acceptodds
Under review as a conference paper at ICLR 2027

LLM Output Detectability and Task Performance Can be Jointly Optimized

Abstract

Detecting machine-generated text is essential for transparency and accountability when deploying LLMs. Watermarking enables statistically reliable detection by biasing token distributions to embed detectable signals into LLM outputs. However, it has been reported that watermarked LLMs often perform worse on downstream tasks. We propose PUPPET, a framework that fine-tunes an LLM via DPO to generate text that is both more detectable by a target detector and better performing on downstream tasks. We use two rewards: a detector that outputs a machine-class likelihood and an evaluator that measures a task-specific metric. Just as a watermark is verified with its secret key, this detector-specific design lets an LLM provider track how its published model is used. Experiments on long-form QA, summarization, and essay writing show that LLMs trained with PUPPET achieve detectability competitive with watermarking methods—even at strict low FPRs—while outperforming them on downstream tasks. Moreover, this optimization requires only a few thousand samples and 1–2 GPU hours, and its gains hold across out-of-domain tasks, six detectors of diverse architectures, and different LLM families and sizes, and are even robust to paraphrasing attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine-generated text detection, watermarking, large language models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jzTtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a DPO method to fine-tune an LLM and train a detector simultaneously, enabling the detector to detect text generated by the LLM and improving the LLM's downstream performance. It offers a potential new approach to machine-generated text detection.

Critical strengths
  1. Introducing DPO as a solution to achieve a better trade-off between detectability and downstream task performance of machine-generated text.
  2. Experiments cover many aspects, providing a comprehensive assessment of the method's applicability and potential limitations.
Critical weaknesses
  1. The potential use of the proposed method is limited. It requires fine-tuning an LLM on a specific task.
  2. The metrics for task performance evaluation and the DPO training appear to use the same metric. This may introduce a systematic bias in evaluation and be unfair to baseline methods.
  3. Using the OpenAI detector on the base model without PUPPET fine-tuning as Vanilla may exaggerate the improvement brought by PUPPET proposed by this paper. The authors may consider comparing other fine-tuning baselines such as [1],
  4. The trade-off between detectability and downstream task performance is a challenge for watermarking. But the method proposed in this paper is not watermarking. So, the paper should identify its own potential trade-off or limitation rather than compare it with those challenges of watermarking.

[1] Xiaojun Xu, Yuanshun Yao, Yang Liu. “Learning to Watermark LLM-generated Text via Reinforcement Learning.” 2024.

Questions for authors
  1. Are there any other task-irrelevant metrics that can be used to inspect the influence of fine-tuning an LLM with PUPPET, or compare it with an LLM without PUPPET?
  2. Do different tasks have different difficulty levels to become detectable via DPO? Can the authors provide any insight?
  3. How long of a text is used to evaluate the detectability and downstream performance in this paper?
  4. How do the authors calibrate the false positive rate of detection during evaluation?
4 comments
AuthorsOpenReview ↗

Response to Reviewer jzTt (1/2)

We thank you for recognizing PUPPET as a stable way to improve both detectability and task performance, and the coverage of our experiments. We would like to clarify one point in the summary: PUPPET does not train a detector. The detector is fixed, and only the LLM is fine-tuned (§2). We respond to each comment below. To address each point fully, we reply in two parts.

W3.1

We respectfully disagree and consider PUPPET sufficiently practical. PUPPET is designed to be fine-tuned per task, but the cost is small: the whole pipeline, from building the preference data to fine-tuning, fits within 12 hours, and fine-tuning itself takes 1–2 GPU hours (§5.5). Because the preference data are built offline, ablating α for a given use case requires no regeneration. Moreover, the gain in detectability generalizes to tasks not used in training (§5.1). Our results also indicate that a single model can be trained on a mixture of tasks (App. J).

W3.2, Q3.1

Watermarking methods do not optimize a task metric. Optimizing it explicitly, together with detectability, is what PUPPET contributes. The intended use case is therefore one where the metric to optimize is given.

As an additional analysis, to check that the gain is not specific to the reward metric, we evaluated with metrics that were not used as rewards.

