SeqWM: Sequential Watermarking for LLM Agent Behavior
Researchers propose SeqWM, a framework for embedding watermarks into the sequential decision-making of LLM-based agents. Unlike text watermarking, which cannot capture action-level choices, SeqWM embeds signals into history-conditioned transition patterns and verifies trajectories position-ag. This addresses the fragility of prior agent watermarking methods that treat each action as independent, making them vulnerable to perturbation or truncation. The approach aims to establish provenance, ownership, and detect unauthorized reuse of agent policies.
Key facts
- SeqWM is a sequential behavioral watermarking framework for LLM agents.
- It embeds signals into history-conditioned transition patterns.
- It verifies trajectories position-ag.
- Prior agent watermarking methods treat each action step as independent.
- Those methods become fragile when trajectories are perturbed, truncated, or observed without reliable alignment.
- Text watermarking cannot capture action-level decisions.
- The goal is to establish provenance, ownership, and detect unauthorized reuse.
- The paper is from arXiv:2605.11036.
Entities
Institutions
- arXiv