PostMark: A Robust Blackbox Watermark for Large Language Models

TL;DR

PostMark introduces a novel post-hoc watermarking approach that enables third-party watermarking of LLM outputs without access to model logits. It uses a semantic Embedder, a cryptographically mapped SecTable, and an Inserter to rewrite text with watermark words, enabling detection by a separate verifier with a presence score and a fixed false-positive threshold. Across multiple baselines, base models, and datasets, PostMark demonstrates strong robustness to paraphrasing—particularly on low-entropy, RLHF-aligned models—while maintaining text quality better than many logit-based watermarks. Human judgments indicate watermark words are hard to detect and watermarked text remains at least as coherent and relevant, though stronger watermarks reduce quality, highlighting a quality-robustness trade-off with practical implications for deployment and policy.

Abstract

The most effective techniques to detect LLM-generated text rely on inserting a detectable signature -- or watermark -- during the model's decoding process. Most existing watermarking methods require access to the underlying LLM's logits, which LLM API providers are loath to share due to fears of model distillation. As such, these watermarks must be implemented independently by each LLM provider. In this paper, we develop PostMark, a modular post-hoc watermarking procedure in which an input-dependent set of words (determined via a semantic embedding) is inserted into the text after the decoding process has completed. Critically, PostMark does not require logit access, which means it can be implemented by a third party. We also show that PostMark is more robust to paraphrasing attacks than existing watermarking methods: our experiments cover eight baseline algorithms, five base LLMs, and three datasets. Finally, we evaluate the impact of PostMark on text quality using both automated and human assessments, highlighting the trade-off between quality and robustness to paraphrasing. We release our code, outputs, and annotations at https://github.com/lilakk/PostMark.

Figures (5)

• Figure 1: The PostMark watermarking and detection procedure. Given some unwatermarked input text, we generate its embedding using the Embedder and compute its cosine similarity with all word embeddings in the SecTable, performing top-k selection and additional semantic similarity filtering to choose a list of words. Then, we instruct the Inserter to watermark the text by rewriting it to incorporate all selected words. During detection, we similarly obtain a watermark word list and check how many of these words are present in the input text.
• Figure 2: Pairwise preference human evaluation results on PostMark@12 and PostMark@6. For both configurations, the watermarked text is at least as good as its unwatermarked counterpart the majority of the time in all aspects.
• Figure 3: Watermark word frequency distribution over 500 OpenGen outputs. The majority of the words are chosen as watermark words less than 5% of the time. There are only two major hub words that are selected more than 20% of the time.
• Figure 4: Human annotation interface for the pairwise comparison task.
• Figure 5: Human annotation interface for the watermark word identification task.