Learning to Follow In-Context Watermark Instructions via Self-Distillation

• URL: http://arxiv.org/abs/2608.29030v1
• Date: Sat, 29 Aug 2026 03:42:28 GMT
• Title: Learning to Follow In-Context Watermark Instructions via Self-Distillation
• Abstract summary: In-context watermarking (ICW) prepends an instruction to a query asking the model to embed a statistically detectable signal in its response.<n>We introduce $mathsfICWBench$, a benchmark of three verifiable ICW instruction families, each scored on both detectability and answer quality.<n>We propose a self-contained two-stage training method, requiring no distillation from a stronger model, no manual annotation, and no pre-existing ICW IF ability.
• Score: 117.19153619258644
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: In-context watermarking (ICW) prepends an instruction to a query asking the model to embed a statistically detectable signal in its response. It thus equips LLMs with a watermarking interface that third parties can invoke without access to model internals. Its reliability hinges on the LLM following the instruction without degrading answer quality, yet how well current LLMs do so has not been measured. We introduce $\mathsf{ICWBench}$, a benchmark of three verifiable ICW instruction families, each scored on both detectability and answer quality. Evaluating 14 frontier proprietary and open-source LLMs, we find that none of the evaluated LLMs achieves both objectives across all three families. To address this, we propose a self-contained two-stage training method, requiring no distillation from a stronger model, no manual annotation, and no pre-existing ICW IF ability. The first stage, self-distillation with logits perturbation (SDLP), uses the same base LLM as both teacher and student: an instruction-equivalent decoding-time logits perturbation makes the teacher follow the ICW instruction, and the student is trained to match the teacher's output distribution. The second stage applies reinforcement learning with the automatic verifier as the reward. Applied to Qwen3-14B and GPT-OSS-20B, our method raises average TPR@$1\%$FPR across three ICW instructions from $0.100$ to $0.974$ and from $0.337$ to $0.968$, respectively, while maintaining high response quality under both perplexity evaluation and LLM-as-a-Judge.

Related papers

• Human-in-the-Loop LLM Grading for Handwritten Mathematics Assessments [4.291185382337384]
  We present a scalable, end-to-end workflow for LLM-assisted grading of short, pen-and-paper assessments.<n>We deploy the system in two undergraduate mathematics courses using six low-stakes in-class tests.
  arXiv  Detail & Related papers  (2026-03-13T15:32:09Z)
• Teaching LLMs to Abstain via Fine-Grained Semantic Confidence Reward [4.921470220575384]
  Mitigating hallucinations in Large Language Models (LLMs) is critical for their reliable deployment.<n>We propose a novel reinforcement learning framework built on $textbfunderlineFine-grained underlineSemantic underlineConfidence underlineReward (Ours)
  arXiv  Detail & Related papers  (2025-10-28T03:00:35Z)
• e3: Learning to Explore Enables Extrapolation of Test-Time Compute for LLMs [49.01449646799905]
  We show that most existing reasoning models do not extrapolate well.<n>Our recipe e3 produces the best known 1.7B model according to AIME'25 and HMMT'25 scores.<n>Our e3-1.7B model not only attains high pass@1 scores, but also improves pass@k over the base model.
  arXiv  Detail & Related papers  (2025-06-10T17:52:42Z)
• Robustness via Referencing: Defending against Prompt Injection Attacks by Referencing the Executed Instruction [68.6543680065379]
  Large language models (LLMs) are vulnerable to prompt injection attacks.<n>We propose a novel defense method that leverages, rather than suppresses, the instruction-following abilities of LLMs.
  arXiv  Detail & Related papers  (2025-04-29T07:13:53Z)
• Ensemble ToT of LLMs and Its Application to Automatic Grading System for Supporting Self-Learning [0.8490659704051299]
  Ensemble Tree-of-Thought (ToT) is a framework that enhances LLM outputs by integrating multiple models.<n>Our grading system first evaluates the grading tendencies of LLMs, then generates multiple results, and finally integrates them via a simulated debate.
  arXiv  Detail & Related papers  (2025-02-23T01:17:46Z)
• S$^2$R: Teaching LLMs to Self-verify and Self-correct via Reinforcement Learning [51.84977135926156]
  We introduce S$2$R, an efficient framework that enhances LLM reasoning by teaching models to self-verify and self-correct during inference.<n>Our results demonstrate that Qwen2.5-math-7B achieves an accuracy improvement from 51.0% to 81.6%, outperforming models trained on an equivalent amount of long-CoT distilled data.
  arXiv  Detail & Related papers  (2025-02-18T13:40:22Z)
• Understanding the Dark Side of LLMs' Intrinsic Self-Correction [58.12627172032851]
  Intrinsic self-correction was proposed to improve LLMs' responses via feedback prompts solely based on their inherent capability.<n>Recent works show that LLMs' intrinsic self-correction fails without oracle labels as feedback prompts.<n>We identify intrinsic self-correction can cause LLMs to waver both intermedia and final answers and lead to prompt bias on simple factual questions.
  arXiv  Detail & Related papers  (2024-12-19T15:39:31Z)
• TICKing All the Boxes: Generated Checklists Improve LLM Evaluation and Generation [24.954629877691623]
  TICK (Targeted Instruct-evaluation with ChecKlists) is a fully automated, interpretable evaluation protocol. We first show that, given an instruction, LLMs can reliably produce high-quality, tailored evaluation checklists. We then show that STICK can be used to improve generation quality across multiple benchmarks via self-refinement and Best-of-N selection.
  arXiv  Detail & Related papers  (2024-10-04T17:09:08Z)
• Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models [54.14602121129874]
  We introduce AutoIF, the first scalable and reliable method for automatically generating instruction-following training data. AutoIF transforms the validation of instruction-following data quality into code verification.
  arXiv  Detail & Related papers  (2024-06-19T13:29:53Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.