# Random remasking improves masked diffusion decoding in study

> An arXiv paper proposes a way to correct errors during masked diffusion generation, reporting stronger results on logic and coding tasks. Independent validation is not provided in the abstract.

Oossa · 2026-10-23 · https://oossa.com/en/random-remasking-improves-masked-diffusion-decoding-in-study

Researchers describe Random Remask-Denoise (R²D), a test-time method that sometimes masks generated tokens again before denoising them. The paper says this lets a model revisit decoding errors without fine-tuning or reward feedback. The authors report more than 97% accuracy on hard Sudoku puzzles and more than 99% on 3-SAT; they also say the method beat iterative-correction baselines and LLaDA2.1-mini’s official Q-mode accuracy on HumanEval+. These are results reported in the paper’s abstract, and it does not provide independent validation there.

## The facts

- The paper proposes R²D for inference and Trajectory Correction Training for training denoisers on their own generation errors.
- The arXiv record lists the paper as submitted on October 22, 2026.

## Why it matters

Researchers working on masked diffusion models can investigate whether remasking improves generated code and logic answers, but the abstract does not establish how well the results generalize beyond the reported tests.

## Sources & references

1. [Random Remasking Scales Masked Diffusion Inference](https://arxiv.org/abs/2610.30744) – arXiv cs.LG, 2026-10-23

Last updated: 2026-10-23