SemTrace: Source-Grounded Semantic Signatures for Document Exposure Detection and Source-Copy Attribution
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Large language models are increasingly used to transform sensitive or controlled documents into downstream text, creating a provenance problem of detecting exposure to a protected copy and attributing the text to its source copy. Existing generation-time watermarks often rely on decoder access, surface-form patterns, or source-irrelevant artifacts. Therefore, we propose SemTrace, a source-grounded semantic signature that encodes provenance through copy-specific choices among alternative source-supported facts. SemTrace verifies and pairs atomic propositions, assigns them through recipient-specific codewords, and uses erasure-aware NLI to recover the semantic pattern for detection and attribution. Experiments under controlled and relaxed generation show strong detection and resilience to paraphrasing, while controlled-generation experiments further demonstrate multi-copy source attribution. Under relaxed generation, SemTrace further outperforms the evaluated ICW baselines in detection and robustness while maintaining competitive text quality. More broadly, SemTrace provides a source-grounded approach to document provenance, demonstrating that semantic content choices can serve as robust and attributable provenance signals in downstream generation.