PaperScope
LIVE · 2026-09-03 05:40 UTC

SemTrace: Source-Grounded Semantic Signatures for Tracing LLM Exposure to Protected Documents

Junyan Zhang, Yudong Zeng, Yongwei Huang, Zuhao Ouyang, Hong Chen, Xuming Hu

Latestcs.CLcs.LGcs.AIcs.CV
arXiv ID
2608.29575 v1
Category
Submitted
2026-08-30

Abstract

Large language models are increasingly used to read documents and produce downstream text, creating a provenance problem when the document owner cannot control or inspect the model that performs the generation. We introduce SemTrace, a source-grounded semantic watermark for detecting whether a generated review was influenced by a known protected manuscript copy. Rather than biasing token probabilities or imposing surface-form patterns, SemTrace constructs a document-specific binary signature from factual propositions that are directly supported by the manuscript itself. A protected PDF invisibly carries a content contract that selects one fact from each binary pair and asks an instruction-following reviewer to express those facts in fixed review slots without changing its independent evaluation. A frozen natural language inference model then decodes the resulting semantic evidence with explicit erasures and scores the recovered bits against the codeword assigned to that copy. This design targets model-agnostic, assigned-copy exposure detection while keeping the watermark semantically tied to the source document.

arXiv abs page · PDF · same-day batch