acceptodds
Under review as a conference paper at ICLR 2027

Reformulating Data-Adaptive LLM-Generated Text Detection: A Class-Conditional Gaussian Approximation

Abstract

Data-adaptive detection of LLM-generated text aims to learn from human and machine passages a scoring rule for a specified deployment criterion. We formulate the detection as a hypothesis-testing problem that maximizes machine-text detection power while controlling the false-positive rate on human writing. Existing logit-based detectors construct evidence from token log probability without considering the context, yet each token with different prefixes actually contributing differently to the detector. We therefore introduce an entropy-conditioned detection method that jointly models token log probability and contextual entropy through a two-dimensional tensor-spline witness and aggregates the resulting evidence into a passage-level score. The central challenge is learning this witness for the fixed-FPR power objective, whose direct formulation leads to an ill-posed optimization problem. We address this challenge by analyzing how token-level evidence accumulates across a passage and applying this sequential structure to approximate the human and machine scores as Gaussian distributions. This reformulation converts the original objective into a tractable problem with a regularized closed-form solution and shows that fixed-FPR power, fixed-FNR human TNR, and AUC share the same optimization direction. Across four source models and five domains, our method achieves relative AUROC-error reductions of 7.4% and 10.9% over the previous state-of-the-art method in white-box and strict proxy-model black-box evaluations, respectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM-generated text detection, data-adaptive detection, optimization, Tensor-spline witness function, Gaussian approximation, AUC, closed-form solution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WbCXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an entropy-conditioned data-adaptive method for detecting LLM-generated text. The authors argue that existing logit-based detectors overlook contextual uncertainty, as tokens with the same log probability may carry different discriminative evidence under different entropy levels. To address this, they introduce a two-dimensional tensor-spline witness that jointly models token log probability and contextual entropy. The paper also identifies a theoretical issue in the optimization objective of AdaDetectGPT and reformulates detection as maximizing detection power under a fixed human FPR. By applying a martingale decomposition, the authors derive a class-conditional Gaussian approximation for passage-level features. Under a shared-covariance assumption, they show that fixed-FPR detection power, fixed-FNR human TNR, and AUC share the same optimal projection direction, leading to a regularized closed-form solution. An independent human calibration set is further used to provide finite-sample FPR control. Experiments are conducted across five datasets and four source models in both white-box and black-box settings. The proposed method outperforms existing baselines, achieving AUROC scores of 0.8374 and 0.7701 in white-box and black-box settings, respectively. Ablation studies suggest that entropy conditioning contributes most of the performance improvements.

Critical strengths
  1. Well-motivated problem formulation. The paper identifies limitations in existing logit-based detectors and highlights a theoretical issue in AdaDetectGPT's optimization objective, providing motivation for the proposed reformulation.
Critical weaknesses
  1. The theoretical results rely on strong assumptions, particularly Gaussianity and shared covariance. However, the experiments only examine the Gaussianity of the final detection scores, without validating the shared-covariance assumption. This raises concerns about whether the theoretical guarantees hold in practice.
  2. The experiments do not clearly demonstrate the benefits of the proposed theoretical reformulation. Ablation results suggest that most improvements come from entropy conditioning rather than other components. Comparisons with alternative optimization methods using the same features would help clarify the contribution of the theory.
Questions for authors
  1. Fast-DetectGPT suffers from length sensitivity due to its normalization scheme. This paper removes the passage-specific variance normalization but instead uses an unnormalized sum of token-level evidence, whose expected value may scale linearly with passage length. As a result, longer passages may systematically receive higher or lower detection scores than shorter ones, even when their average token-level evidence is similar, potentially leading to length-dependent false-positive rates under a shared threshold. Since the experiments use length-matched passages of 150–200 tokens, could the authors evaluate performance on mixed-length texts and report length-stratified FPR and detection power under a shared calibration threshold?
  2. The paper argues that tokens with the same log probability but different contextual entropy may provide different discriminative evidence for human-vs-machine detection. However, this claim is not directly validated by the current experiments. Could the authors provide an analysis showing whether contextual entropy affects discriminative evidence when controlling for token log probability, and visualize how the learned witness varies across entropy levels?
  3. all evaluated source models were released in 2022–2023, with no experiments on more recent LLMs (close-source LLMs such as GPT6). This raises concerns about whether the proposed method generalizes to modern LLM-generated text, particularly given the rapid improvements in generation quality.
  4. The baseline evaluation mainly focuses on earlier logit-based methods and overlooks recent detectors such as SurpMark [1] and Uncertainty++ [2]. SurpMark is relevant as both methods leverage labeled human and machine reference text to construct detection statistics, making it a meaningful comparison. Uncertainty++ is particularly relevant to the paper's motivation of exploiting uncertainty information beyond token log probabilities. Comparisons with these methods under matched experimental settings would better establish the competitiveness and novelty of the proposed approach.
  5. The paper lacks evaluation on established public benchmarks for LLM-generated text detection, such as DetectRL (NeurIPS2024) and RAID (ACL2024). These benchmarks cover more diverse generators, text lengths, decoding strategies, and adversarial settings. Could the authors evaluate their method on these benchmarks to better demonstrate its generalization and robustness in realistic scenarios?

