acceptodds
Withdrawn from ICLR 2027

The Signal Beneath the Words: Paraphrase-Resistant Watermarking for LLM Distillation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unauthorized distillation poses a growing threat to the intellectual property of large language models. Existing tracing methods commonly embed watermarks through token-level statistical biases that can be inherited by student models. Yet even simple paraphrasing before distillation can substantially weaken this inheritance: paraphrasing readily rewrites lexical choices while preserving the useful content of the response. This exposes a fundamental weakness of token-level watermarks: their carrier lies precisely in the part of language that paraphrasing is designed to change. We take a different perspective: rather than watermarking what words a model chooses, we watermark a faint but persistent affective tendency in how it speaks. Our key observation is that language carries more than explicit content. Responses can exhibit a subtle emotional undertone that accompanies the text without dominating its meaning or compromising response quality, much like a weak aspect of personality. When expressed consistently across teacher responses, such an affective tendency can become part of what a distilled student learns. Based on this observation, we propose an emotion-circuit watermarking framework that implants a weak, quality-preserving affective tendency, e.g., a slight tendency toward disgust, by steering emotion-related internal activations during teacher generation. The watermark remains subtle in individual responses but becomes statistically detectable across outputs. Verification determines whether this tendency has been inherited by comparing the student's target-emotion responses and their contrast with other emotions against matched unwatermarked reference students. Across two student families and five token-watermark baselines, our watermark remains detectable in 100% of students after neutral paraphrasing of the teacher data, including when distillation questions come from a second source.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Intellectual Property Protection, Knowledge Distillation, Model Watermarking, Emotion Circuit Steering, Model Ownership Verification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sKBrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes EmotionMark, which aims to protect the IP of LLM generated content. The idea of EmotionMark is to embed emotional tune into LLM generated content. Such emotional tune could survive LLM paraphrase attack so that any student model which distillate on it could be identified. Further experiments demonstrate the effectiveness of EmotionMark.

Critical strengths
  1. The anti-distillation problem is real and urgent, making the motivation of this paper sound and convincing
  2. The proposed method is simple and clear.
Critical weaknesses
  1. The soundness of EmotionMark is not explained. E.g. why such emotional tune can survive LLM paraphrase?
  2. The verification stage relies on the accuracy of public emotion verifier, whose performance constructs a big concern for the reliability of EmotionMark
  3. The novelty of EmotionMark is limited, it is just a domain transfer of steering LLMs.
Questions for authors
  1. Why EmotionMark could survive LLM paraphrase? It is likely that a simple emotional style paraphrase can erase EmotionMark.
  2. What is the main challenge of applying LLM steering on anti-distillation? If it is simply a domain transfer, the novelty of EmotionMark is a big concern for me.
Reviewer_k2MgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To improve watermark robustness against paraphrasing attacks in unauthorized LLM distillation, this paper proposes EmotionMark, which embeds watermark signals in the subtle emotional tone of teacher-generated responses rather than token-level patterns. Experimental results demonstrate that the proposed watermark remains detectable in distilled student models even after paraphrasing, outperforming existing token-level watermarking methods.

Critical strengths
  • The paper explores an interesting direction for distillation watermarking by using subtle affective tendencies rather than token-level patterns as watermark carriers.

  • The experimental results demonstrate substantial improvements in robustness against paraphrasing compared to existing token-level watermarking baselines.

  • The paper provides extensive ablation studies and verification analyses, including evaluations under various attack settings and false-positive assessments.

Critical weaknesses
  • My biggest concern is the practical applicability of using disgust as the watermark carrier. Such a negative emotional tendency may be undesirable for real-world LLM services. It remains unclear whether more service-compatible emotions can achieve comparable detectability without degrading user experience.

  • The robustness against adaptive, emotion-targeted paraphrasing remains unclear. The neutral paraphrasing attack preserves approximately 48% of the original affective signal (Table 11), while stronger attacks that explicitly suppress disgust or rewrite responses in a cheerful tone before distillation are not evaluated. In practice, a distiller would likely remove such undesirable emotional tendencies to prevent the student model from inheriting a disgust response style.

  • The verification reliability depends on the reference pool and the suspect student's training configuration. With only 20 reference students per family, it is unclear whether the reported detection rates and FPR remain reliable across different reference pools and unseen student configurations.

  • The experiments focus on general instruction-following datasets (Alpaca and WildChat), without evaluating reasoning-intensive distillation. Since extracting reasoning capabilities is an important distillation scenario, as also considered in ADFP, it is unclear whether the watermark remains effective on emotion-neutral tasks such as mathematical reasoning and coding without degrading task performance.

Questions for authors

Please see the weaknesses above. My primary concern is the use of disgust as a watermarking signal, which may discourage service providers from adopting the method and distillers from using such responses. The authors should investigate whether more service-compatible emotional signals can achieve comparable detectability without compromising user experience.

Reviewer_WbSuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors propose EmotionMark, a watermarking scheme uses activation steering to embed subtle affective tendencies into teacher responses. Experiments across two student families show robustness against paraphrasing and several rewriting attacks, with relatively modest perplexity increases.

Critical strengths
  1. The problem is important and well-motivated, as paraphrasing can substantially weaken existing token-level watermarks which have been proven by previous works.

  2. The affective watermarking idea is interesting, and the authors conduct extensive experiments and ablations.

Critical weaknesses
  1. The adaptive attack evaluation is limited to predefined rewriting strategies and classifier-guided best-of-eight rewriting. I would like to see evaluation against stronger learned attacks such as RLCracker, which optimizes paraphrasing to remove watermark signals while preserving semantic content. It remains unclear whether EmotionMark can survive such attacks.

  2. The baseline comparison is not sufficiently convincing. SynthID-Text, TextSeal, and Gumbel-max show little or no inheritance even without attacks. Moreover, EmotionMark uses best-of-five candidate selection, while the baselines do not, making it difficult to isolate the advantage of the proposed watermark carrier.

  3. The quality-preservation claim is not fully supported. The method introduces measurable perplexity increases, and downstream benchmark results are not uniformly conclusive. Furthermore, detecting an affective tendency does not necessarily establish that a student was distilled from the specific teacher, particularly when the verifier is keyless.

Questions for authors
  1. Could the authors evaluate EmotionMark against RLCracker or a similar learned adaptive paraphrasing attack?

  2. Could the authors provide stronger baseline comparisons under matched generation budgets and more convincing evidence for quality preservation and source attribution?

  3. Could the authors evaluate whether EmotionMark generalizes to other target emotions beyond disgust? How sensitive is detection performance to the choice of emotion and steering strength?

  4. Since the verifier relies on public emotion classifiers, how can the authors distinguish genuine watermark inheritance from independently induced affective tendencies or distribution shifts? Could they provide more comprehensive false-positive evaluations under these conditions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.