MethodPerplexity under Qwen3-8B ↓ (ELI5 / Multi-News / IELTS)BARTScore [1] ↑ (ELI5 / Multi-News)
Vanilla4.68 / 7.36 / 3.32−3.72 / −2.98
KGW3.95 / 6.88 / 2.92−3.66 / −2.96
PUPPET3.80 / 5.59 / 3.00−3.54 / −2.86

Perplexity and BARTScore also show the improvement in task performance over Vanilla (BARTScore: Tab. 7, App. A). The model tuned with PUPPET is thus not overfitted to the given metric.

W3.3

Vanilla as the reference. PUPPET does not compete with the detector. Whether the detector is already strong or still weak, PUPPET boosts its detection performance while improving task performance. For Llama-3 on ELI5, Vanilla already shows high detectability, and PUPPET improves task performance without lowering it (Tab. 1). For Qwen3-8B on Multi-News, Vanilla shows relatively low detectability, and PUPPET largely improves it together with task performance (§5.2, Tab. 3). Fine-tuning alone does not explain this: the task reward alone provides no reliable gain in detection (§5.4).

Other fine-tuning baselines. As a training-based baseline, we added Wong et al. [2] on Llama-3. Each cell shows AUROC / task metric (ROUGE-L for ELI5 and Multi-News, judge score for IELTS).

MethodELI5Multi-NewsIELTS
Vanilla97.1 / 22.992.8 / 26.587.0 / 6.77
KGW99.6 / 23.599.7 / 26.699.9 / 6.75
Wong et al. [2]99.3 / 23.696.5 / 26.596.2 / 6.54
PUPPET100.0 / 25.899.8 / 28.099.3 / 7.19

PUPPET is above Wong et al. in both AUROC and task performance on the three tasks. Wong et al. also train with a detection objective and a quality objective, yet their task performance stays close to that of Vanilla. The results of PUPPET therefore do not follow trivially from existing training-based methods.

We did not run Xu et al. [3] because its 2023 implementation targets OPT-1.3B and Llama-2, and stabilizing its RL-based training on our models was not feasible within the rebuttal period.

W3.4

In machine-generated text detection, watermarking is the de facto standard, and its trade-off with task performance is the central challenge (§1). Our claim is that optimizing the LLM for both detectability and task performance can resolve this challenge and may open a new paradigm for machine-generated text detection.

As for the trade-off of PUPPET itself, the reward weight α balances the two objectives (§5.4, App. I): a trade-off exists along α, but PUPPET occupies a more favorable region than Vanilla, and it exceeds the watermarking baselines in task performance while matching their detection (Tab. 1). We will also state the limitations of PUPPET explicitly: its detectability is established empirically for a target detector, not guaranteed statistically as in watermarking, and robustness beyond naive paraphrasing is left for future work.


References

[1] Yuan et al. "BARTScore: Evaluating Generated Text as Text Generation." NeurIPS 2021.

[2] Wong et al. "An End-to-End Model for Logits-Based Large Language Models Watermarking." ICML 2025.

[3] Xu et al. "Learning to Watermark LLM-generated Text via Reinforcement Learning." arXiv 2024.

AuthorsOpenReview ↗

Response to Reviewer jzTt (2/2)

This is the second part of our response, covering the questions.

Q3.2

Yes. Across base LLMs and detectors, the detectability of Vanilla on IELTS is consistently lower than on the other tasks, which makes joint optimization more challenging (App. I). IELTS thus demands the largest AUROC gains (up to +28.5 points for Qwen3; §5.2) and more training samples to approach its ceiling (App. E). These results suggest that, when Vanilla is hard to detect, raising detectability is harder and tends to limit the gain in task performance.

Q3.3

All methods generate under the same limit on new tokens: 300 for ELI5 and 512 for Multi-News, following WaterBench [1], and 2,500 for IELTS, which covers 1.5 times the longest human essay. Detectability and task performance are measured on the full generated text.

The average length in words is as follows.

MethodELI5Multi-NewsIELTS
Vanilla237281400
KGW244291406
PUPPET253334411

Q3.4

In the submission, the thresholds for 1% and 0.1% FPR were determined on the evaluation set itself, with its human references as negatives (200–600 texts; §5.3).

We increased the human negatives to 15,000 texts for ELI5 and 5,000 for Multi-News and recomputed TPR for every generated text. Brackets show 95% Wilson confidence intervals.

