AgentMark: Behavioral Watermarking for LLM Agents

Researchers propose AgentMark, a framework to embed multi-bit identifiers into the planning behaviors of LLM-based agents, such as tool and subgoal choices, for IP protection and regulatory provenance. Unlike content watermarking, which attributes outputs, AgentMark targets the high-level decision-making layer. It addresses challenges like utility degradation from distributional deviations and black-box agent operation by eliciting an explicit behavior distribution and applying distribution-preserving conditional sampling. The paper is available on arXiv.