STELA Framework Introduces Linguistics-Aware Watermarking for LLMs

A new framework called STELA addresses the persistent challenge of balancing text quality with detection robustness in large language model watermarking. Unlike recent approaches that rely on model-specific signals like token-level entropy, STELA modulates watermark strength based on linguistic degrees of freedom inherent in language. This method uses part-of-speech n-gram modeling to assess linguistic indeterminacy, weakening the signal in grammatically constrained contexts to preserve quality and strengthening it elsewhere. The framework aims to overcome the barrier to public verification posed by methods requiring access to underlying model logits. As LLMs advance rapidly, publicly verifiable watermarking has become critical for fostering a trustworthy AI ecosystem. The research was published on arXiv with identifier 2510.13829v4 under announcement type replace-cross. STELA represents a novel approach to aligning watermark strength with linguistic predictability rather than model output distributions.