Research map / Large language models / Reasoning and chain of thought
Distilling the Thought, Watermarking the Answer: A Principle Semantic Guided Watermark for Reasoning Large Language Models
Shuliang Liu, Xingyu Li, Hongyi Liu, Dong Fang, Duan Bingchen, Zheng Qi, Lingfeng Su, Xuming Hu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a principle-semantic-guided watermark for reasoning LLMs that distills the thought process and watermarks only the final answer, aiming to preserve reasoning coherence at low cost.
Problem
Watermarking reasoning LLMs is hard: token-level pseudo-random biases can corrupt reasoning flow, while semantic-aware watermarks add latency or need auxiliary models.
Approach
Distills the reasoning thought and embeds the watermark in the final answer rather than the reasoning trace, using principle-based semantic guidance; the abstract is truncated so exact mechanism is not fully stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select critical tokens with Criticality Score instead of random tokens (ablation) | ReasonMark · C4 (Qwen3-32B) | perplexity: 12.8801 -> 10.3080 | Improved |
| add Global Causal Contribution to criticality score (ablation) | ReasonMark · C4 (Qwen3-32B) | perplexity: 11.1510 -> 10.3080 | Improved |
| add Competitive Persistence Scoring (ablation) | ReasonMark · C4 (Qwen3-32B) | AUC: 98.69 -> 99.31 | Improved |
| use ReasonMark instead of KGW | Qwen3-32B · C4 | perplexity: 12.15 -> 10.31 | Improved |
| use ReasonMark instead of KGW | Qwen3-32B · WMT16 DE-EN | BLEU: 7.351 -> 9.916 | Improved |
| use ReasonMark instead of KGW on AIME | Qwen3-32B · AIME | math accuracy: 69.23 -> 69.86 | Improved |
| watermarking with ReasonMark vs no watermark | Qwen3-32B · GSM8K | math accuracy: 94.01 -> 93.96 | No clear effect |
| watermarking with ReasonMark vs no watermark | Qwen3-32B · C4 detection | AUC: 98.78 (KGW) -> 99.31 | Improved |
| raise PSV update rate beta0 and top-k beyond optimum (ablation) | ReasonMark · C4 perplexity | perplexity: 11.1 at 1e-3, about 10.3 at 0.1 | Mixed |
| ReasonMark under translation attack | ReasonMark · C4 attacked (Qwen3-32B) | AUC: 99.31 -> 82.58 | Hurt |
| ReasonMark vs other methods on added latency | ReasonMark · C4 latency | time per token: 0.06613 vs 0.06109 (no watermark) | Mixed |
Closest papers by meaning
- Echoes within the Reasoning: Stealthy and Effective Watermarking via Chain of Thought (ICML 2026 · similarity 0.92)
- A Semantic Invariant Robust Watermark for Large Language Models (ICLR 2024 · similarity 0.88)
- Token-Specific Watermarking with Enhanced Detectability and Semantic Coherence for Large Language Models (ICML 2024 · similarity 0.87)
- Unbiased Watermark for Large Language Models (ICLR 2024 · similarity 0.87)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.87)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
