acceptodds
Under review as a conference paper at ICLR 2027

Gradient-Guided Radioactive Watermarking for Detecting Fine-Tuning on Teacher-Generated Data

Abstract

Model distillation transfers capabilities from powerful large language models (LLMs) to smaller downstream models, raising an important provenance question: Has a downstream model been fine-tuned on protected teacher-generated data? Existing methods embed watermark into generated text to detect such data use. However, their watermark statistics are evaluated on training-related or same-prompt contexts. In real-world scenarios, model owners may not know which data is used for distillation. To address this gap, we introduce Gradient-Guided Radioactive Watermarking (GGRW), a generation-time watermarking framework that links token selection to downstream fine-tuning dynamics. GGRW uses a proxy model to identify and sample tokens whose induced fine-tuning updates reinforce a private watermark direction, enabling detection on disjoint held-out contexts used in neither watermark construction nor student fine-tuning. Experimental results show that the watermark is detectable under dilution attacks and distillation, where student architectures are unknown. Furthermore, our method effectively detects the watermark on disjoint held-out contexts, whereas the baseline methods fail.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Radioactive Watermarking, Training Data Provenance, Fine-Tuning Detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gzcrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Gradient-Guided Radioactive Watermarking (GGRW), a method for detecting whether a downstream language model was fine-tuned on protected teacher-generated data. The method first constructs a private, key-dependent watermark response on a probe set and computes its gradient in the parameter space of a proxy model. During teacher generation, candidate tokens are rewarded when the proxy-model update induced by training on that token aligns with the watermark direction. A symmetric finite-difference approximation allows the alignment scores for all tokens to be estimated using two perturbed copies of the proxy model.

After a student has been fine-tuned, GGRW tests whether it assigns increased probability to alignment-weighted, key-dependent carrier tokens on a disjoint held-out set. Experiments on GSM8K cover one teacher, one proxy, and three student configurations. The reported results show detection across Qwen and Llama students, including under clean-data dilution, while KGW, ADFP, and BWM are not detected under the paper’s disjoint held-out protocols.

Critical strengths
  • The paper studies an important and practically relevant provenance problem that is not handled well by ordinary output watermarking.
  • GGRW directly connects token selection to a desired downstream training direction, rather than assuming that an output watermark will automatically survive fine-tuning.
  • The use of disjoint probe, training, and detection examples is a meaningful improvement over evaluations that reuse training prompts.
  • The finite-difference construction turns an otherwise expensive token-by-token gradient computation into two proxy forward passes per decoding step.
  • The dilution results show a fairly clear dose-response relationship between the protected-data fraction and the detection statistic.
  • The ablation supports the claim that the gradient-weighted, key-dependent reward is the component responsible for the observed signal.
  • The appendix is appropriately transparent that ADFP and BWM can succeed under their native or same-prompt evaluation protocols.
Critical weaknesses
  1. The wrong-key test is not sufficiently justified.
    For a wrong key, the paper changes the green partition while keeping the target-key watermark direction and alignment scores fixed. The paper does not establish the exchangeability required for interpreting the resulting value as a calibrated p-value. A single unwatermarked run per architecture is insufficient to demonstrate a 5% false-positive rate.

  2. Cross-tokenizer detection is not formally defined.
    The candidate set is defined using teacher tokens, but the Llama students use a different tokenizer. A Qwen token may correspond to multiple Llama tokens, and candidate strings may overlap by prefix. The paper does not explain how these probabilities are computed, making the cross-tokenizer results difficult to interpret or reproduce.

  3. The experimental scope is narrow.
    All experiments use GSM8K, one teacher, one proxy, roughly 3B-parameter students, 1,024 training examples, and rank-8 LoRA. Although the held-out examples are disjoint, they come from the same narrow distribution. This evidence does not yet support broad claims about unknown downstream architectures or realistic distillation settings.

  4. Run-to-run variability is not measured.
    The experiments appear to use one key, one data split, and one training run per condition. Moreover, the nominally similar Qwen, λwm=1.5\lambda_{\mathrm{wm}}=1.5, 100%-protected settings in Tables B.1 and B.4 report student accuracies of 69.53% and 51.17%, respectively. This discrepancy makes repeated runs and confidence intervals necessary.

  5. The utility claim is not adequately supported.
    The paper states that utility does not vary significantly, but it reports neither repeated trials nor significance tests. Qwen student accuracy ranges from 53.91% to 69.53% across nonzero watermark strengths, which is a substantial variation.

  6. The teacher-gap term confounds watermark strength with sampling temperature.
    After removing the token-independent constant, Equations 13–14 scale the teacher logits by 1+λwmμwm1+\lambda_{\mathrm{wm}}\mu_{\mathrm{wm}}. The gap penalty is therefore equivalent to lowering the effective temperature. As a result, the strength sweep changes both the watermark reward and the sampling entropy.

