Advances in generative models have made it possible for AI-generated text, code, and images to mirror human-generated content in many applications. W atermark-ing, a technique that aims to embed information in the output of a model to verify its source, is useful for mitigating the misuse of such AI-generated content. However, we show that common design choices in LLM watermarking schemes make the resulting systems surprisingly susceptible to attack--leading to fundamental trade-offs in robustness, utility, and usability. To navigate these trade-offs, we rigorously study a set of simple yet effective attacks on common watermarking systems, and propose guidelines and defenses for LLM watermarking in practice.

Implicitly or explicitly, this data is composed of domains--varying distributions of language.

This intervention significantly improves the performance of LLaMA models on the TruthfulQA benchmark.

China's Renewable Energy Revolution Is a Huge Mess That Might Save the World A global onslaught of cheap Chinese green power is upending everything in its path. No one is ready for its repercussions. There's a particular kind of sci-fi nerd who equates fusion tech with utopia. If we could only harness the engine of the stars, it would uncork near limitless energy and neatly sweep away a whole mess of humanity's problems. But how would that work exactly? What would the transition look like?