LaTER-14B

LaTER-14B is a latent-reasoning language model trained from Qwen3-14B. The model modifies the original Qwen3-14B architecture for LaTER-style latent exploration and explicit verification, then trains the modified architecture to support efficient test-time reasoning.

This checkpoint is not intended to be loaded as a vanilla Qwen3ForCausalLM model. Please refer to the LaTER code examples at https://github.com/TioeAre/LaTER for the required model implementation and inference usage.

Model Details

• Base model: Qwen/Qwen3-14B
• License: Apache-2.0

Usage

Use the model with the custom LaTER implementation rather than the default Qwen3 Transformers class. Example code and any required custom model registration should follow the repository:

https://github.com/TioeAre/LaTER

Intended Use

This model is intended for research on efficient test-time reasoning, latent exploration, and explicit verification.

Citation

@misc{li2026later,
      title={LaTER: Efficient Test-Time Reasoning via Latent Exploration and Explicit Verification},
      author={Xuan Li and Yining Wang and Yuchen Liu and Guanjun Liu and Delai Qiu and Shengping Liu and Jiaen Liang and Wei Huang and Jun Yu and Junnan Zhu},
      year={2026},
      eprint={2605.07315},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.07315},
}