  7. The baseline comparison is not fully controlled.
    ADFP and BWM are evaluated outside their primary settings, and the protocols use different numbers and types of effective observations. The reported results therefore do not establish superiority at matched false-positive rates, query budgets, and utility levels.

  8. Robustness and computational cost are underexplored.
    Detection fails for all three students when protected data constitute at most 40% of the training set. The paper does not test paraphrasing, continued clean fine-tuning, full-parameter fine-tuning, model merging, quantization, or adaptive watermark removal. It also does not report the throughput and memory overhead of two additional proxy forward passes per generated token.

Questions for authors
  1. How exactly are the candidate sets CiC_i, green partitions, and key-dependent hash constructed? What are the candidate-set size, green fraction, and value of ξ\xi?

  2. What exchangeability argument justifies Equation 18 when a wrong key changes the green partition but not the watermark direction or alignment scores?

  3. What is the empirical false-positive rate across multiple clean models, target keys, training seeds, and data splits?

  4. How is pm(v∣c,Ci)p_m(v\mid c,C_i) computed when vv is a Qwen token but the suspect model uses the Llama tokenizer?

  5. Can the authors report results over multiple independent generation and training runs? What explains the Qwen student accuracy difference between Tables B.1 and B.4?

  6. Do the conclusions remain unchanged when each watermark strength is compared with a control using the same effective sampling temperature?

  7. How does GGRW perform after paraphrasing, continued clean fine-tuning, full-parameter fine-tuning, model merging, quantization, or watermark scrubbing?

  8. What are the generation throughput, memory overhead, and total computational cost relative to KGW and ADFP?

6 comments

Question 1:

How exactly are the candidate sets CiC_i, green partitions, and key-dependent hash constructed? What are the candidate-set size, green fraction, and value of ξ\xi?

Answer 1:

Thank you for identifying these missing implementation details. Section 3.2 defines the candidate sets and key-dependent partitions conceptually. We clarify their implementation and the numerical value of ξ\xi below.

For context cic_i, let zT(v∣ci)z_T(v\mid c_i) denote the teacher's unscaled next-token logit. We construct candidate sets as

Ci={v∈Top64⁡(zT(⋅∣ci)):max⁡uzT(u∣ci)−zT(v∣ci)≤3.0}.C_i = \left\lbrace v\in\operatorname{Top64}\bigl(z_T(\cdot\mid c_i)\bigr): \max_u z_T(u\mid c_i)-z_T(v\mid c_i)\leq 3.0 \right\rbrace.

Thus, the candidate-set size is context-dependent and capped at 64.

To construct the partition, let aia_i be the last Qwen token ID in cic_i. For each candidate vv, represented by its Qwen token ID, we compute

hk(ci,v)=Hash⁡64(s,k,ai,v)264−1,h_k(c_i,v) = \frac{\operatorname{Hash}_{64}(s,k,a_i,v)}{2^{64}-1},

where ss is a fixed salt. The function Hash⁡64\operatorname{Hash}_{64} computes the SHA-256 digest of the serialized tuple (s,k,ai,v)(s,k,a_i,v) and interprets the first eight digest bytes as an unsigned 64-bit integer in big-endian order. We set

Gik={v∈Ci:hk(ci,v)<0.5},Rik=Ci∖Gik.G_i^k = \lbrace v\in C_i:h_k(c_i,v)<0.5\rbrace, \qquad R_i^k=C_i\setminus G_i^k.

The green assignment probability is 0.5; Within the candidate set, the numbers of green and red tokens may differ.

We use ξ=10−8\xi=10^{-8} in Eq. (3). These settings are shared across the Qwen, Llama, and Llama-Instruct experiments.

Question 2:

What exchangeability argument justifies Equation 18 when a wrong key changes the green partition but not the watermark direction or alignment scores?

Answer 2:

Thank you for raising this point. Equation (18) is intended to quantify how extreme the target-key statistic is relative to the statistics obtained using random wrong keys. Computing this rank-based comparison does not require exchangeability. The target key is used to construct the watermark direction and alignment scores, whereas the wrong keys only change the green partition. Holding the direction and alignment scores fixed therefore provides an empirical comparison of sensitivity to the green partition, rather than an exchangeable null reference.

Question 3:

What is the empirical false-positive rate across multiple clean models, target keys, training seeds, and data splits?

Answer 3:

Thank you for this question. We observe 2 false positives among 90 evaluation combinations, corresponding to an empirical false-positive rate of 2.22%, using the decision rule Mkm>0M_k^m>0 and pwmm≤0.05p_{\mathrm{wm}}^m\leq 0.05.

