wonjKim/self-conditioned-fmlm

Self-Conditioned Flow Map Language Models via Fixed-Point Flows

This repository contains checkpoints for ELF, ELF*, and FMLM* introduced in the paper Self-conditioned Flow Map Language Models via Fixed-point Flows.

All models use the ELF-B backbone (12 layers, hidden 768, 12 heads, ~105M parameters).

Released Models

• ELF: Self-conditioned flow teacher.
• ELF*: Self-conditioning-free fixed-point denoiser (CDEQ).
• FMLM*: Few-step flow map (two-time denoiser).

Code and Usage

The official codebase for training, evaluation, and sampling instructions can be found on GitHub: 👉 self-conditioned-fmlm GitHub Repository

Citation

If you find this work useful, please consider citing:

@article{yoo2026selfconditioned,
    title={Self-Conditioned Flow Map Language Models via Fixed-Point Flows},
    author={Jaehoon Yoo and Wonjung Kim and Floor Eijkelboom and Chanhyuk Lee
            and Nicholas M. Boffi and Seunghoon Hong and Jinwoo Kim},
    journal={arXiv preprint arXiv:2607.00714},
    year={2026},
}