I loved the analysis in the CANDI (text diffusion) paper. Wonderful to see new works in this direction. However, beyond perplexity, we urgently need to start seeing benchmark results after post-training, and comparisons to autoregressive LLMs. We must measure the performance gap to know whether we’re making progress.

@IAzangulov: The main difference with CADD/CCDD/CANDI is that MLFM fully decouples masking from continuous corruption. This gives us several advantages:

1. We can decide when to promote tokens based on problem difficulty, rather than following a fixed unmasking rate.

2. Token promotion is a hard commitment, and MLFM makes the risk explicit: the error can be bounded directly and does not depend on the number of denoising steps.

3. This makes the design distillation-friendly: distilled flow steps can be combined with token promotion, so faster continuous updates do not force premature commitments.

Intuitively, online token promotion acts like signal amplification: it removes noise from resolved corrupted tokens. In spirit, this is closer to a self-conditioning mechanism.

We updated the blog post to discuss the comparison.