The evaluation covers:

  • Clean models and training seeds: Six Qwen2.5-3B students are trained separately on unwatermarked data using LoRA ranks 8 and 16, with three distinct training seeds per rank.
  • Target keys: Five distinct target keys, with the watermark direction and detection features reconstructed for each target key.
  • Detection-data splits: Three mutually disjoint GSM8K detection sets. The student-training and direction-construction data are held fixed.

Each student is evaluated using all five target keys on all three detection sets, yielding 90 combinations. Each LoRA configuration produces 1 false positive among 45 combinations. For each detection set, we evaluate all six student models using all five target keys, yielding 30 evaluation combinations. Across the three detection sets, the false-positive counts are 2/30, 0/30, and 0/30, respectively. These results assess empirical false-positive behavior across clean models, training seeds, target keys, and detection-data splits within the Qwen2.5-3B/GSM8K setting.

Question 4:

How is pm(v∣c,Ci)p_m(v \mid c,C_i) computed when vv is a Qwen token but the suspect model uses the Llama tokenizer?

Answer 4:

Thank you for pointing out this ambiguity. In the cross-tokenizer setting, we use an explicit mapping between the Qwen and Llama vocabularies based on exact token-string matches. Let ϕ(v)\phi(v) denote the Llama token ID whose vocabulary token string matches that of Qwen token vv, and define the retained candidate set as

C~i={v∈Ci:ϕ(v) exists and is non-special}.\widetilde C_i=\lbrace v\in C_i:\phi(v)\text{ exists and is non-special}\rbrace.

We re-tokenize the same context text with the Llama tokenizer and verify that the Llama and Qwen tokenizations end at the same character position, so that the next-token probabilities are evaluated at the same textual boundary. Writing this Llama-tokenized context as cmc_m, we compute

pm(v∣c,C~i)=exp⁡zm(ϕ(v)∣cm)∑u∈C~iexp⁡zm(ϕ(u)∣cm),v∈C~i,p_m(v\mid c,\widetilde C_i) = \frac{\exp z_m(\phi(v)\mid c_m)} {\sum_{u\in\widetilde C_i}\exp z_m(\phi(u)\mid c_m)}, \qquad v\in\widetilde C_i,

where zmz_m is the suspect model's next-token logit. Thus, Qwen token IDs are never used directly to index Llama logits. Candidates without a shared single-token representation are excluded. We also renormalize the reference probabilities over the same retained set and preserve the original Qwen IDs for watermark green/red assignment. The same procedure is used for Llama-3.2-3B and Llama-3.2-3B-Instruct.

Question 5:

Can the authors report results over multiple independent generation and training runs? What explains the Qwen student accuracy difference between Tables B.1 and B.4?

Answer 5:

Thank you for raising these two questions. We report results from additional independent generation and training runs and examine the Qwen student accuracy discrepancy between Tables B.1 and B.4.

1. Results across independent generation and training runs.

We conduct independent generation and training experiments under the Qwen2.5-3B/GSM8K configuration with watermark strength 1.5. Each paired run uses a distinct generation seed to regenerate teacher responses and a distinct training seed to train each student from the base model. Within each pair, the watermarked and unwatermarked conditions use the same 1024 training questions and share the generation and training seeds. All students are evaluated on the same 512 GSM8K test questions.

The table below presents two completed paired runs, numbered as Runs 1 and 2. We apply the same detection rule to every student: detection is Yes if Mkm>0M_k^m>0 and pwmm≤0.05p_{\mathrm{wm}}^m\leq0.05; otherwise, detection is No. **The detection key, detection contexts, and wrong-key set remain fixed across runs.

RunStudentAccuracyMkmM_k^mpwmmp_{\mathrm{wm}}^mDetected
1Watermarked60.74%1.12380.0020Yes
1Unwatermarked64.65%−0.0952-0.09520.9760No
2Watermarked56.84%1.17860.0020Yes
2Unwatermarked60.94%−0.1274-0.12740.9721No

These two paired runs show variation in student accuracy across independent generation and training runs. The watermarked students have lower accuracy than their corresponding unwatermarked controls by 3.91 and 4.10 percentage points in Runs 1 and 2, respectively. In both runs, the watermark is detected in the watermarked student but not in the unwatermarked control.

2. Investigation of the Qwen accuracy discrepancy between Tables B.1 and B.4.

We check the reported values against our original experimental records and find that the previously reported Qwen2.5-3B result in Table B.4 was obtained from an incomplete training run. We rerun the experiment to completion and update the corresponding results for Qwen2.5-3B in Table B.4 of the revised manuscript. We apologize for this reporting error. The corrected results are shown below.