MethodELI5 TPR@1%ELI5 [email protected]%Multi-News TPR@1%Multi-News [email protected]%
Vanilla48.533.347.016.5
KGW94.8 [92.8, 96.3]86.3 [83.4, 88.9]93.5 [89.2, 96.2]73.5 [67.0, 79.1]
PUPPET95.5 [93.5, 96.9]89.3 [86.6, 91.6]93.0 [88.6, 95.8]75.0 [68.6, 80.5]

PUPPET and KGW lie within each other's confidence intervals at both FPR levels on both tasks. We will replace Tab. 5 with this table.


References

[1] Tu et al. "WaterBench: Towards Holistic Evaluation of Watermarks for Large Language Models." ACL 2024.

Thank the authors for their response. My questions have been partially addressed, and I now have a better understanding of the underlying principles and mechanisms of PUPPET. I agree that the task-specific fine-tuning approach proposed by PUPPET has the potential to simultaneously improve both detectability and downstream task performance. However, I still feel that the current work lacks sufficient depth in analyzing the underlying sources of these gains as well as the potential trade-offs or costs involved. Therefore, I have decided to maintain my original score.

AuthorsOpenReview ↗

Response to Reviewer jzTt follow-up comment

Thank you for the follow-up. We would be glad to address your interest in PUPPET as fully as we can, and we respond to the two remaining points below.

Underlying sources of the gains

We read this point as the following question: why does applying PUPPET (DPO with a detection reward and a task reward) make the outputs of a model both more detectable and better on the task? If you have a specific analysis in mind, please let us know.

Changes in the outputs. We compared the outputs of Llama-3 on ELI5 before and after PUPPET (Fig. 2, §4; setup in App. A.6). After PUPPET, the coverage of the spans that the OpenAI Detector relies on increases by 1.8%, and the coverage of the spans that ROUGE-L rewards increases by 13%. This suggests that the model learns expressions that serve each objective, which makes its outputs more detectable and better on the task.

Compatibility of the two rewards. The reward ablation shows that both rewards are needed (Fig. 4, §5.4). We also analyzed the detector scores and the task scores of the candidate responses in the training data. Within each prompt, the two scores are nearly uncorrelated (Pearson r between −0.04 and +0.09 across four base LLMs and three tasks), so the two rewards do not systematically conflict. As a result, in about 90% of the preference pairs built with the combined score (§2), the chosen response is higher than the rejected one in both scores. DPO therefore steers the model toward outputs that satisfy both objectives.

Potential trade-offs or costs

The paper already discusses these points as summarized below. If any of them remains unclear, please tell us which one.

Trade-off between detectability and task performance. This trade-off is a central challenge in machine-generated text detection (§1). With joint optimization, PUPPET improves both detection and task performance over Vanilla on the three tasks (Tab. 1, §4). A trade-off remains along the reward weight α, but PUPPET occupies a more favorable region than Vanilla (App. I; W3.4 in our first response).

Task specificity: performance on other tasks and computational cost. PUPPET optimizes an explicitly given target task. On tasks not used in training, the gain in detection transfers without degrading task performance (Tab. 2, §5.1). We also observed no significant degradation on math, coding, and agent benchmarks after applying PUPPET (see our response to Reviewer fv4T on 19 Nov). PUPPET requires training for each target task, but the cost in data and compute is small (Fig. 5, §5.5), and a single model can also be trained on a mixture of tasks (App. J).

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

This paper introduces PUPPET, a method that fine-tunes generative models to simultaneously enhance both the detectability of their outputs and the quality of their task performance. For each prompt, the method generates five candidate responses; these are then scored for detectability and quality, normalized, and combined into a weighted score. Preference pairs are constructed by selecting the highest- and lowest-scoring responses, which are subsequently used for training via DPO (Direct Preference Optimization) and LoRA (Low-Rank Adaptation). During inference, the model follows a standard generation process, while detection is performed by a fixed detector that assigns a score to the text and applies a threshold for classification.

Experiments cover ELI5 question answering, Multi-News summarization, and IELTS essay writing. The scope is further extended to include the Qwen3 model, various detection tools, cross-task transfer, and robustness against rewriting attacks. Llama-3 demonstrated improvements in both metrics across all three primary tasks; for instance, on the ELI5 dataset, the AUROC increased from 97.1 to 100.0, and the ROUGE-L score rose from 22.9 to 25.8. Experiments involving multiple detectors require task-specific adaptation, and the findings are primarily specific to the detector used during the training phase.

