Agnibh Dasgupta, Abdullah Tanvir, Xin Zhong · arXiv (Cornell University) · 2024
Watermarking the outputs of large language models (LLMs) is critical for provenance tracing, content regulation, and model accountability. Existing approaches often rely on access to model internals or are constrained by static rules and token-level perturbations. Moreover, the idea of steering generative behavior via prompt-based instruction control remains largely underexplored. We introduce a prompt-guided watermarking framework that operates entirely at the input level and requires no access to model parameters or decoding logits. The framework comprises three cooperating components: a Prompting LM that synthesizes watermarking instructions from user prompts, a Marking LM that generates watermarked outputs conditioned on these instructions, and a Detecting LM trained to classify whether a response carries an embedded watermark. This modular design enables dynamic watermarking that adapts to individual prompts while remaining compatible with diverse LLM architectures, including both proprietary and open-weight models. We evaluate the framework over 25 combinations of Prompting and Marking LMs, such as GPT-4o, Mistral, LLaMA3, and DeepSeek. Experimental results show that watermark signals generalize across architectures and remain robust under fine-tuning, model distillation, and prompt-based adversarial attacks, demonstrating the effectiveness and robustness of the proposed approach.
Cited by 0Open access
Agnibh Dasgupta, Abdullah All Tanvir, Xin Zhong · IEEE Transactions on Artificial Intelligence · 2025
Watermarking the outputs of large language models (LLMs) is critical for provenance tracing, content regulation, and model accountability. Existing approaches often rely on access to model internals or are constrained by static rules and token-level perturbations. Moreover, the idea of steering generative behavior via prompt-based instruction control remains largely underexplored. We introduce a prompt-guided watermarking framework that operates entirely at the input level and requires no access to model parameters or decoding logits. The framework comprises three cooperating components: a Prompting LM that synthesizes watermarking instructions from user prompts, a Marking LM that generates watermarked outputs conditioned on these instructions, and a Detecting LM trained to classify whether a response carries an embedded watermark. This modular design enables dynamic watermarking that adapts to individual prompts while remaining compatible with diverse LLM architectures, including both proprietary and open-weight models. We evaluate the framework over 25 combinations of Prompting and Marking LMs, such as GPT-4o, Mistral, LLaMA3, and DeepSeek. Experimental results show that watermark signals generalize across architectures and remain robust under fine-tuning, model distillation, and prompt-based adversarial attacks, demonstrating the effectiveness and robustness of the proposed approach.
Cited by 0
Ethan Perez, Saffron Huang, Francis Song et al. · arXiv (Cornell University) · 2022
Language Models (LMs) often cannot be deployed because of their potential to harm users in hard-to-predict ways. Prior work identifies harmful behaviors before deployment by using human annotators to hand-write test cases. However, human annotation is expensive, limiting the number and diversity of test cases. In this work, we automatically find cases where a target LM behaves in a harmful way, by generating test cases ("red teaming") using another LM. We evaluate the target LM's replies to generated test questions using a classifier trained to detect offensive content, uncovering tens of thousands of offensive replies in a 280B parameter LM chatbot. We explore several methods, from zero-shot generation to reinforcement learning, for generating test cases with varying levels of diversity and difficulty. Furthermore, we use prompt engineering to control LM-generated test cases to uncover a variety of other harms, automatically finding groups of people that the chatbot discusses in offensive ways, personal and hospital phone numbers generated as the chatbot's own contact info, leakage of private training data in generated text, and harms that occur over the course of a conversation. Overall, LM-based red teaming is one promising tool (among many needed) for finding and fixing diverse, undesirable LM behaviors before impacting users.
Cited by 17Open access
Joss Moorkens, Andy Way, Séamus Lankford · 2024
Cited by 0
Lucas Torroba Hennigen, Yoon Kim · 2023
Masked language models (MLM) do not explicitly define a distribution over language, i.e., they are not language models per se. However, recent work has implicitly treated them as such for the purposes of generation and scoring. This paper studies methods for deriving explicit joint distributions from MLMs, focusing on distributions over two tokens, which makes it possible to calculate exact distributional properties. We find that an approach based on identifying joints whose conditionals are closest to those of the MLM works well and outperforms existing Markov random field-based approaches. We further find that this derived model’s conditionals can even occasionally outperform the original MLM’s conditionals.
Cited by 3Open access
Ethan Perez, Saffron Huang, Francis Song et al. · 2022
Ethan Perez, Saffron Huang, Francis Song, Trevor Cai, Roman Ring, John Aslanides, Amelia Glaese, Nat McAleese, Geoffrey Irving. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.
Cited by 227Open access
Picone, Rico · Figshare · 2023
Block diagram representation of a trained language model.
Cited by 0Open access
Philipp Koehn · Cambridge University Press eBooks · 2009
One essential component of any statistical machine translation system is the language model, which measures how likely it is that a sequence of words would be uttered by an English speaker. It is easy to see the benefits of such a model. Obviously, we want a machine translation system not only to produce output words that are true to the original in meaning, but also to string them together in fluent English sentences.
Cited by 8
Physics Subject Headings (PhySH) · 2018
Cited by 0
SpringerReference · 2011
Cited by 0
Hermann Ney · RWTH Publications (RWTH Aachen) · 1997
Cited by 0
Alexander Koplenig (2552053), Sascha Wolfer (3830233), Jan Oliver Rüdiger (20660533) et al. · Figshare · 2025
The table lists each investigated LM along with its implementation techniques, source, and time required to train it on a document of median length. Training times are provided in seconds and categorised based on the hardware used: Central Processing Unit (CPU) and High-Performance Computing (HPC) cluster with Graphics Processing Unit (GPU) support. The first five LMs were run exclusively on a CPU, while the remaining two LMs were run on a GPU. For comparison, we also include the computation time required if these two models are run on a CPU. Further implementation details are given in <a href="https://journals.plos.org/complexsystems//article/info:doi/10.1371/journal.pcsy.0000032#pcsy.0000032.s001" target="_blank">S1 Text</a>.
Cited by 0Open access
Ney, Hermann · RWTH Publications (RWTH Aachen) · 1997
Cited by 0
Djoerd Hiemstra · Encyclopedia of Database Systems · 2018
Cited by 1
Picone, Rico · Figshare · 2023
Block diagram representation of a trained language model.
Cited by 0Open access
Djoerd Hiemstra · Encyclopedia of Database Systems · 2009
Cited by 12Open access
Ángel Ramos · Zenodo (CERN European Organization for Nuclear Research) · 2016
English and Spanish language models for the NLP Assistant
Cited by 0Open access
Philip Resnik · ArXiv.org · 2026
Futrell and Mahowald claim LMs "serve as model systems", but an assessment at each of Marr's three levels suggests the claim is clearly not true at the implementation level, poorly motivated at the algorithmic-representational level, and problematic at the computational theory level. LMs are good candidates as tools; calling them cognitive models overstates the case and unnecessarily feeds LLM hype.
Cited by 0Open access
Philip Resnik · arXiv (Cornell University) · 2026
Futrell and Mahowald claim LMs "serve as model systems", but an assessment at each of Marr's three levels suggests the claim is clearly not true at the implementation level, poorly motivated at the algorithmic-representational level, and problematic at the computational theory level. LMs are good candidates as tools; calling them cognitive models overstates the case and unnecessarily feeds LLM hype.
Cited by 0Open access
Yi Gu, Zhiting Hu, Tianmin Shu et al. · 2023
Cited by 0