- 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
-
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. -
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. -
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. -
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, , 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. -
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. -
The teacher-gap term confounds watermark strength with sampling temperature.
After removing the token-independent constant, Equations 13–14 scale the teacher logits by . 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. -
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. -
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
-
How exactly are the candidate sets , green partitions, and key-dependent hash constructed? What are the candidate-set size, green fraction, and value of ?
-
What exchangeability argument justifies Equation 18 when a wrong key changes the green partition but not the watermark direction or alignment scores?
-
What is the empirical false-positive rate across multiple clean models, target keys, training seeds, and data splits?
-
How is computed when is a Qwen token but the suspect model uses the Llama tokenizer?
-
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?
-
Do the conclusions remain unchanged when each watermark strength is compared with a control using the same effective sampling temperature?
-
How does GGRW perform after paraphrasing, continued clean fine-tuning, full-parameter fine-tuning, model merging, quantization, or watermark scrubbing?
-
What are the generation throughput, memory overhead, and total computational cost relative to KGW and ADFP?
6 comments
Question 4:
How is computed when 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 denote the Llama token ID whose vocabulary token string matches that of Qwen token , and define the retained candidate set as
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 , we compute
where 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 and ; otherwise, detection is No. **The detection key, detection contexts, and wrong-key set remain fixed across runs.
| Run | Student | Accuracy | Detected | ||
|---|---|---|---|---|---|
| 1 | Watermarked | 60.74% | 1.1238 | 0.0020 | Yes |
| 1 | Unwatermarked | 64.65% | 0.9760 | No | |
| 2 | Watermarked | 56.84% | 1.1786 | 0.0020 | Yes |
| 2 | Unwatermarked | 60.94% | 0.9721 | No |
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 model | Student accuracy | Detected | ||||
|---|---|---|---|---|---|---|
| Qwen2.5-3B | 1.5 | 1.0 | 69.92% | 1.2197 | 0.0020 | Yes |
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
For each watermark strength, we first calculate the corresponding effective temperature . We then use the teacher model to generate unwatermarked responses at this temperature for the control group. We preserve the candidate-set construction, top-, training prompts, and student fine-tuning settings.
With and , the resulting temperature-matched controls achieve the following accuracies on the same 512 GSM8K test questions:
| Teacher acc. (%) | Qwen acc. (%) | Llama acc. (%) | Llama-Instruct acc. (%) | Detected controls | ||
|---|---|---|---|---|---|---|
| 0.5 | 0.5600 | 72.66 | 57.81 | 24.22 | 66.21 | 0/3 |
| 1 | 0.4667 | 74.22 | 58.40 | 24.80 | 66.99 | 0/3 |
| 1.5 | 0.4000 | 71.48 | 58.79 | 31.64 | 67.19 | 0/3 |
| 2 | 0.3500 | 73.44 | 64.84 | 26.56 | 66.60 | 0/3 |
| 2.5 | 0.3111 | 73.24 | 68.95 | 29.30 | 70.31 | 0/3 |
| 3 | 0.2800 | 71.88 | 65.62 | 32.23 | 68.95 | 0/3 |
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, and . Their detection statistics range from to , with 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 . 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.
Question 1:
How exactly are the candidate sets Ci, green partitions, and key-dependent hash constructed? What are the candidate-set size, green fraction, and value of ξ?
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 below.
For context ci, let zT(v∣ci) denote the teacher's unscaled next-token logit. We construct candidate sets as
Ci={v∈Top64(zT(⋅∣ci)):maxuzT(u∣ci)−zT(v∣ci)≤3.0}.
Thus, the candidate-set size is context-dependent and capped at 64.
To construct the partition, let ai be the last Qwen token ID in ci. For each candidate v, represented by its Qwen token ID, we compute
hk(ci,v)=264−1Hash64(s,k,ai,v),
where s is a fixed salt. The function Hash64 computes the SHA-256 digest of the serialized tuple (s,k,ai,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.
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 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>0 and pwmm≤0.05.
The evaluation covers:
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.