OpenMarkX: An Extensible Platform for Watermarking Open-Source Model and Agent Outputs
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: We're looking at the paper: "OpenMarkX: An Extensible Platform for Watermarking Open-Source Model and Agent Outputs".
Elias: Here's the gist of the paper. Text watermarking for large language models (LLMs) is important for model owners to verify the origin and protect the intellectual property of AI-generated text While watermarking methods for closed-source LLMs’ text generation are relatively mature, watermarking open-source LLMs’ text generation remains challenging The primary challenge lies in embedding watermarks directly into model weights without compromising detection accuracy The core problem addressed Text watermarking for open-source LLMs is challenging because developers have no control over how decoding occurs, making traditional closed-source methods ineffective The main difficulty lies in embedding watermarks directly into model weights without compromising detection accuracy Existing solutions face two critical challenges: (i) Reduced detectability due to inconsistency between the watermark patterns learned by the model and the predefined patterns used during detection (ii) Vulnerability to modifications by downstream users, such as fine-tuning or model merging, which may weaken or completely remove the embedded watermark How the training signal is built The proposed PRO introduces a trainable watermark policy model that is jointly optimized with the LLM during training This co-optimization helps generate watermark patterns that are easier for the model to learn, significantly reducing inconsistencies between generated patterns and predefined detection criteria The watermark policy consists of a pre-trained Embedding Model E (e.g., BERT) and a trainable Watermark Mapping Model M (e.g., an MLP) The training objective is to minimize the mean KL divergence between teacher and student next token distributions on all prefixes in D Where the gains come from The PRO framework improves detectability and robustness while maintaining generation quality This is achieved through two main components: 1.
Nadia: Co-Adaptive Watermark Policy (CAWP): This co-optimizes a trainable watermark policy model with the student LLM, allowing the watermark patterns to adapt to the LLM’s learning dynamics and become more learnable Detection leverages the co-optimized patterns rather than the original predefined ones, ensuring detection aligns with the watermark signals internalized by the LLM 2.
Elias: Forgotten Perturbation-aware Learning (FPL): This component explicitly reduces model sensitivity to forgotten perturbation, as illustrated in Figure 3 (bottom) The training objective of the LLM is defined as arg min πθ Lsim(πθ) + β Lanti(πθ) − Lanti(πθ − α∇Lanti(πθ))∥∇Lanti(πθ) Robustness against model modifications The FPL component is designed to enhance robustness against user modifications, such as fine-tuning or model merging The training objective of the LLM is defined as arg min πθ Lsim(πθ) + β Lanti(πθ) − Lanti(πθ − α∇Lanti(πθ))∥∇Lanti(πθ) This formulation encourages the model to not only learn a strong watermark but also be robust against potential forgetting induced by user-side modifications Results reveal that PRO consistently preserves high detectability (AUC ≥ 0.80) under aggressive model modifications, including high-ratio merging and long-step fine-tuning Overall performance improvements PRO achieves high detectability, low quality degradation and large n-gram lengths (i.e., n ≥ 5), as shown in Figure 2 (left) PRO bridges the gap between existing open-source watermarks by discovering learning-friendly watermark patterns that LLMs can internalize under low-distortion conditions It maintains high detectability, even at low error tolerance levels, making it more suitable for security-critical applications PRO consistently outperforms prior methods in terms of watermark detectability (i.e., AUC) and generation quality (i.e.
Paper summary: Nadia: So, this paper is about something called OpenMarkX, which is an extensible platform for watermarking open-source model and agent outputs. The core idea here is that you can finally verify if AI-generated text came from an open-source model because the previous methods just don't work for these models.
Elias: Exactly. The thesis is that while we have good watermarking techniques for closed-source LLMs, they don't translate well to open-source models because developers can’t control how the decoding happens. They embed things during decoding in a closed system, but in open source, there is no control over that process at all.
Priya: So what does this mean for the actual security and privacy side? If we can't reliably check the origin of text from these models, it makes verifying provenance incredibly difficult when dealing with AI-generated content.
Nadia: Right. And what OpenMarkX claims is that the main problem isn't just embedding a watermark into the weights, but finding a way to do that without messing up how well we can actually detect it later on.
Elias: The paper points out two big hurdles with existing methods: first, there’s this Generation-Detection Inconsistency where the model learns a version of the pattern different from what we use for detection. Second, there’s vulnerability to things like fine-tuning or model merging which can actually remove the watermark entirely.
Priya: That inconsistency sounds worrying because if you train a model on something, it might learn to ignore that specific watermark pattern, making it useless later on for verification.
