Don't Settle at the Mode! Mitigating Diversity Collapse in Pretrained Flow Models via Feature Self-Guidance

Researchers have introduced two new methods to improve flow models and language generation, addressing issues of diversity collapse and parallel generation limitations, with papers submitted in 2026^[1][2].

State-of-the-art flow models suffer from diversity collapse when generating multiple samples under the same conditioning, according to a paper submitted to arXiv on June 25, 2026^[1]. To address this, researchers introduced a self-guidance mechanism that disperses internal features during batch generation and projects them back onto the data manifold, ensuring diverse generation without sacrificing alignment with input conditions. This method integrates seamlessly into pretrained flow models as a plug-and-play module, adding only marginal inference cost. Experiments demonstrated significant improvements in diversity while preserving fidelity across several conditional flow models. Another paper, also submitted in 2026, introduced Masked Language Flow Models (MLFMs) to improve language generation by incorporating masking into Flow Language Models (FLMs)^[2]. MLFMs enable conditional generation via continuous flows and allow pretrained Masked Diffusion Models (MDMs) to be converted into MLFMs through a simple adaptation. A novel sampler was proposed that alternates continuous denoising with discrete unmasking of confident tokens. Flow-based language models can be scaled to solve downstream reasoning and instruction-following tasks.