Clin-REACT-31B

Model name in the manuscript: Clin-REACT 31B
Base model: google/gemma-4-31B-it
Release format: merged full-parameter checkpoint

Clin-REACT is a family of large language models fine-tuned for clinical reasoning using the ICU-REACT framework. The models are designed to reason over clinically relevant information and produce responses for tasks spanning ICU decision-making and broader clinical-reasoning benchmarks.

This repository contains the merged, self-contained checkpoint. The LoRA adapter used during supervised fine-tuning has been merged into the corresponding base-model weights for distribution and inference.

The underlying Gemma 4 architecture is multimodal. The Clin-REACT fine-tuning and benchmark results reported in this model card are text-based; inherited multimodal capabilities were not evaluated as part of these results.

Benchmark results

Clin-REACT 31B is compared with its Gemma 4 31B IT backbone and selected strong open models, including substantially larger baselines.

Scores below are the mean primary benchmark scores, reported as percentages. Higher is better. The macro average is the unweighted mean of the five benchmark primary scores. Bold indicates the best result within the comparison set shown for each benchmark.

The benchmarks contain 71 ICU-REACT cases, 174 SCT-Bench cases, 72 ER-Reason cases, 1,254 MedRBench cases, and 934 VivaBench cases. Because the benchmarks use different task formulations and primary scoring procedures, individual benchmark scores should primarily be interpreted within each benchmark.

Model	ICU-REACT (n=71)	SCT-Bench (n=174)	ER-Reason (n=72)	MedRBench (n=1254)	VivaBench (n=934)	Macro avg.
Clin-REACT 31B	44.8	75.5	51.4	46.9	33.5	50.4
Gemma 4 31B IT	40.6	77.6	47.4	44.0	31.6	48.3
Baichuan M2 32B	36.9	63.0	46.2	38.4	26.9	42.3
Llama 3.3 70B Instruct	30.9	59.9	45.2	45.1	26.7	41.6
GPT-OSS-120B	40.1	74.3	48.3	47.0	32.5	48.5

Installation

pip install -U torch transformers accelerate

Load and use the model

Gemma 4 uses a multimodal processor/model interface, including text-only prompts.

import torch
from transformers import AutoModelForMultimodalLM, AutoProcessor

MODEL_ID = "macontreras98/Clin-REACT-31B"

processor = AutoProcessor.from_pretrained(MODEL_ID)

model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    device_map="auto",
)

# Example ICU-REACT-style clinical reasoning sample
messages = [
    {
        "role": "system",
        "content": (
            "You are an ICU clinician. Write one coherent paragraph explaining "
            "which variables are most relevant to the decision and why."
        ),
    },
    {
        "role": "user",
        "content": (
            "Patient context: 75-year-old female ICU patient with recurrent high "
            "fevers and no clear infection source.\n\n"
            "Decision question: Is the frequency and severity of fever in this "
            "patient worrisome for ongoing infection or non-infectious etiology?"
        ),
    },
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

input_length = inputs["input_ids"].shape[-1]

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,
    )

generated_tokens = outputs[0][input_length:]
response = processor.decode(
    generated_tokens,
    skip_special_tokens=True,
)

print(response)

For exact benchmark reproduction, use the same prompting, processor/chat template, preprocessing, thinking-mode configuration, and decoding settings used in the ICU-REACT evaluation pipeline.

Intended use and limitations

Clin-REACT is released for research on clinical reasoning and medical AI. It is not a medical device and should not be used as a substitute for professional clinical judgment, diagnosis, treatment decisions, or other autonomous patient-care decisions.

Benchmark performance does not establish clinical safety, prospective effectiveness, or suitability for deployment. Users are responsible for evaluating the model for their own setting and for complying with applicable privacy, security, institutional, and regulatory requirements.

License

The base model google/gemma-4-31B-it is licensed under the Apache License 2.0. This repository distributes a modified derivative checkpoint. See the LICENSE file and retain any applicable upstream copyright and attribution notices.

Citation

If you use Clin-REACT in your research, please cite the associated ICU-REACT / Clin-REACT manuscript.

@article{contreras2026icureact,
  title   = {Teaching LLMs How ICU Physicians Approach Clinical Reasoning Through OMOP-Aligned Retrieval Improves Reasoning Across Clinical Domains},
  author  = {Contreras, Miguel and Siegel, Scott and Nerella, Subhash and Sena, Jessica and Zhang, Jiaqing and
             Sun, Heng and Akkaladevi, Hruday Tej and Lu, Peiyu and Rosen, Jordan and Kapoor, Sumit and
             Desaraju, Sasank and Thompson, Grace R. and Purcell, Jacob and Petrauskis, Michael and
             Hong, Philip KW and Brennan, Meghan and Chrabaszcz, Sarah and Smith, Tierra and Ren, Ronnie and
             Kabbash, Michel S. and Haziroglu, Ceyhun and Patel, Rushi and Gomez, Gabriel and Chaiklin, Charlotte and
             Leung, Randy and John, Kenneth N. and Wiggins, Whitman and Kayser, Philip and Bird, Vincent and
             Bruzzone, Maria and Loftus, Tyler J. and Bihorac, Azra and Rashidi, Parisa},
  journal = {arXiv preprint arXiv:2608.22622},
  year    = {2026},
  doi     = {10.48550/arXiv.2608.22622},
  url     = {https://arxiv.org/abs/2608.22622}
}