Nadia: Well, OpenMarkX tackles this by proposing something called PRO, which stands for Precise and Robust Open-source LLM Watermark. They introduce a trainable watermark policy model that gets optimized together with the LLM during training.
Elias: The idea is that this co-optimization helps create watermark patterns that the model actually learns better, which should cut down on that inconsistency between what's generated and what we use to detect it.
Priya: So instead of just sticking to a fixed pattern for detection, the system learns its own pattern based on how the LLM is actually learning during training?
Nadia: Precisely. They call this Co-Adaptive Watermark Policy, or CAWP. This component co-optimizes with the student LLM to make those watermark patterns adapt to the model's specific learning dynamics.
Elias: And they also introduce Forgotten Perturbation-aware Learning, or FPL, which is designed to handle those modifications we talked about—like when someone fine-tunes a model or merges it with another one.
Priya: How does FPL actually help with robustness against those intentional changes? It sounds like it's actively fighting against the watermark being erased.
Elias: The training objective for the LLM involves minimizing something called Lsim, which is related to similarity, and also an anti-Lanti term that tries to keep the model stable even when we apply a perturbation, which is basically what fine-tuning or merging does.
Paper summary: Nadia: The paper shows that this approach improves robustness significantly; they report achieving an AUC of zero point eight seven under aggressive model modifications, compared to zero point seven nine with existing methods when merging watermarked and unwatermarked models at a ratio of zero point five <ref:2510.23891#pg3,achieving an AUC of 0.87>.
Priya: An AUC of zero point eight seven is solid, but I have to ask, what about the actual quality of the output <ref:2510.23891#pg3,An AUC of 0.87>? Does this improvement in detection come at a cost to how good the text actually is?
Elias: They claim they maintain high generation quality while improving detectability and robustness. They also show improvements in n-gram lengths, specifically for n greater than or equal to five.
Nadia: So, if you're looking at the big picture right now, this work bridges a gap that existed between closed-source and open-source watermarking. It shows that we can discover patterns that LLMs can internalize even under low distortion conditions without sacrificing detection accuracy.
Elias: It’s about discovering these learning-friendly watermark patterns so the LLMs actually take them in, rather than just adopting whatever pattern they see during a fixed process. That is what PRO aims to achieve with OpenMarkX.
Priya: So, for someone who just listens to the show and isn't deep into the math, this paper means that we are getting closer to a reliable way of knowing if AI text has a secret marker from its source model.
Nadia: That’s right. It moves the goal from just trying to embed something into the weights blindly to actively designing a system where the watermark is learned and protected against tampering.
Elias: The authors acknowledge that this platform is extensible, meaning it’s not just one fixed technique, but a framework that can handle different kinds of model and agent outputs.
Priya: What do you think the implications are if this becomes more widely available? Could this become a standard part of how we check AI content?
Nadia: It suggests that for open-source models, verification isn't just about finding one specific trick; it’s about building a flexible system that anticipates how the model will evolve and how users might try to break the watermark.
Elias: The authors stress that this method is better suited for security-critical applications because it maintains high detectability even at low error tolerance levels, which is important when you need to be sure about the text.
Priya: I think the main limitation they point out is that it’s still focused on embedding into model weights, and as long as the LLM decoding process is completely opaque, there's still a dependency there.
Nadia: Exactly. They show that PRO outperforms prior methods in terms of detection AUC and generation quality, but they are focused on the internal mechanism of watermarking rather than solving every problem with AI output verification.
Elias: So OpenMarkX is less about a single perfect solution and more about providing an extensible platform for researchers to develop better ways to embed and detect these signals in open-source models.
Conclusion: Nadia: So we're wrapping up on OpenMarkX, which is this new platform for open-source model watermarking and agent outputs.
Elias: Yeah, looking at the title, "OpenMarkX: An Extensible Platform," it really suggests they’re building something adaptable rather than just a single trick.
Priya: It implies that this isn't some one-off technique; it’s meant to be a system you can plug different types of models into.
Nadia: Exactly, and the authors are focused on creating this framework so it handles the messy reality of open-source development.
Elias: They’re tackling that challenge where developers don't have control over how the model decodes, which is a big hurdle for watermarking here.
Priya: So what’s the simple version of what they accomplished? They managed to make the watermark learnable by optimizing it alongside the LLM during training.
Nadia: Right, so instead of just forcing a pattern in, they let the model help figure out patterns that are actually easier for it to internalize.
Elias: And they also added this part called Forgotten Perturbation-aware Learning, which is designed to keep that watermark from getting wiped out by things like fine-tuning or merging.
