New method ActHook watermarks LLM agent trajectories for copyright protection

Researchers have developed ActHook, a novel watermarking technique designed to protect the copyright of large language model (LLM) agent trajectory datasets. This method embeds hidden 'hook actions' into the data that are activated by a secret key, allowing for the detection of unauthorized use without affecting the agent's performance. Experiments demonstrated ActHook's effectiveness in identifying watermarked trajectories with high accuracy on various agent types, including those used for coding and web searching. AI

IMPACT Introduces a method to protect intellectual property in LLM agent training data, potentially impacting dataset creators and users.

RANK_REASON Academic paper introducing a new method for watermarking LLM agent trajectories. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• ActHook
• arXiv
• LLM
• Qwen-2.5-Coder-7B

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →