Self-Distillation Improves LLM In-Context Watermarking Reliability

In-context watermarking (ICW) offers a method for third parties to embed a statistically detectable signal into an LLM's response by simply prepending an instruction to a query, without needing access to the model's internal workings. However, the reliability of this technique hinges on the LLM's ability to consistently follow these instructions while maintaining high answer quality, a capability that current frontier models often lack. To address this, researchers introduced \mathsf{ICWBench}, a benchmark designed to measure both detectability and answer quality across various ICW instruction families. Their evaluation of 14 leading proprietary and open-source LLMs revealed that none could consistently achieve both objectives across all instruction types. To overcome this limitation, a novel two-stage training method was developed, utilizing self-distillation. The first stage, self-distillation with logits perturbation (SDLP), uses the base LLM itself as both teacher and student, where the teacher is guided by an instruction-equivalent decoding-time perturbation. The student then learns to match the teacher's output distribution. The second stage employs reinforcement learning with an automatic verifier providing rewards. This method significantly improved watermarking detectability (from 0.100 to 0.974 for Qwen3-14B and 0.337 to 0.968 for GPT-OSS-20B) while preserving response quality.