Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Implicit Unwatermarking in Pseudo-Reference Reranking for LLM Watermarking

AbstractText watermarking for large language models (LLMs) requires balancing output quality and detection reliability. We study a practically relevant reranking setting in which multiple watermarked outputs are compared against unwatermarked pseudo-references. We show that quality-only reranking can weaken watermark signals by favoring candidates closer to unwatermarked references, effectively acting as implicit unwatermarking. We propose constrained Minimum Bayes Risk (MBR) selection, which treats detection as a constraint rather than a weighted objective: it filters out weakly watermarked candidates and then selects the highest-quality remaining output. Experiments across watermarking methods, datasets, and models show that constrained selection substantially improves detectability over quality-only reranking while remaining competitive in task quality. It also achieves a more favorable matched-detection tradeoff than normalized weighted MBR.

More from EMNLP 2026

All EMNLP 2026 papers →