Critical strengths
  1. Clear evidence for dual-objective improvement. The method yields gains in both detectability and quality metrics across all three primary tasks. Ablation studies on the reward components, along with comparisons against a baseline of "quality-only fine-tuning + KGW," further substantiate the value of joint optimization.

  2. Elegant and practical approach. The method leverages existing scorers and requires only a small amount of preference data for offline training. It introduces no additional overhead during the decoding/inference stage. Its applicability is further supported by experiments across diverse models, tasks, and detection systems.

Critical weaknesses
  1. Methodological innovation is primarily incremental. Nicks et al. (2024), already cited in Section 6 of the manuscript, construct DPO preferences using detector scores with the goal of reducing detectability. Xu et al. [1] have also proposed joint training with detection and task-specific rewards. The unique value of this paper’s fixed detector, offline preference construction, and task adaptation requires further clarification relative to these existing approaches. [1] Xu, X., Yao, Y., & Liu, Y. (2024). Learning to Watermark LLM-generated Text via Reinforcement Learning. arXiv preprint arXiv:2403.10553.

  2. Insufficient comparison with training-based baselines. While Table 1 focuses on decoding-time watermarking, and Section 5.4 includes a "quality optimization + KGW" baseline, the study lacks a controlled comparison against similar training-based methods. Although the current experiments demonstrate the effectiveness of the overall framework, it remains difficult to isolate the specific performance and cost advantages contributed by the individual design choices.

  3. Lack of detail regarding evaluation at extremely low false-positive rates (FPR). Section 3 and Table 9 indicate that only 200 evaluation prompts were used per task; without additional human-generated negative samples, this sample size is insufficient to reliably estimate an FPR of 0.1%. For Table 5, it is necessary to specify the size of the calibration set, the method used to determine thresholds, and the confidence intervals, in order to substantiate conclusions drawn under the condition of a strictly low false-positive rate.

Questions for authors
  1. Given the same base model, task data, and a comparable total budget, what are the advantages of PUPPET—in terms of the detection-quality trade-off and end-to-end costs—when compared to the training-based framework proposed by Xu et al. or a PPO-based approach utilizing the same dual-reward objective?

  2. Regarding the 0.1% FPR results in Table 5: how many human-generated negative samples were used, on which dataset was the threshold determined, and what are the confidence intervals for the FPR and TPR during independent testing (after the threshold was fixed)?

1 comment
AuthorsOpenReview ↗

Response to Reviewer KGrM

We thank you for recognizing that the method improves both objectives on all three tasks, and that it is simple and adds no cost at inference. We respond to each comment below.

W2.1

Thank you for raising this important comparison.

Nicks et al. [1] use no task reward, and their goal is to evade detectors. Xu et al. [2] can combine a task reward with the score of a jointly trained detector, but their goal and claim are to preserve the utility of the non-watermarked model ("negligible distortion on the original utility"; caption of their Tab. 1). Our motivation is that existing methods, including these and conventional watermarking, at most preserve the quality of non-watermarked text and do not explicitly improve task performance. We therefore proposed a framework that optimizes detectability and task performance jointly. It keeps an existing detector fixed and builds preference pairs offline, which makes training stable and inexpensive (W2.2).

Our results are also not readily expected from these studies. That an LLM can be trained to be detectable while preserving its utility [2] does not imply that its task performance can also be improved. Likewise, that an LLM can be tuned to evade a detector [1] does not imply that it can be made more detectable while also improving task performance.

We will state these differences clearly in §1 and §6.

W2.2, Q2.1

Training-based baselines. We evaluated Wong et al. [3] on Llama-3. Each cell shows AUROC / task metric (ROUGE-L for ELI5 and Multi-News, judge score for IELTS).

MethodELI5Multi-NewsIELTS
Vanilla97.1 / 22.992.8 / 26.587.0 / 6.77
KGW99.6 / 23.599.7 / 26.699.9 / 6.75
Wong et al. [3]99.3 / 23.696.5 / 26.596.2 / 6.54
PUPPET100.0 / 25.899.8 / 28.099.3 / 7.19

PUPPET is above Wong et al. in both AUROC and task performance on the three tasks.

We did not run Xu et al. [2]: the released implementation targets OPT-1.3B and Llama-2 with a software stack from 2023, and stabilizing its RL-based training on our models was not feasible within the rebuttal period.