Student modelλwm\lambda_{\mathrm{wm}}αmix\alpha_{\mathrm{mix}}Student accuracyMkmM_k^mpwmmp_{\mathrm{wm}}^mDetected
Qwen2.5-3B1.51.069.92%1.21970.0020Yes

Question 6:

Do the conclusions remain unchanged when each watermark strength is compared with a control using the same effective sampling temperature?

Answer 6:

Thank you for highlighting this potential confound. Expanding Eqs. (13)–(14), the teacher-gap penalty rescales the teacher logits, giving an effective temperature

Teff=τ1+λwmμwm.T_{\mathrm{eff}} =\frac{\tau}{1+\lambda_{\mathrm{wm}}\mu_{\mathrm{wm}}}.

For each watermark strength, we first calculate the corresponding effective temperature TeffT_{\mathrm{eff}}. We then use the teacher model to generate unwatermarked responses at this temperature for the control group. We preserve the candidate-set construction, top-pp, training prompts, and student fine-tuning settings.

With τ=0.7\tau=0.7 and μwm=0.5\mu_{\mathrm{wm}}=0.5, the resulting temperature-matched controls achieve the following accuracies on the same 512 GSM8K test questions:

λwm\lambda_{\mathrm{wm}}TeffT_{\mathrm{eff}}Teacher acc. (%)Qwen acc. (%)Llama acc. (%)Llama-Instruct acc. (%)Detected controls
0.50.560072.6657.8124.2266.210/3
10.466774.2258.4024.8066.990/3
1.50.400071.4858.7931.6467.190/3
20.350073.4464.8426.5666.600/3
2.50.311173.2468.9529.3070.310/3
30.280071.8865.6232.2368.950/3

The strength in the first column identifies the corresponding watermarked setting; the controls contain no key-dependent watermark reward.

For each matched temperature, the “Detected controls” column shows how many of the three control students meet the watermark detection criterion. The watermark-inheritance conclusion remains unchanged. None of the 18 students trained on temperature-matched unwatermarked outputs satisfies the paper's detection criterion, Mkm>0M_k^m>0 and pwmm≤0.05p_{\mathrm{wm}}^m\leq0.05. Their detection statistics range from −0.400-0.400 to −0.240-0.240, with pwmmp_{\mathrm{wm}}^m ranging from 0.998 to 1.000. In comparison, all 18 watermarked students reported in Table B.1 in appendix satisfy the criterion, with positive statistics ranging from 0.492 to 1.333 and pwmm≈0.002p_{\mathrm{wm}}^m\approx0.002. Thus, the observed watermark inheritance cannot be explained solely by the effective temperature reduction. The accuracy results also show that temperature affects teacher and student utility, with different patterns across student backbones.

Reviewer_x5rTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Gradient-Guided Radioactive Watermarking (GGRW), a method for identifying whether a downstream model has been trained on protected teacher-generated data. It uses a proxy model to guide token selection during generation, favoring tokens that are expected to strengthen a key-dependent watermark signal through subsequent fine-tuning. The authors also develop a detection method that operates on held-out contexts not used during watermark construction or student training. Experiments on GSM8K demonstrate watermark inheritance across different student architectures, robustness to clean-data dilution, and improved held-out detectability compared with existing baselines.

Critical strengths
  1. Rather than relying solely on token-level watermark signals, the proposed method uses proxy-model gradients to guide watermark generation. This allows the embedding process to account for how the signal may be transferred during downstream fine-tuning.
  2. The method supports watermark detection using held-out contexts that are separate from those used for watermark construction and student training.
  3. The experiments include both Qwen and Llama student models and show that watermark signals can be detected across these architectures.
Critical weaknesses
  1. Limited dataset evaluation. All experiments are conducted on GSM8K, leaving the generalizability of GGRW to other tasks and open-domain text generation unclear.
  2. Lack of multi-key evaluation. Although the detector uses random wrong keys for significance testing, the paper does not evaluate whether multiple watermarks can be independently detected when a student is trained on data protected by different keys.
  3. Limited training settings. The experiments focus on LoRA fine-tuning. It remains unclear whether the watermark persists under continued pre-training on other unwatermarked datasets, limiting the conclusions about broader training scenarios.
Questions for authors
  1. Dataset generalization: Could the authors evaluate GGRW on an additional dataset, such as an instruction-following or open-domain generation benchmark, to demonstrate generalization beyond GSM8K?
  2. Multi-key detection: How does GGRW perform when a student is fine-tuned on a mixture of datasets watermarked with different keys? Can the detector reliably identify each key independently?
  3. Training settings: Does the watermark remain detectable under continued pre-training? Could the authors provide additional experiments or analysis to assess its persistence under different training regimes?
5 comments

Question 1:

Limited dataset evaluation. All experiments are conducted on GSM8K, leaving the generalizability of GGRW to other tasks and open-domain text generation unclear.

