LP-SFT: Keeping Plausible Alternatives Alive in Supervised Fine-Tuning

Research overview

LP-SFT separates the supervised token from other plausible candidates and preserves relative preferences among the alternatives. It studies adaptation quality alongside retention of sampling-accessible solutions.

LP-SFT was previously titled “Local-Preserving Supervised Fine-Tuning via Multimodal Entropy Structure”. This record follows arXiv v4 (October 3, 2026), including its updated author list.

Metadata checked against arXiv and Google Scholar on October 6, 2026. This overview summarizes the work; consult the linked paper for results and limitations.

Citation

@misc{lpsft2026,
  title={LP-SFT: Keeping Plausible Alternatives Alive in Supervised Fine-Tuning},
  author={Wang, Yueyang and Bi, Baolong and Lu, Shuo and Zhang, Jingyuan and Shi, Jiajun and Zhang, Yuxuan},
  year={2026},
  eprint={2607.04733},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2607.04733}
}