Priya: That sounds like they’re trying to build a defense against the inevitable tampering that happens when users play around with open models.
Nadia: It seems the authors are pointing toward a future where verification isn't just about finding one specific secret pattern, but building an entire adaptable system.
Elias: It really shifts the focus from just embedding something into weights to designing a whole process that anticipates how the model might change and how people might try to break it.
Priya: So for me, this means that if we want to trust AI text provenance, we need these kinds of flexible systems instead of just hoping a single watermark works perfectly.
Nadia: It’s about making the verification process robust enough for real-world use in open environments without completely destroying the quality of what the AI is actually generating.
Elias: We’ve seen how they handle detection accuracy and generation quality, but we need to keep an eye on those long n-gram lengths they are achieving.
Priya: That leads us into a different area—what these results actually mean for checking the data itself.
Jiaqi Xue, Yifei Zhao, Mansour Al Ghanim, Shangqian Gao, Ruimin Sun, Qian Lou, Mengxin Zheng
University of Central Florida · Florida State University · Florida International University
cs.CR, cs.AI, cs.LG
Submitted: 2025-10-27
Updated: 2026-11-07
Comments: Accepted by NeurIPS 2026
Project page: https://mobiusml.github.io/hqq_blog
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 85/100
The gist: Text watermarking for large language models (LLMs) is important for model owners to verify the origin and protect the intellectual property of AI-generated text While watermarking methods for…
Key concepts
- Trainable Watermark Policy Model
- This is a model component that learns how to generate watermarks alongside the main LLM during its training process. It consists of an embedding model and a trainable mapping model. By optimizing this policy jointly with the LLM, it ensures the generated watermark patterns are easier for the LLM to learn and integrate effectively into its behavior.
- Co-Adaptive Watermark Policy (CAWP)
- CAWP is a technique where the watermark policy model is co-optimized with the student LLM. This allows the watermark patterns to adapt dynamically to how the LLM learns. Instead of using fixed, predefined patterns, CAWP generates watermarks that align with the specific learning dynamics of different LLMs, making them more robust and detectable.
- Forgotten Perturbation-aware Learning (FPL)
- FPL is a training objective designed to make the LLM robust against changes made after training. It explicitly trains the model to maintain its watermark signal even when subjected to 'forgotten perturbations,' such as fine-tuning or merging. This enhances the watermark's resilience against user modifications, ensuring detection remains accurate.
- Watermark Detectability (AUC)
- Watermark detectability is a measure of how easily an external system can correctly identify whether a piece of text was generated by a watermarked model. It is typically measured using the Area Under the Curve (AUC) metric. High AUC scores indicate that the watermark is robust and detectable even when the model has been modified, which is crucial for security applications.
Terminology
Summary
Text watermarking for large language models (LLMs) is important for model owners to verify the origin and protect the intellectual property of AI-generated text While watermarking methods for closed-source LLMs’ text generation are relatively mature, watermarking open-source LLMs’ text generation remains challenging The primary challenge lies in embedding watermarks directly into model weights without compromising detection accuracy
The core problem addressed
Text watermarking for open-source LLMs is challenging because developers have no control over how decoding occurs, making traditional closed-source methods ineffective The main difficulty lies in embedding watermarks directly into model weights without compromising detection accuracy Existing solutions face two critical challenges: (i) Reduced detectability due to inconsistency between the watermark patterns learned by the model and the predefined patterns used during detection (ii) Vulnerability to modifications by downstream users, such as fine-tuning or model merging, which may weaken or completely remove the embedded watermark
How the training signal is built
The proposed PRO introduces a trainable watermark policy model that is jointly optimized with the LLM during training This co-optimization helps generate watermark patterns that are easier for the model to learn, significantly reducing inconsistencies between generated patterns and predefined detection criteria The watermark policy consists of a pre-trained Embedding Model E (e.g., BERT) and a trainable Watermark Mapping Model M (e. p. 4 The watermark policy consists of a pre-trained Embedding
Model E (e. , BERT) and a trainable Watermark Mapping
Model M (e.Where the paper says it
Where the gains come from
The PRO framework improves detectability and robustness while maintaining generation quality This is achieved through two main components:
-
Co-Adaptive Watermark Policy (CAWP): This co-optimizes a trainable watermark policy model with the student LLM, allowing the watermark patterns to adapt to the LLM’s learning dynamics and become more learnable Detection leverages the co-optimized patterns rather than the original predefined ones, ensuring detection aligns with the watermark signals internalized by the LLM