[1] Black-Box Detection of LLM-Generated Text Using Generalized Jensen-Shannon Divergence ICML26 [2] On the Salience of Low-Probability Tokens for AI-Generated Text Detection: A Multiscale Uncertainty Perspective ICML26

4 comments

Thank you for your careful reading and your advices. Here we restate our contributions, test the covariance assumption and broaden the experimental coverage.

Weakness 1: The covariance assumption and power

Thank you for raising the shared-covariance assumption. Our previous claim of equal covariance is actually unnecessary for the AUROC evaluation. We fit it using the regularized sum of two separately estimated class covariances. Under the Gaussian model, AUROC is the probability that an independent machine score exceeds a human score. Their difference has mean θ⊤(μM−μH)\theta^\top(\mu_M-\mu_H) and variance θ⊤(ΣH+ΣM)θ\theta^\top(\Sigma_H+\Sigma_M)\theta. AUROC increases with the ratio of this mean to its standard deviation. Cauchy–Schwarz gives the maximizing direction θ0∝(ΣH+ΣM)−1(μM−μH)\theta_0\propto(\Sigma_H+\Sigma_M)^{-1}(\mu_M-\mu_H) even with unequal covariances.

At fixed FPR α\alpha, human variance sets the threshold’s offset from the human mean, and machine variance affects the fraction of machine scores above it. So, still, the AUROC direction can lose TPR when the variance split changes with direction. We bound this loss here:

0≤sup⁡θπαG(θ)−παG(θ0)≤O(Δτ2).0\le \sup_\theta \pi_\alpha^G(\theta)-\pi_\alpha^G(\theta_0)\le O(\Delta_\tau^2).

Here παG\pi_\alpha^G is Gaussian TPR at FPR α\alpha, and Δτ\Delta_\tau is the range of the human share of total projected variance across directions. The bound holds when those shares stay away from 0 and 1 and the standardized mean gap along θ0\theta_0 is bounded above and away from zero. The loss will be zero as long as the covariances of the scores of human and machine are proportional.

To make things clearer we also do experiments on whether this fitting error will influence our test TPR. Here our setting fits one direction for both targets(human and machine) while Gaussian Neyman–Pearson (NP) uses both class covariances to fit separate linear directions on the training data by optimizing Gaussian power at 1% and 5% FPR. Both use the same cached passage features and data splits in 20 white-box and black-box settings. For each target, we compare Ours and the corresponding NP fit at the same FPR on the test ROC.

We measure covariance mismatch on held-out features. The residual in the table is the relative Frobenius distance after rescaling one covariance to best match the other (zero if proportional); it is distinct from Δτ\Delta_\tau. The human variance share is the human score variance divided by the sum of the two class score variances along our fitted direction. Each row averages the settings in that regime.

RegimeCovariance residualHuman variance shareOurs: TPR at 1% / 5% FPRSeparate-covariance NP: TPR at 1% / 5% FPR
White box14.4%55.1%21.13% / 42.97%19.78% / 42.91%
Black box14.2%50.0%11.33% / 29.79%11.19% / 27.52%

The covariance residual is about 14% in both regimes, but it does not directly measure a loss in power. At matched FPR, Ours leads the NP fits by 0.06–2.27 percentage points across the four comparisons. This means the shared-covariance fitting direction is reasonable.

Weakness 2: What the reformulation and the score each contribute

We agree that the previous experiments cannot show where the gains come from. Here we restate our contribution that the score design accounts for much of the gain. Previously we have done ablation over entropy and here we provide the ablation of document-level covariance modeling, or in other words, removing the passage-specific denominator. The motivation of this change is from the failed attempt of AdaDetectGPT for its theoretical claim. The previous difficulty is to deal with the denominator to form a Gaussian style score. Here we remove it and prove that the new score directly fall in the Gaussian distribution. The means and variances of this new score can therefore determine its Gaussian power. A regularized linear equation solve fits the Gaussian AUROC surrogate and supplies one direction for all FPR targets.

Let SθS_\theta denote the score that divides our linear score by the passage-specific reference standard deviation. We keep the fitted direction and the length-matched passages fixed and calibrate the two scores separately on the same human set. The table averages 20 white-box and black-box settings.

