Fisher-R1-7B is an open-weight LLM agent for reliable hypothesis testing. It is post-trained from Qwen2.5-Coder-7B-Instruct on synthetic executable statistical tasks, using supervised fine-tuning followed by reinforcement learning with verified statistical rewards.

The model is designed to inspect data, select and execute an appropriate statistical test, report a p-value, and draw a conclusion. It is evaluated on P-Bench.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "May2222/Fisher-R1-7B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

For the training setup, evaluation protocol, and results, see Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing.

Citation

bibtex

@article{miao2026fisherr1,
  title   = {Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing},
  author  = {Miao, Jiacheng and Mu, Jin and Chen, Guanhua and Zou, James},
  journal = {arXiv preprint arXiv:2608.07437},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.07437}
}