Agents
arXiv: Refined Near-Optimal Detection for Gumbel Watermarking — Provable Agent Output Attribution
Lattimore proposes a provably near-optimal detection mechanism for the Gumbel watermarking scheme (Aaronson 2022) used to attribute LLM-generated text. The refined detection is simpler than previous approaches while achieving stronger theoretical guarantees. Directly applicable to multi-agent systems and agentic workflows where attribution of which agent or model produced specific outputs is critical for audit trails, accountability, and debugging agent chains.
Source
↳ Follow the thread