Watermarking has emerged as a promising solution for tracing and authenticating text generated by large language models (LLMs). A common approach to LLM watermarking is to construct a green/red token list and assign higher or lower generation probabilities to the corresponding tokens, respectively. However, most existing watermarking algorithms rely on heuristic green/red token list designs, as directly optimizing the list design with techniques such as reinforcement learning (RL) comes with several challenges. First, desirable watermarking involves multiple criteria, i.e., detectability, text quality, robustness against removal attacks, and security against spoofing attacks. Directly optimizing for these criteria introduces many partially conflicting reward terms, leading to an unstable convergence process. Second, the vast action space of green/red token list choices is susceptible to reward hacking. In this paper, we propose an end-to-end RL framework for robust and secure LLM watermarking. Our approach adopts an anchoring mechanism for reward terms to ensure stable training and introduces additional regularization terms to prevent reward hacking. Experiments on standard benchmarks with two backbone LLMs show that our method achieves a state-of-the-art trade-off across all criteria, with notable improvements in resistance to spoofing attacks without degrading other criteria.

This paper describes our translation systems for the WMT23 shared task. We participated in the discourse-level literary translation task - constrained track. In our methodology, we conduct a comparative analysis between the conventional Transformer model and the recently introduced MEGA model, which exhibits enhanced capabilities in modeling long-range sequences compared to the traditional Transformers. To explore whether language models can more effectively harness document-level context using paragraph-level data, we took the approach of aggregating sentences into paragraphs from the original literary dataset provided by the organizers. This paragraph-level data was utilized in both the Transformer and MEGA models. To ensure a fair comparison across all systems, we employed a sentence-alignment strategy to reverse our translation results from the paragraph-level back to the sentence-level alignment. Finally, our evaluation process encompassed sentence-level metrics such as BLEU, as well as two document-level metrics: d-BLEU and BlonDe.