.com/mims-harvard/ToolUniverse) (FDA labeling,
Open Targets, ChEMBL, EuropePMC, etc.).

Given a clinical question, the model performs *multi-step* tool calls,
synthesises the evidence, and returns a free-form answer grounded in
authoritative biomedical sources.

## Quick start

The model is exposed through the
[`athena-r1`](https://github.com/mims-harvard/ATHENA) Python package,
which handles the tool-call protocol and conversation management. Two
services back the agent: vLLM (model server) and ToolUniverse (tool server).

```bash
# 1. Install
pip install "athena-r1[vllm,web] @ git+https://github.com/mims-harvard/ATHENA.git"

# 2. Start backing services
bash scripts/launch_tooluniverse.sh         # → :8080
bash scripts/launch_vllm.sh 8000 mims-harvard/ATHENA-R1-Qwen3-8B

# 3. Run the agent (Python)
python -c "
from athena_r1 import AthenaR1
agent = AthenaR1(
    model='mims-harvard/ATHENA-R1-Qwen3-8B',
    vllm_url='http://0.0.0.0:8000/v1',
    tool_server='http://0.0.0.0:8080',
)
print(agent.answer('Dose adjustment for metformin in CKD eGFR 35?').answer)
"
```

For a chat UI (bundled browser demo with live-streamed reasoning):

```bash
python web/agui_server.py    # → http://localhost:8090/ (AG-UI server + demo)
```

For an OpenAI-compatible API endpoint:

```bash
python web/openai_server.py    # → http://localhost:9000/v1/chat/completions
```

## Inference settings (paper-canonical)

| Parameter | Value |
|---|---|
| temperature | 0.7 |
| top_p | 0.95 |
| top_k | 20 |
| min_p | 0.0 |
| presence_penalty | 0 |
| max_round | 40 |
| concurrent Qs | 4 |

## Evaluation

Open-ended setting: each question is answered free-form, then mapped to one of
the original answer choices.

| Benchmark | n | ATHENA-R1 | GPT-5 |
|---|---|---|---|
| DrugPC (open-ended drug reasoning) | 3,168 | **94.7%** | 76.9% |
| TreatmentPC (patient-specific treatment) | 456 | **82.9%** | 72.2% |

ATHENA-R1 exceeds GPT-5 by 17.8 points on DrugPC and 10.7 on TreatmentPC.

See the
[`docs/eval_results.md`](https://github.com/mims-harvard/ATHENA/blob/main/docs/eval_results.md)
file in the code repo for the full benchmark tables and the two-level
self-learning ablation.

## How it works

1. **Stage 1 — multi-step tool reasoning**: the model emits
   `<tool_call>...</tool_call>` blocks; the runtime dispatches them through
   ToolUniverse, appends results to the conversation, and re-prompts. Loop
   continues until `[FinalAnswer]` or `max_round` is hit.
2. **Stage 2 (eval only) — option mapping**: a *separate* function call
   maps the free-form answer to an MCQ letter. Two backends supported:
   the local ATHENA-R1 model (self-extraction) or Azure GPT-5
   (external reader).

## Intended use

ATHENA-R1 is a research artifact for treatment-reasoning research and
decision support. It is not a medical device and must not be used for direct
patient care.

## Citation

```bibtex
@article{gao2026athena,
  title   = {An AI agent for treatment reasoning over a biomedical tool universe},
  author  = {Gao, Shanghua and ... and Zitnik, Marinka},
  journal = {arXiv preprint},
  year    = {2026}
}
```

## License

MIT.

## Acknowledgements

Evidence retrieval is powered by
[ToolUniverse](https://github.com/mims-harvard/ToolUniverse), a library of
curated biomedical tools.