Why DPO rather than PPO. In a pilot study, we used PPO instead. Each run took 15–18 GPU hours, training was hard to stabilize, and the generations collapsed to a few tokens in several runs. DPO with offline preference pairs is stable and takes 1–2 GPU hours per run (§5.5).

Training cost. PUPPET requires training for each design choice, but the whole pipeline fits within 12 hours (§5.5). The preference data are built once: ablations and adjustments of the reward weight α require no regeneration. A single model trained on a mixture of the three tasks reaches detectability comparable to per-task training after 5,000 samples (App. J).

W2.3, Q2.2

Thank you for this important feedback.

Submission. Tab. 5 used no separate calibration set. The threshold was determined on the evaluation set itself, with its human references as negatives: 600 texts for ELI5 (200 prompts × 3 references), 200 for Multi-News, and 200 for IELTS. TPR at 1% and 0.1% FPR was read from the ROC curve by linear interpolation, without confidence intervals.

New evaluation. We increased the human negatives to 15,000 texts for ELI5 and 5,000 for Multi-News, as these two datasets include enough human texts to measure TPR at 0.1% FPR. TPR at 1% and 0.1% FPR is computed against these negatives, with every generated text of the evaluation set scored individually. Brackets show 95% Wilson confidence intervals.

MethodELI5 TPR@1%ELI5 [email protected]%Multi-News TPR@1%Multi-News [email protected]%
Vanilla48.533.347.016.5
KGW94.8 [92.8, 96.3]86.3 [83.4, 88.9]93.5 [89.2, 96.2]73.5 [67.0, 79.1]
PUPPET95.5 [93.5, 96.9]89.3 [86.6, 91.6]93.0 [88.6, 95.8]75.0 [68.6, 80.5]

PUPPET and KGW lie within each other's confidence intervals at both FPR levels on both tasks: with enough negatives, PUPPET matches KGW in the high-stakes setting. We will replace Tab. 5 with this table and revise §5.3 accordingly.

Thresholds fixed on separate data. These negatives are separate from the evaluation set, so the thresholds are fixed before testing. The table shows FPR (%) on the human references of the evaluation set.

MethodELI5 @1%ELI5 @0.1%Multi-News @1%Multi-News @0.1%
KGW2.33 [1.39, 3.88]0.17 [0.03, 0.94]0.50 [0.09, 2.78]0.00 [0.00, 1.88]
PUPPET0.17 [0.03, 0.94]0.00 [0.00, 0.64]0.00 [0.00, 1.88]0.00 [0.00, 1.88]

The observed FPR of PUPPET does not exceed the nominal level, so the conclusion above is unchanged.


References

[1] Nicks et al. "Language Model Detectors Are Easily Optimized Against." ICLR 2024.

[2] Xu et al. "Learning to Watermark LLM-generated Text via Reinforcement Learning." arXiv 2024.

[3] Wong et al. "An End-to-End Model for Logits-Based Large Language Models Watermarking." ICML 2025.

Reviewer_fv4TOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PUPPET, a framework that fine-tunes an LLM through DPO to jointly improve output detectability and downstream task performance. For each query, candidate responses are ranked by a weighted combination of normalized detector and task-evaluation scores to construct preference pairs. Experiments on long-form QA, summarization, and essay generation show improved task scores and strong detection performance, with additional evaluations across model families, detectors, and paraphrasing attacks.

Critical strengths
  1. The problem is relevant, and the proposed framework is simple and easy to follow. Jointly optimizing detectability and task quality is a reasonable direction for improving the practical utility of detectable text generation.
  2. The experimental analysis covers several useful settings, including Qwen3 models of different sizes, alternative detectors, low-FPR detection, and reward ablations. The task-only training followed by KGW comparison in Section 5.4 is helpful for examining whether the two objectives should be optimized together.
  3. The reported training cost is modest. Section 5.5 also separates candidate generation, reward scoring, and fine-tuning costs, which helps assess the full pipeline.
Critical weaknesses