Answer 1:

Thank you for pointing out the limited dataset coverage. To evaluate GGRW beyond mathematical reasoning, we have added an experiment on OASST1, an open-domain conversational dataset. We use DeepSeek-R1-Distill-Qwen-7B as the teacher and Qwen2.5-3B as the proxy model and the student architecture. We use 256 questions for watermark-direction construction, 1024 for protected generation and matched student fine-tuning, and 512 separate questions for evaluation. Both the protected and unwatermarked students are fine-tuned using rank-8 LoRA for two epochs.

As shown in Table 1, the protected student satisfies the detection rule used in our paper (Mkm>0M_k^m>0 and pwmm≤0.05p_{\mathrm{wm}}^m \leq 0.05), with Mkm=0.0006M_k^m=0.0006 and pwmm=0.0100p_{\mathrm{wm}}^m=0.0100. Neither the unwatermarked student nor the base model satisfies this rule.

Table 1. Evaluation on OASST1. The unwatermarked and protected students are fine-tuned on unwatermarked and protected teacher completions, respectively, generated from the same 1024 questions. All models are evaluated on the same 512 held-out questions. Evaluation NLL is computed on the unwatermarked teacher completions for these questions; lower is better.

ModelMkmM_k^mpwmmp_{\mathrm{wm}}^mDetectedEvaluation NLL ↓\downarrow
Base student model-0.00250.0519No1.3349
Unwatermarked student-0.00200.1557No1.1455
Protected student+0.00060.0100Yes1.1619

We additionally report NLL measurements to characterize distributional changes associated with watermarking. The original unwatermarked teacher assigns mean per-token NLLs of 0.3653 and 0.3071 to the protected and unwatermarked training completions, respectively, consistent with a shift in generation preferences induced by watermarking. On the unwatermarked teacher completions for the 512 evaluation questions, the protected student obtains an NLL of 1.1619, compared with 1.1455 for the unwatermarked student and 1.3349 for the base model. These results indicate improved fitting to the reference responses after fine-tuning relative to the base model, with a modest observed NLL increase of 0.0164 for the protected student compared with the unwatermarked student.

These results extend our empirical evaluation beyond GSM8K to open-domain dialogue. We have added the experimental setup, detection results, and NLL measurements to Appendix C, “Additional Evaluation on OASST1.”

Question 2:

Lack of multi-key evaluation. Although the detector uses random wrong keys for significance testing, the paper does not evaluate whether multiple watermarks can be independently detected when a student is trained on data protected by different keys.

Answer 2:

Thank you for pointing out this missing evaluation. We agree that random wrong-key testing alone does not demonstrate whether multiple watermarks remain detectable after training on mixed-key data. To address this concern, we conduct an additional three-key experiment on GSM8K.

We fine-tune a single Qwen2.5-3B student for two epochs on a balanced mixture of datasets protected by three different keys, with 1024 responses per key. We then evaluate each key separately using its corresponding watermark gradient and 512 held-out detection events. All detection questions are disjoint from the training questions. Following the paper's detection protocol, we use 500 wrong-key trials per key and apply the detection criterion Mkm>0M_k^m>0 and pwmm≤0.05p_{\mathrm{wm}}^m\leq 0.05 separately to each key. As shown in Table 2, all three constituent watermarks are detected in the jointly trained student.

Table 2. Detection results for the student jointly trained on three watermarked datasets.

Detection keyMkmM_k^mpwmmp_{\mathrm{wm}}^mResult
k1k_10.48070.0020Detected
k2k_20.27770.0339Detected
k3k_30.30940.0140Detected

To further examine key specificity, we evaluate three students trained separately on the individual watermarked datasets using the same held-out detection inputs. As shown in Table 3, each student satisfies the detection criterion only for its own key, but not for the other two keys. The unmodified base model does not satisfy the detection criterion for any of the three keys.

Table 3. Single-key student controls. Bold entries satisfy Mkm>0M_k^m>0 and pwmm≤0.05p_{\mathrm{wm}}^m\leq 0.05.

