RENT is an unsupervised method for training reasoning LLMs by minimizing entropy. We demonstrate on a variety of datasets and models that RENT improves model performance without using any ground truth labels!

RENT is featured in our paper "Maximizing Confidence Alone Improves Reasoning" (link)

RENT is built on top of the verl library.

[10/31]: we release a checkpoint RENT-Qwen-7B of Qwen2.5-7B-Instruct trained with RENT on AIME24. The model can be found at this huggingface link

Please refer to the existing verl quickstart for installation : verl Installation

For the specific vllm version we used (along with other packages), see the requirements.txt file.

Prepare AIME data:

python ./examples/data_preprocess/aime.py --local_dir {path_to_your_dataset}

Run Training:

Adjust the configuration in ppo_trainer.yaml to match your desired training configuration (number of gpus, batch size, etc.). To override this config somewhere else, see "Creating Custom Configurations"

python -m verl.trainer.main_ppo exps="[grpo, entropy, format, sampleval, aime]" base_model=Qwen/Qwen2.5-7B-Instruct

See verl's documentation on how to prepare data and implement custom reward functions

• Data and Reward Preparation
  • Prepare Data for Post-Training
  • Implement Reward Function for Dataset

We use an extensible config setup, allowing you to override default configurations for specific tasks/jobs.

To define a custom configuration, create a new yaml file in verl/trainer/config/exps. NOTE: you MUST include # @package _global_ at the beginning of the file in order to override other configs.

To use different configuration files, simply add them to the exps="[...]" argument to verl.trainer.main_ppo. Note: configurations are applied from left-to-right order, so configs to the right will override configs to the left!

@article{prabhudesai2025rent,
    title={Maximizing Confidence Alone Improves Reasoning},
    author={Prabhudesai, Mihir and Chen, Lili and Ippoliti, Alex and Fragkiadaki, Katerina and Liu, Hao and Pathak, Deepak},
    journal={arXiv preprint arXiv:2505.22660},
    year={2025}
}