This repository contains the NRP (Necessary Reasoning Prefix) detector model used in the DECS algorithm, as presented in the paper Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling.

The NRP detector is designed to determine whether a given reasoning chunk contains the ground truth signal, enabling surgically precise token-level rewards to reduce "overthinking" in reasoning models.

Project Page: https://pixas.github.io/decs-iclr26-site/
Repository: https://github.com/pixas/DECS
Paper: arXiv:2509.25827

Usage

According to the official repository, you can deploy the NRP detector using vLLM:

vllm serve --model pixas/DECS_NRP_DETECTOR --port 10041

Citation

If you use this model, please cite the following work:

bibtex

@inproceedings{jiang2026decs,
  title     = {Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling},
  author    = {Jiang, Shuyang and Tao, Xiaofeng and Zhang, Kui and Xiao, Yanghua},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  note      = {Oral},
  url       = {https://arxiv.org/abs/2509.25827}
}