WeaveMark

Properties

category: watermarking
maturity: emerging
approach: Multi-bit LLM text watermarking scheme that spreads a coded payload across multiple bits per token, uses a soft-decision error-correcting code to improve extraction accuracy under edits and paraphrase, and applies an unbiased multilayer reweighting of the token-sampling distribution to preserve text quality while embedding a longer, user-identifiable message than prior multi-bit schemes.
access: open-source
owns: An MIT-licensed reference implementation (GitHub) of a multi-bit LLM watermarking scheme, including the watermarking library, generation/detection scripts, attack simulations, and evaluation tools for downstream tasks like summarization and translation.
distinctFrom: Unlike synth-id (Google DeepMind's tournament-sampling text watermark, used by Gemini and now Claude) or claude-text-watermark (Anthropic's own SynthID-Text-derived implementation), WeaveMark is a separate, independently-published academic scheme whose specific claim is improved multi-bit payload capacity and extraction accuracy under edits via coded payload spreading and error-correcting decoding, benchmarked directly against a prior multi-bit scheme (BiMark) rather than a single-bit detection-only watermark.
source: arXiv:2609.02177, WeaveMark: Robust and Scalable Multi-bit LLM Watermarking via Coded Payload Spreading
notes: Submitted 2026-09-02, two days before this lens's 2026-09-04 window start; included as a gap-fill following the same precedent as previously-tracked pre-window entries (audio-watermark-fragility-probe at two days pre-window, nvidia-synthetic-video-detector at two weeks). Code confirmed on GitHub (qkrrkd90-source/WeaveMark), MIT licensed, 0 stars / 0 forks at time of check, so real-world adoption is effectively nil despite the working code release. Vendor/author-reported result '89.8% match rate for 32-bit messages at 200 tokens, compared with 20.8% for BiMark' is the paper's own benchmark, not independently reverified here. Authors: Gang-Hyun Park, Ju-Hyeong Lee, Hee-Youl Kwak, Dae-Young Yun.

Provenance

Source: https://arxiv.org/abs/2609.02177

Properties

category

watermarking

maturity

emerging

approach

Multi-bit LLM text watermarking scheme that spreads a coded payload across multiple bits per token, uses a soft-decision error-correcting code to improve extraction accuracy under edits and paraphrase, and applies an unbiased multilayer reweighting of the token-sampling distribution to preserve text quality while embedding a longer, user-identifiable message than prior multi-bit schemes.

access

open-source

owns

An MIT-licensed reference implementation (GitHub) of a multi-bit LLM watermarking scheme, including the watermarking library, generation/detection scripts, attack simulations, and evaluation tools for downstream tasks like summarization and translation.

distinctFrom

Unlike synth-id (Google DeepMind's tournament-sampling text watermark, used by Gemini and now Claude) or claude-text-watermark (Anthropic's own SynthID-Text-derived implementation), WeaveMark is a separate, independently-published academic scheme whose specific claim is improved multi-bit payload capacity and extraction accuracy under edits via coded payload spreading and error-correcting decoding, benchmarked directly against a prior multi-bit scheme (BiMark) rather than a single-bit detection-only watermark.

source

arXiv:2609.02177, WeaveMark: Robust and Scalable Multi-bit LLM Watermarking via Coded Payload Spreading

notes

Submitted 2026-09-02, two days before this lens's 2026-09-04 window start; included as a gap-fill following the same precedent as previously-tracked pre-window entries (audio-watermark-fragility-probe at two days pre-window, nvidia-synthetic-video-detector at two weeks). Code confirmed on GitHub (qkrrkd90-source/WeaveMark), MIT licensed, 0 stars / 0 forks at time of check, so real-world adoption is effectively nil despite the working code release. Vendor/author-reported result '89.8% match rate for 32-bit messages at 200 tokens, compared with 20.8% for BiMark' is the paper's own benchmark, not independently reverified here. Authors: Gang-Hyun Park, Ju-Hyeong Lee, Hee-Youl Kwak, Dae-Young Yun.