Model / training dataDetect k1k_1: Mkm/pwmmM_k^m / p_{\mathrm{wm}}^mDetect k2k_2: Mkm/pwmmM_k^m / p_{\mathrm{wm}}^mDetect k3k_3: Mkm/pwmmM_k^m / p_{\mathrm{wm}}^m
Unmodified base model−0.1583/0.4271-0.1583 / 0.4271−2.0845/0.7685-2.0845 / 0.7685−1.9076/0.6128-1.9076 / 0.6128
k1k_1 only1.3264/0.0020\mathbf{1.3264 / 0.0020}−0.6176/0.6467-0.6176 / 0.6467−0.4912/0.6707-0.4912 / 0.6707
k2k_2 only−0.3773/0.9980-0.3773 / 0.99800.9854/0.0379\mathbf{0.9854 / 0.0379}0.8706/0.08980.8706 / 0.0898
k3k_3 only−0.2603/0.9222-0.2603 / 0.92220.9427/0.18360.9427 / 0.18360.9581/0.0020\mathbf{0.9581 / 0.0020}

These results show that the three constituent watermarks can be detected separately after mixed-key training, with the single-key controls providing additional evidence of key specificity.

Question 3:

Limited training settings. The experiments focus on LoRA fine-tuning. It remains unclear whether the watermark persists under continued pre-training on other unwatermarked datasets, limiting the conclusions about broader training scenarios.

Answer 3:

Thank you for highlighting this limitation. We conduct an additional experiment using the watermarked Qwen2.5-3B student (λwm=1.5\lambda_{\mathrm{wm}}=1.5). We merge its LoRA adapter into the base model and perform full-parameter continued pre-training on unwatermarked FineWeb-Edu text, using the standard next-token prediction objective. Continued pre-training covers 10 million tokens that contribute to the training loss, approximately 10.83 times the cumulative number of supervised tokens used during the original fine-tuning of the watermarked student.

We evaluate each checkpoint using the same held-out detection set, watermark key, and detection procedure. Following the paper, we report MkmM_k^m and pwmmp_{\mathrm{wm}}^m, with detection requiring Mkm>0M_k^m>0 and pwmm≤0.05p_{\mathrm{wm}}^m\leq 0.05.

Table 4. Watermark detection after full-parameter continued pre-training on unwatermarked FineWeb-Edu text.

Continued pre-training tokensMkmM_k^mpwmmp_{\mathrm{wm}}^mDetected
01.20730.0020Yes
1M1.28320.0020Yes
5M1.21940.0020Yes
10M1.22770.0020Yes

As shown in Table 4, the watermark satisfies the detection criterion at every checkpoint, and the detection statistic remains close to its initial value. The reported pp-values reach the minimum attainable with 500 wrong-key trials (1/501). These results provide evidence of watermark persistence under full-parameter continued pre-training on unwatermarked data, extending the evaluation beyond LoRA fine-tuning.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts in addressing my concerns. However, a few issues remain.

First, I noticed two seemingly different trends between the main paper and the rebuttal. In the multi-key experiment, all three keys remain detectable despite each accounting for only one-third of the training data, whereas detection fails at 30–40% protected-data ratios in the dilution experiments. Similarly, the watermark remains detectable after 10M clean pretraining tokens, while substantial clean-data mixing during initial fine-tuning prevents detection. More importantly, I share another reviewer's concern about the statistical calibration of the wrong-key significance test.

I appreciate the additional evidence, but given these remaining concerns, I maintain my original score for now.

Reviewer_rNSbOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a new way to embed "radioactive" watermarks in student models to establish distillation provenance. The method is based on those of prior work on antidistillation sampling and fingerprinting, with a new objective for the proxy model based on a hidden probe set of context segments.

Critical strengths
  • The use of a private probe set and the calculation of the reward with respect to it is nice.
  • The method maintains teacher utility quite well, which is good. I think the use of the probe set is a good way to maintain utility.
  • It's a substantial improvement over antidistillation fingerprinting in terms of detectability while also maintaining utility.
Critical weaknesses
  • They do require quite a lot of the training data to be marked with their technique; around half, it would appear from Figure 4.
  • Would be good to see tests on watermark removal attacks.
Questions for authors
  • It would be interesting to highlight the tokens that are part of the private probe set in the appendix.
  • I think including graphs for both the statistic and the p-value is quite redundant. I would rather see plots of teacher / student utility vs. detectability (or some such thing).
  • Can you provide more details on how the private probe set was chosen / more info on this generally.
  • I don't really understand the inclusion of the teacher gap penalty term. It's also referred to as the "teacher gap penalty" despite not being referred to as such in Section 3.3. This feels like it does not actually change the sampling distribution of tokens, i.e., the effect is surely absorbed by modulating lambda.
  • I don't understand why they took only the positive part of the delta. It makes some intuitive sense but is not really addressed in the paper.
1 comment

Question 1:

It would be interesting to highlight the tokens that are part of the private probe set in the appendix.