1.The related-work discussion and baseline comparison omit Wong et al., An End-to-End Model for Logits-Based Large Language Models Watermarking (ICML 2025). This work jointly optimizes a watermark encoder and decoder to balance text quality and detection robustness, directly addressing a motivation central to PUPPET. 2.The analogy with keyed watermarking is insufficiently justified. PUPPET optimizes outputs toward an AI-text detector but does not establish a provider-specific watermark or a statistical detection guarantee analogous to the green-token count test in randomized green-list watermarking. 3. The paraphrasing results do not clearly isolate robustness learned through PUPPET from detectability introduced by the paraphraser. In Table 17, vanilla Llama-3 outputs already achieve AUROCs of 98.8, 98.8, and 92.9 after Dipper paraphrasing. The model do not explicitly model the text editing process. 4. The evidence for general task-quality improvement is limited by evaluation using the same metrics that provide the training rewards. ROUGE-L measures reference overlap, while the IELTS experiment uses the same judge model for reward construction and evaluation. 5. The method is less flexible to deploy than decoding-time watermarking. PUPPET requires access to model weights, task-specific preference construction, and DPO fine-tuning; applying it to an already deployed model therefore requires an additional tuning and model-update step. In contrast, conventional logit- or sampling-based watermarking can often be added to an existing generation pipeline without retraining the base model, provided that the required decoding access is available.

Questions for authors
  1. Can PUPPET distinguish between two models trained to maximize the same detector score? How does attribution perform when the negative class includes other PUPPET-trained models and paraphrased human text?
  2. Do the task-quality and paraphrasing gains persist under independent quality evaluation and other paraphrasers, including an adaptive attack against the target detector?
8 comments
AuthorsOpenReview ↗

Response to Reviewer fv4T (1/2)

We thank you for recognizing the relevance of the problem, the breadth of the analysis, and the separate reporting of the training cost. We respond to each comment below. To address each point fully, we reply in two parts.

W1.1

Thank you for the pointer. We will add Wong et al. [1] to §6.

Relation to PUPPET. The motivation is similar: Wong et al. also optimize a detection objective and a quality objective jointly. The goal differs. They aim to keep the quality of watermarked text on par with that of non-watermarked text. PUPPET instead adds task quality to the reward and explicitly optimizes both detectability and task performance.

What is optimized also differs. Wong et al. train an external encoder and decoder, whereas PUPPET fine-tunes the LLM itself. Their encoder runs at every decoding step and is reported to increase generation time by 60.6% on Llama2-7B (their Tab. 18). PUPPET adds no step at decoding.

Results. We also ran it as a baseline on Llama-3, using the checkpoint released by the authors through MarkLLM. Each cell shows AUROC / task metric (ROUGE-L for ELI5 and Multi-News, judge score for IELTS).

MethodELI5Multi-NewsIELTS
Vanilla97.1 / 22.992.8 / 26.587.0 / 6.77
KGW99.6 / 23.599.7 / 26.699.9 / 6.75
Wong et al. [1]99.3 / 23.696.5 / 26.596.2 / 6.54
PUPPET100.0 / 25.899.8 / 28.099.3 / 7.19

Under this setting, PUPPET is above Wong et al. in both AUROC and task performance on the three tasks.

These results suggest that joint optimization alone is not sufficient. The results of PUPPET therefore do not follow trivially from those of Wong et al.

W1.2

We would like to clarify the intent of this analogy. PUPPET is designed to enhance detectability with respect to a specific target detector, not detectability against all detectors. This is analogous to watermarking, which embeds signals detectable only with a specific key. The analogy thus describes the use case: a model developer builds a private detector alongside the model, tunes the model toward it before publishing, and can then trace the outputs of its own model, just as the holder of a watermark key can.

These are not fully equivalent notions of detectability: a watermark's detectability is guaranteed statistically by design, whereas that of PUPPET is established empirically. We will make the intent of the analogy and this distinction explicit in the abstract and §1.

W1.3, Q1.2 (paraphrasing)

Thank you for this insightful comment, which helped us sharpen the scope of our robustness claim.

What the Dipper results show. Watermarking methods such as KGW [2] rely on surface-level token distributions only, which paraphrasing disrupts. The OpenAI Detector can also use other features such as meaning and style, so it still detects Vanilla outputs after Dipper [3] (Tab. 17). PUPPET is built on this detector and inherits the property: it keeps its AUROC after Dipper, whereas the watermarking baselines drop sharply (Fig. 3).

A newer and stronger paraphraser. Dipper is a widely used paraphrasing attack in this field (§5.2). To examine robustness further, we also ran Qwen3-8B as a paraphraser, which is newer and stronger than Dipper.

