Sampling from Flow Language Models via Marginal-Conditioned Bridges

Official implementation of MCB sampler for flow language models, based on david3684/flm.

We propose a new sampler for FLMs which exploits additional probabilistic structure of the denoiser, and admits temprature scaled and nucleus (top-p) sampling.

Setup

Install dependencies with uv:

uv sync
# install flash attention
uv pip install flash-attn==2.8.3 --no-build-isolation

Running Experiments

To reproduce the experiments, run notebooks in notebooks/ folder.

Reference

@article{azangulov2026mcb,
  title={Sampling from Flow Language Models via Marginal-Conditioned Bridges},
  author={Iskander, Azangulov and Leo Zhang},
  journal={arXiv preprint arXiv:2605.13681},
  year={2026}
}