-
Forgotten Perturbation-aware Learning (FPL): This component explicitly reduces model sensitivity to forgotten perturbation, as illustrated in Figure 3 (bottom) The training objective of the LLM is defined as arg min πθ Lsim(πθ) + β Lanti(πθ) − Lanti(πθ − α∇Lanti(πθ))∥∇Lanti(πθ)
Robustness against model modifications
The FPL component is designed to enhance robustness against user modifications, such as fine-tuning or model merging The training objective of the LLM is defined as arg min πθ Lsim(πθ) + β Lanti(πθ) − Lanti(πθ − α∇Lanti(πθ))∥∇Lanti(πθ) This formulation encourages the model to not only learn a strong watermark but also be robust against potential forgetting induced by user-side modifications Results reveal that PRO consistently preserves high detectability (AUC ≥ 0.80) under aggressive model modifications, including high-ratio merging and long-step fine-tuning
Overall performance improvements
PRO achieves high detectability, low quality degradation and large n-gram lengths (i.e., n ≥ 5), as shown in Figure 2 (left) PRO bridges the gap between existing open-source watermarks by discovering learning-friendly watermark patterns that LLMs can internalize under low-distortion conditions It maintains high detectability, even at low error tolerance levels, making it more suitable for security-critical applications PRO consistently outperforms prior methods in terms of watermark detectability (i. p. 7 As shown
in Figure 6, existing open-source watermarks exhibit quality degradation, while PRO shows the
lowest effect on the quality of text. PRO bridges this gap: by co-optimizing a dynamic watermark
policy with the watermarked LLM, it discovers learning-friendly watermark patterns that LLMs can
internalize under low-distortion conditions. 84 AUC, while achieving 0. 99 AUC with δ = 2 results in significant degradation of generation
quality, i.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Watermark Policy Optimization: The system can improve detection accuracy by introducing a trainable watermark policy model M that is jointly optimized with the LLM during training, ensuring
the watermark patterns are easier for the model to learn
and mitigatingGeneration-Detection Inconsistency.
-
Robustness Against Modifications: The system gains resilience against downstream user changes through Forgotten Perturbation-aware Learning (FPL), which trains the model to
withstand forgotten perturbation by minimizing its disruptive effects while maintaining watermark detectability.
-
Enhanced Detectability and Quality Trade-off: The proposed method achieves high detection accuracy without degrading generation quality, as PRO
bridges this gap: by co-optimizing a dynamic watermark policy with the watermarked LLM, it discovers learning-friendly watermark patterns that LLMs can internalize under low-distortion conditions.
-
Model Attribution Precision: The system enables model-level attribution by using the watermarking policy model trained with one LLM to detect outputs from others, achieving
model-level attribution
and maintaining low false positive rates when distinguishing between specific models.
Abstract
We introduce OpenMarkX, an extensible platform for watermarking open-source model and agent outputs under standard inference. It takes models, compatible methods, private keys, and evaluation settings, and produces watermarked weights, detection results, and quality and robustness reports. Its 16 text-watermark configurations are complemented by agent, vision-language, image, and physics-field workflows. Our proposed method, OpenMark, addresses learnability and robustness through joint model--policy learning and perturbation-aware training. On several open-source LLMs, OpenMark achieves 0.92 TPR at 1% FPR while reducing perplexity by 20.5% on average relative to the strongest prior method, and remains more detectable after such modifications, e.g., retaining 0.57 TPR at 5% FPR after 0.5-ratio model merging, compared with at most 0.34 for prior methods.
Sources
- OpenCodeInstruct: A Large-scale Instruction Tuning Dataset for Code LLMs
- GaussMark: A Practical Approach for Structural Watermarking of Language Models
- On the Opportunities and Risks of Foundation Models
- Composition-contrastive Learning for Sentence Embeddings
- Provably Robust Watermarks for Open-Source Language Models
- Beyond Size: How Gradients Shape Pruning Decisions in Large Language Models
- Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus
- GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers
- Evaluating the Robustness and Accuracy of Text Watermarking Under Real-World Cross-Lingual Manipulations
- Towards Watermarking of Open-Source LLMs
- On the Learnability of Watermarks for Language Models
- DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence
- Factuality Beyond Coherence: Evaluating LLM Watermarking Methods for Medical Texts
- Unbiased Watermark for Large Language Models
- On the Reliability of Watermarks for Large Language Models
- Robust Distortion-free Watermarks for Language Models
- DeepSeek-V3 Technical Report
- A Robust Semantics-based Watermark for Large Language Model against Paraphrasing
- A Simple and Effective Pruning Approach for Large Language Models
- Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data