Answer 1:

Thank you for this suggestion. We have added “Illustrative Examples from the Private Probe Set” to Appendix D, including four probe events with their context prefixes, candidate sets, and reference next tokens. Green candidates are highlighted in bold. For convenience, we reproduce the examples in Table 1.

Table 1. Representative private probe events. All listed candidates belong to the corresponding probe event; green candidates are shown in bold. G and R indicate green and red assignments, respectively. “Reference next token” denotes the continuation in the original reference answer, rather than a token sampled by the protected generation procedure.

Context prefix ending before the probe positionGreen candidatesRed candidatesReference next token
… She receives an extra 7 dollars from a customer, increasing thetotalprofit, profits, amounttotal (G)
… On Thursday and Friday, there is asteadytotal, consistent, constanttotal (R)
… to get how much money is left over afterpurchasing, buyingshe, thepurchasing (G)
… After 10 years, growing at the rate of 3 inches every year, Haley will haveadded, increasedgrown, gainedgrown (R)

All listed candidates belong to the corresponding private probe event. Context prefixes are shortened. G and R denote green and red assignments, respectively. The reference next token is taken from the original reference answer.

These examples illustrate that the same token can receive different green/red assignments across contexts under the same key. For example, total is green in the first row but red in the second.

Question 2:

I think including graphs for both the statistic and the p-value is quite redundant. I would rather see plots of teacher / student utility vs. detectability (or some such thing).

Answer 2:

Thank you for this suggestion. Figures 3(a) and 3(b) share the same watermark-strength settings, allowing teacher/student utility and detectability to be compared at each setting. We have expanded the discussion in Section 4.2 to analyze these results jointly. Specifically, increasing λwm\lambda_{\mathrm{wm}} from 1.5 to 2 raises the detection statistics for all three students, while teacher accuracy decreases and student accuracy changes vary across models. This highlights the need to consider both utility and detectability when selecting the watermark strength. We have retained both the detection statistic and the wrong-key pp-value because the former measures the watermark response, while the latter indicates how extreme this response is relative to the statistics obtained using random wrong keys.

Question 3:

Can you provide more details on how the private probe set was chosen / more info on this generally.

Answer 3:

Thank you for requesting these details. We use GSM8K examples with indices ranging from 0 to 255 to supply reference solutions for constructing the watermark-gradient probes. A probe is a next-token context formed by concatenating a question prompt with a prefix of its GSM8K reference solution, retaining at most the last 256 tokens. For a reference solution of length LL tokens, we scan prefix offsets with a stride of max⁡(1,⌊L/8⌋)\max(1,\lfloor L/8\rfloor) and retain at most four eligible positions per solution. At each position, the candidate set consists of tokens among the teacher's top-64 next-token candidates whose unscaled logits are within 3.0 of the highest teacher logit. We retain the position only if this set contains both green and red tokens under the watermark key.

These probes are used to compute a watermark-gradient direction in the proxy model, which guides token selection during protected data generation. Requiring both green and red candidates allows the gradient to capture the relative probability preference. The resulting direction encourages the generated training data to carry a key-specific signal that can transfer to the student through distillation.

Comments on the paper

4 comments

Question 4:

I don't really understand the inclusion of the teacher gap penalty term. It's also referred to as the "teacher gap penalty" despite not being referred to as such in Section 3.3. This feels like it does not actually change the sampling distribution of tokens, i.e., the effect is surely absorbed by modulating lambda.

Answer 4:

Thank you for pointing this out. We have explicitly named the second term in Eq. (13) the “teacher-gap penalty” in Section 3.3.

To clarify its effect, let

rk(v∣c)=I[v∈Gk][Δwmk(v∣c)]+.r_k(v\mid c) = \mathbb{I}[v\in G_k]\bigl[\Delta_{\mathrm{wm}}^k(v\mid c)\bigr]_+.

Expanding Eqs. (13)–(14) gives

zwmk(v∣c)=(1+λwmμwm)zT(v∣c)τ+λwmrk(v∣c)−b(c),z_{\mathrm{wm}}^k(v\mid c) = (1+\lambda_{\mathrm{wm}}\mu_{\mathrm{wm}})\frac{z_T(v\mid c)}{\tau} + \lambda_{\mathrm{wm}}r_k(v\mid c) - b(c),

where

b(c)=λwmμwmzT(vmax⁡∣c)τb(c) = \lambda_{\mathrm{wm}}\mu_{\mathrm{wm}}\frac{z_T(v_{\max}\mid c)}{\tau}

is constant across candidates and cancels under softmax. Therefore, for a fixed candidate set, the distribution before top-pp sampling is

qk(v∣c)∝exp⁡[zT(v∣c)τeff+λwmrk(v∣c)],q_k(v\mid c) \propto \exp\left[\frac{z_T(v\mid c)}{\tau_{\mathrm{eff}}} + \lambda_{\mathrm{wm}}r_k(v\mid c)\right],

with

τeff=τ1+λwmμwm.\tau_{\mathrm{eff}} = \frac{\tau}{1+\lambda_{\mathrm{wm}}\mu_{\mathrm{wm}}}.

