README.md

metadata

language:
  - en
license: apache-2.0
task_categories:
  - text-generation
pretty_name: BeliefTrack
configs:
  - config_name: task_a_7b
    data_files:
      - split: train
        path: task_a_7b/train.json
      - split: test
        path: task_a_7b/test.json
  - config_name: task_a_9b
    data_files:
      - split: train
        path: task_a_9b/train.json
      - split: test
        path: task_a_9b/test.json
  - config_name: task_b_7b
    data_files:
      - split: train
        path: task_b_7b/train.json
      - split: test
        path: task_b_7b/test.json
  - config_name: task_b_9b
    data_files:
      - split: train
        path: task_b_9b/train.json
      - split: test
        path: task_b_9b/test.json

BeliefTrack

📄 arXiv • 🤗 HFPaper • 🤗 HF Collection

BeliefTrack is a closed-world benchmark for Contextual Belief Management (CBM) in multi-turn language model interactions. Each example asks a model to maintain the set of hypotheses that remain consistent with formal evidence.

Citation

@article{xu2026whenshouldmodelschange,
  title={When Should Models Change Their Minds? Contextual Belief Management in Large Language Models},
  author={Xu, Haoming and Xu, Weihong and Li, Zongrui and Wang, Mengru and Yao, Yunzhi and Wu, Chiyu and Shang, Jin and Gong, Yu and Deng, Shumin},
  journal={arXiv preprint arXiv:2605.30219},
  year={2026}
}