ELI5 AUROCELI5 TPR@1%Multi-News AUROCMulti-News TPR@1%
Vanilla80.916.085.019.5
KGW78.018.079.316.0
PUPPET90.542.394.645.0

All values are measured after paraphrasing. On these two tasks, PUPPET retains higher detectability than both Vanilla and KGW. The detector flags only 8.2% (ELI5) and 4.0% (Multi-News) of human text paraphrased by it, so the detectability introduced by this paraphraser is small. Accordingly, with this paraphrased human text as the negative class, the AUROC of PUPPET remains 98.5 (ELI5) and 98.8 (Multi-News), which also relates to Q1.1.

Scope of the claim. Our experiments establish robustness to naive paraphrasing. We do not overclaim: this does not imply robustness to stronger, detector-evading attacks [4], which is a separate, challenging problem. We will clarify this scope in the abstract and §5.2, and note adversarial rewriting attacks as future work.


References

[1] Wong et al. "An End-to-End Model for Logits-Based Large Language Models Watermarking." ICML 2025.

[2] Kirchenbauer et al. "A Watermark for Large Language Models." ICML 2023.

[3] Krishna et al. "Paraphrasing Evades Detectors of AI-Generated Text, but Retrieval Is an Effective Defense." NeurIPS 2023.

[4] Cheng et al. "Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text." arXiv 2025.

AuthorsOpenReview ↗

Response to Reviewer fv4T (2/2)

This is the second part of our response.

W1.4, Q1.2 (independent quality evaluation)

Unlike watermarking methods, PUPPET explicitly optimizes a task metric together with detectability. We agree that the gain should also be checked with metrics that were not used as rewards.

MethodPerplexity under Qwen3-8B ↓ (ELI5 / Multi-News / IELTS)BARTScore [1] ↑ (ELI5 / Multi-News)
Vanilla4.68 / 7.36 / 3.32−3.72 / −2.98
KGW3.95 / 6.88 / 2.92−3.66 / −2.96
PUPPET3.80 / 5.59 / 3.00−3.54 / −2.86

In both metrics, PUPPET improves over Vanilla, in agreement with the reward metrics (BARTScore is reported in Tab. 7, App. A).

For IELTS, a judge model different from the reward model is a further check. We could not complete it within the rebuttal period and will add it to the final version.

W1.5

PUPPET assumes a model provider as its user (§1), for whom access to the model weights is not a burden. On the other points:

Task-specific construction. A single model can be trained on a mixture of the three tasks. It reaches detectability comparable to per-task training after 5,000 samples (App. J).

Cost of fine-tuning. The preference data are built once and are reused across ablations, and fine-tuning takes 1–2 GPU hours (§5.5).

Constraints at inference. Decoding-time watermarking needs no training, but it intervenes at every decoding step. This constrains the inference stack: in MarkLLM, the vLLM integration covers 3 methods (UPV, KGW, and Unigram), and the others run on Transformers. In our ELI5 runs, K-SemStamp, which generates sentence by sentence with rejection sampling, took about 290 s per response, whereas vLLM generated all 600 unwatermarked responses in about 50 s (not a controlled comparison). Once trained, a PUPPET model is served as an ordinary model with no constraint on decoding. The provider bears a one-time cost, and users bear none.

Q1.1

As stated in §5.6, model attribution with PUPPET is an exploratory analysis, and the results there suggest its feasibility. The use case of PUPPET is a model developer who builds a detector alongside its model and tunes the model toward it. Training several models toward the same detector and telling them apart is therefore outside the main use case. As an exploratory analysis, we ran attribution between Llama-3 and Qwen3, both trained with PUPPET toward the same detector, and observed that AUROC varies across tasks: 55.3 (ELI5), 81.6 (Multi-News), and 87.2 (IELTS). We found that this variation depends on how far apart the detector score distributions of the two models are. In practice, we therefore consider attribution feasible when the two models are trained toward different detectors. We will add this analysis to App. H.


References

[1] Yuan et al. "BARTScore: Evaluating Generated Text as Text Generation." NeurIPS 2021.

Thanks to the authors for their response. Some of my concerns are addressed, including the comparison with the end-to-end baseline, the robustness under unseen paraphrasing edits, and the advantages over standard watermarking paradigms.

However, key limitations remain. The proposed method still lacks statistical detection guarantees comparable to watermarking, and the LLM output quality evaluation remains in-domain (i.e., training and evaluating on the same task types). Therefore, I maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.