A Linguistics-Aware LLM Watermarking via Syntactic Predictability

2025-10-10 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

STELA is a new watermarking framework for large language models (LLMs) designed to balance text quality with detection robustness, addressing the challenge of public verifiability. Unlike previous methods that rely on model-specific output distributions and require access to logits for detection, STELA aligns watermark strength with the linguistic degrees of freedom in text. It dynamically adjusts the watermark signal using part-of-speech (POS) n-gram-modeled linguistic indeterminacy, reducing it in grammatically constrained contexts to maintain text quality and increasing it in flexible contexts to improve detectability. This approach enables publicly verifiable detection without needing access to the underlying model's logits. Extensive experiments across English, Chinese, and Korean demonstrate that STELA outperforms prior watermarking techniques in detection robustness.

Key takeaway

For research scientists developing LLM governance tools, STELA offers a robust, publicly verifiable watermarking solution. You should consider integrating linguistics-aware methods like STELA to overcome the limitations of logit-dependent detection, enhancing the trustworthiness and transparency of AI-generated content. This approach improves detection robustness across diverse languages while preserving text quality.

Key insights

STELA offers a linguistics-aware LLM watermarking method for publicly verifiable detection without model logits.

Principles

• Balance text quality and detection robustness.
• Align watermark strength with linguistic flexibility.

Method

STELA modulates watermark signal strength using part-of-speech (POS) n-gram-modeled linguistic indeterminacy, weakening it in grammatically constrained contexts and strengthening it in flexible contexts.

In practice

• Apply STELA for publicly verifiable LLM content attribution.
• Use POS n-grams to model linguistic indeterminacy.

Topics

• LLM Watermarking
• Syntactic Predictability
• Linguistic Indeterminacy
• Part-of-Speech n-grams
• Public Verifiability

Code references

• Shinwoo-Park/stela_watermark

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Security Engineer

Related on AIssential

• Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection — A mechanism-oriented taxonomy of indirect linguistic encoding significantly enhances LLM-based coded language detection.
• Equity with Efficiency: An Empirical Study of Tokenizers for Multilingual Large Language Models — Cross-lingual fairness and tokenization efficiency are not fundamentally at odds in multilingual LLMs.
• A Structured Benchmark for Text-Guided Anomaly Detection: When Language Stops Conditioning the Decision — Current multimodal anomaly detection systems exhibit superficial text guidance, often relying more on visual features than language conditioning.
• LLM Self-Recognition: Steering and Retrieving Activation Signatures — LLMs can reliably self-recognize their outputs, a capability enhanced by sparse internal activation steering for robust attribution.
• Linguistics-Aware Non-Distortionary LLM Watermarking — LUNA provides a multilingual, non-distortionary LLM watermark using linguistic context for robust, quality-preserving detection.
• Adaptive Conformal Prediction for Improving Factuality of Generations by Large Language Models — Adaptive conformal prediction improves LLM factuality by calibrating uncertainty thresholds based on prompt characteristics, enhancing conditional coverage.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.