Thus, the penalty changes the distribution by sharpening the teacher component. This effective-temperature adjustment generally cannot be reproduced by changing λwm\lambda_{\mathrm{wm}} alone at a fixed τ\tau, because λwm\lambda_{\mathrm{wm}} also scales the key-dependent reward.

The penalty is included to favor candidates with smaller logit gaps from the teacher-preferred token. Table 3 shows that both Key only and Full are detected, whereas Gap only is not. Thus, the penalty serves as an auxiliary control on teacher preference and is not required for detection in the evaluated setting.

Question 5:

I don't understand why they took only the positive part of the delta. It makes some intuitive sense but is not really addressed in the paper.

Answer 5:

Thank you for highlighting this missing explanation. The positive-part operation follows the first-order analysis of token-level training effects in Section 3.2. From Eqs. (6)–(11), the change in the watermark response induced by learning a candidate token is approximately

Rk(θP+η∇θPlog⁡p(v∣c))−Rk(θP)≈η∥∇θPRk(θP)∥2Δwmk(v∣c).R_k\left(\theta_P + \eta\nabla_{\theta_P}\log p(v\mid c)\right) - R_k(\theta_P) \approx \eta\left\Vert\nabla_{\theta_P}R_k(\theta_P)\right\Vert_2\Delta_{\mathrm{wm}}^k(v\mid c).

Thus, a positive alignment score predicts a local increase in the target watermark response, whereas a negative score predicts a decrease. We use

max⁡(0,Δwmk(v∣c))\max(0,\Delta_{\mathrm{wm}}^k(v\mid c))

to provide an additional reward only to green candidates with positive estimated alignment. Candidates with nonpositive scores receive no contribution from this reward term.

Thanks for your response. Maintaining my "clear accept".

One additional thing that might help the presentation would be to note that the derivation essentially proceeds along the same lines as previous work. And point out at a high level before beginning where the derivation differs. This might allow some of the derivation to be moved to the appendix in exchange for more experimental details in the body. It would also highlight the exact novelty of this work.

Thank you for your helpful suggestions and continued support of our work. We will incorporate your feedback into the camera-ready version to improve the clarity and presentation of the paper.

Thank you for your detailed feedback and for highlighting these remaining concerns.

Regarding statistical calibration, Equation (18) compares the target-key statistic with statistics obtained using random wrong keys. In the original procedure, the target key contributes to constructing the watermark direction and selecting detection events, while the reference keys are evaluated with the target direction and alignment scores fixed. Independently sampling reference keys therefore does not by itself establish a calibrated p-value. We interpret Equation (18) as an empirical wrong-key ranking score and assess its false-positive behavior through independent clean-model evaluation.

We conducted 500 independent clean-model repetitions. Each repetition used a newly trained unwatermarked Qwen2.5-3B student, a separately sampled target key, and 500 wrong-key references. The clean training corpus, construction and detection inputs, model configuration, and detector implementation were held fixed.

Under the original decision rule, M>0M>0 and pwm≤0.05p_{\mathrm{wm}}\leq0.05, we observed 10 false positives in 500 trials (2.0%). When applying pwm≤0.05p_{\mathrm{wm}}\leq0.05 alone, we also observed 10 rejections in 500 trials (2.0%). For both rules, the one-sided 95% exact binomial upper confidence bound was 3.37%, below 5%. Thus, the evaluation directly supports thresholding the wrong-key ranking score itself, as well as the complete detection rule.

These results provide independent statistical support for a false-positive rate below 5% at the specified decision threshold under the evaluated protocol. We will describe Equation (18) as a wrong-key ranking score whose false-positive behavior at the selected threshold is supported by this independent evaluation.

Regarding practical utility and generation overhead, we believe the practical contribution of GGRW should be assessed jointly in terms of task utility, computational cost, and the detection setting it addresses. The additional computation during generation directly supports watermark inheritance through downstream learning. Proxy-guided token selection favors updates aligned with a private watermark direction. Under the evaluated held-out protocol, GGRW satisfies the detection criterion in settings where the tested KGW and ADFP configurations remain undetected. This detection capability provides a concrete practical benefit that should be weighed against the generation overhead.

Watermark strength also allows configurations to be selected according to the desired balance between task utility and detectability. The practical value of this trade-off depends on the provenance requirements of the intended application.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.