# Researchers improve watermark detection for diffusion language models

_Published Monday, October 26, 2026 at 5:00 PM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report a watermarking method that detected at least 96% of responses from LLaDA-8B-Instruct on WaterBench at a 1% false-positive rate, while matching the LLM-judge score of unwatermarked text. The preprint addresses diffusion language models, which can choose words before the preceding text is settled.

The method learns both the token groups used to embed a watermark and a predictor of how token choices affect neighboring positions. Its detector counts each distinct context-token pair once, preventing repetition from increasing the score. The reported quality comparison uses an LLM judge; the authors also report no significant loss in accuracy on code and question-answering tasks.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2610.32205)

---
Canonical: https://techandbusiness.org/newswire/cz8V-Dag10-mWI79UsVfpF
Published: 2026-10-26T21:00:08.559Z
Story chronology: 2026-10-26T04:00:00.000Z
Retrieved: 2026-10-26T23:40:33.024Z
Publisher: Tech & Business (techandbusiness.org)