重点学习：LLM的token采样机制（top-k, temperature）
推荐资料：研究论文《On the Reliability of Watermarks for Large Language Models》
实践建议：用HuggingFace transformers比较有水印/无水印的输出差异
延伸阅读：Oulipo文学运动与约束写作
06给 Codex 的 Prompt
Analyze the technical implementation differences between standard LLM token sampling and watermarked sampling. Create a comparison table covering: 1) random number generation method, 2) output determinism, 3) quality impact metrics, 4) traceability features. Then write Python code to simulate both sampling approaches using GPT-2, measuring perplexity differences on Shakespeare text generation.