Instructions to use CMU-AIRe/MRT-online with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use CMU-AIRe/MRT-online with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="CMU-AIRe/MRT-online")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("CMU-AIRe/MRT-online")
model = AutoModelForCausalLM.from_pretrained("CMU-AIRe/MRT-online")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use CMU-AIRe/MRT-online with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "CMU-AIRe/MRT-online"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CMU-AIRe/MRT-online",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/CMU-AIRe/MRT-online

SGLang

How to use CMU-AIRe/MRT-online with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "CMU-AIRe/MRT-online" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CMU-AIRe/MRT-online",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "CMU-AIRe/MRT-online" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CMU-AIRe/MRT-online",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use CMU-AIRe/MRT-online with Docker Model Runner:
```
docker model run hf.co/CMU-AIRe/MRT-online
```

MRT-online (R1-Distill-Qwen-1.5B)

DeepSeek-R1-Distill-Qwen-1.5B fine-tuned with the on-policy variant of Meta Reinforcement Fine-Tuning (MRT). Where the offline variant uses an off-policy prefix and a single end-of-trace progress bonus, MRT-online generates the reasoning trace on-policy, segments it into episodes online, forks short forced-termination branches at each episode boundary, and assigns a per-episode dense progress reward. This is the "branched rollouts from a meta-prover policy" direction left as an open problem in the paper.

📄 Paper: Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning
💻 Code / recipe: CMU-AIRe/MRT → train/rl/ (rollout_mrt_online.py, REPRODUCTION.md)
🔁 Offline variant: CMU-AIRe/MRT-offline

v0.1 / research variant. Produced by the open-source v0.1 training code on miles. This is an exploratory on-policy variant, not the configuration reported in the paper.

Evaluation

pass@1 (mean of 64 samples/problem) at a 16K token budget, averaged over AIME 2024 / AIME 2025 / AMC 2023 / MinervaMATH / MATH500:

model	AIME24	AIME25	AMC23	Minerva	MATH500	Avg	gain over base
base (R1-Distill-Qwen-1.5B)	27.34	22.86	67.89	24.94	81.71	44.95	—
GRPO (outcome-reward)	28.12	22.97	67.77	26.45	81.85	45.43	+0.48
MRT-online (this model)	28.59	22.24	68.79	25.87	82.37	45.57	+0.62
MRT-offline (for reference)	28.75	23.59	70.86	24.96	82.61	46.16	+1.20

Finding: MRT-online improves over outcome-reward GRPO (+0.62 vs +0.48) but trails the offline single-scalar MRT (+1.20). The on-policy per-episode signal helps, but with only G=4 termination branches per boundary the per-episode credit is higher-variance — empirical support for the off-policy single-scalar form the paper adopts. A larger branch count or longer training may narrow the gap.

Training

Base: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B; data: 4,000 NuminaMath problems.
On-policy GRPO + per-episode progress reward (α=1.0, ≤6 episodes/rollout, G=4 branches), 248 optimizer steps, 16K budget, temp 0.9.
Framework: miles (Megatron-LM + SGLang). Recipe: CMU-AIRe/MRT train/rl/REPRODUCTION.md.

Citation

@misc{qu2025optimizingtesttimecomputemeta,
      title={Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning},
      author={Yuxiao Qu and Matthew Y. R. Yang and Amrith Setlur and Lewis Tunstall and Edward Emanuel Beeching and Ruslan Salakhutdinov and Aviral Kumar},
      year={2025}, eprint={2503.07572}, archivePrefix={arXiv}, primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.07572},
}

Downloads last month: 15

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for CMU-AIRe/MRT-online

Base model

deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

Finetuned

(640)

this model

Paper for CMU-AIRe/MRT-online

Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuning

Paper • 2503.07572 • Published Mar 10, 2025 • 48