YongqiLi/R-EMID-Estimator

R-EMID Estimator

This is a LoRA adapter for computing the R-EMID (Reasoning-based Effective Mutual Information Difference) metric, introduced in the ACL 2026 Findings paper "Understanding Generalization in Role-Playing Models via Information Theory".

Model Description

The R-EMID Estimator is trained via CoRL (Co-Evolving Reinforcement Learning), the method proposed in the paper. It is used to estimate the R-EMID metric, which measures the generalization performance of role-playing models (RPMs) from an information-theoretic perspective.

This model is intended for evaluation and metric computation rather than direct role-playing.

Citation

@inproceedings{li-2026-RPMG,
  title     = {Understanding Generalization in Role-Playing Models via Information Theory},
  author    = {Yongqi Li and Hao Lang and Fei Huang and Tieyun Qian and Yongbin Li},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026}
}