With the widespread adoption of Large Language Models (LLMs), concerns about potential misuse have emerged. To this end, watermarking has been adapted to LLM, enabling a simple and effective way to detect and monitor generated text. However, while the existing methods can differentiate between watermarked and unwatermarked text with high accuracy, they often face a trade-off between the quality of the generated text and the effectiveness of the watermarking process. In this work, we present a novel type of LLM watermark, Sparse WatermARK (or SpARK), which aims to mitigate this trade-off by applying watermarks to a small subset of generated tokens distributed across the text. To demonstrate this type of watermark, we introduce two novel variants, SpARK-P and SpARK-R, which achieve sparsity by anchoring watermarked tokens to words that have specific Part-of-Speech (POS) tags and specific hash values w.r.t a pseudorandom hash function, respectively. Our experimental results demonstrate that the proposed watermarking schemes, albeit embarrassingly simple, are incredibly effective, achieving high detectability while generating text that outperforms previous LLM watermarking methods in quality across various tasks. SpARK further advances the watermarking capability for LLMs while maintaining their generated text quality.

Code large language model introduces copyright risks when models reproduce protected implementations from training data. Existing inference-time copyright-protection methods overlook the power of using the original LLM capability for a runnable code generation. In this paper, we propose AegisCodeGen (ACG), an inference-time copyright-aware controlled generation framework for code LLMs, effectively extending the text copyright protection towards the code generation domain. We argue that code copyright protection should be formulated as a region-aware generation control problem, where the model first identifies where modification is safe and then decides how strongly copyright-aware fusion should be applied. Based on this insight, ACG mainly consists of two stages. Firstly, ACG performs multiple sampling generations to construct an agreement profile over the target program, indicating which regions are more likely to affect core code logic after modification. Secondly, ACG uses this profile to adaptively control fusion strength, applying stronger copyright protection in low-agreement regions and weaker shifts near executable anchors. Through intense experiments, we demonstrate that ACG reduces protected-content reproduction while preserving runability and generation utility, providing a lightweight copyright-aware module that requires neither model retraining nor access to training data.