RegimeScoreAUROCFPR / TPR, nominal 1%FPR / TPR, nominal 5%
White boxOurs0.83740.84% / 17.72%4.40% / 40.02%
White boxSθS_\theta0.83400.68% / 14.06%4.21% / 37.63%
Black boxOurs0.77541.03% / 10.35%4.59% / 27.87%
Black boxSθS_\theta0.77210.86% / 8.42%4.34% / 25.35%

Removing the denominator adds about 0.003 to AUROC in both regimes. At nominal 5% FPR, it raises calibrated TPR by 2.39 percentage points in white box and 2.52 in black box. The TPR gains persist when we compare at exactly 5% FPR on the test ROC: from 41.91% to 43.06% in white box and from 28.91% to 29.61% in black box.

We agree that the optimal projection direction itself is not our main contribution. Our contribution is to show how Gaussian AUROC provides a tractable fitting objective for a single score used at multiple FPR targets. To test the practical value of this reformulation, we compared our covariance-sum fit with L2L_2-regularized logistic regression and Gaussian Neyman–Pearson (NP) fits using the same features and data splits.

Each entry reports white-box / black-box results, averaged over 20 settings. TPR is measured at matched FPR on the test ROC. Time is the median for model selection and refitting on cached features.

Fitting methodAUROCTPR at 1% FPRTPR at 5% FPRFit time (s)
Ours0.8373 / 0.775321.13% / 11.33%42.97% / 29.79%0.35 / 0.64
L2L_2 logistic regression0.8371 / 0.774420.91% / 11.23%43.18% / 30.16%3.36 / 8.68
Gaussian NP, 1% target0.8226 / 0.765719.78% / 11.19%40.91% / 27.60%1.49 / 1.52
Gaussian NP, 5% target0.8302 / 0.760119.97% / 9.98%42.91% / 27.52%1.43 / 1.49

Ours and logistic regression yield nearly identical AUROC and similar TPR at both FPR levels. The clearest difference is fitting time: Ours is 9.6× faster in white box and 13.6× faster in black box than logistic regression. The comparison will Gaussian NP leads to similar conclusion. This is the practical benefit of the reformulation: the Gaussian AUROC objective gives an optimal linear direction in closed form, and a single linear solve per fit estimates that direction with regularization while retaining the detection performance.

Question 1: Mixed lengths under a shared threshold

Thank you for raising length sensitivity as a problem. We scored the prompt and fresh continuation and used one human calibration threshold across the three lengths for each method, setting, and nominal FPR. The table reports realized rates across 20 white-box settings. Here SθS_\theta uses the same fitted direction as Ours, with the passage-specific denominator.

Continuation tokensMethodAUROCFPR / TPR, nominal 1%FPR / TPR, nominal 5%
AllOurs0.89340.84% / 33.12%4.76% / 57.20%
SθS_\theta0.89320.70% / 30.06%4.66% / 56.07%
Ada0.88590.93% / 29.96%4.84% / 55.54%
SurpMark0.82810.72% / 14.63%4.57% / 37.77%
Uncertainty0.77320.74% / 5.90%4.50% / 25.07%
50–99Ours0.82520.30% / 8.68%3.29% / 30.32%
SθS_\theta0.82180.57% / 10.46%4.46% / 33.94%
Ada0.81430.67% / 11.77%4.97% / 33.13%
SurpMark0.76420.39% / 3.22%2.61% / 18.11%
Uncertainty0.70200.95% / 4.05%5.43% / 19.48%
100–149Ours0.91440.98% / 31.86%4.47% / 60.55%
SθS_\theta0.91230.91% / 28.95%4.34% / 58.12%
Ada0.90291.24% / 28.88%4.56% / 57.77%
SurpMark0.83920.45% / 12.35%4.00% / 38.17%
Uncertainty0.78100.79% / 5.51%4.34% / 24.48%
150–200Ours0.95201.23% / 58.99%6.54% / 80.88%
SθS_\theta0.95080.61% / 50.90%5.18% / 76.27%
Ada0.94680.88% / 49.37%4.99% / 75.84%
SurpMark0.88021.32% / 28.42%7.12% / 57.17%
Uncertainty0.84000.49% / 8.14%3.71% / 31.29%

Across all lengths, Ours has the highest AUROC (0.8934) and the highest TPR at both nominal FPR levels (33.12% at 1% and 57.20% at 5%).

Ours has the highest AUROC in each of the three length bands. At 50–99 tokens, ours has lower TPR than SθS_\theta at both nominal FPR levels, but also lower realized FPR. At 100–149 and 150–200 tokens, Ours has the highest TPR at both nominal FPR levels. Therefore, totally our score has the best performance over mixed-length detection.

Question 2: Entropy at comparable token log probabilities

Thank you for suggesting this comparison. Figure 2 stratifies by log probability, relative position, and passage length. The top panels show held-out log odds ratios for machine authorship (P55–P75 vs. P25–P45 entropy); the bottom panels show witness slopes at P35 and P65. Below are the odds ratios and slopes for [−2,−1.5)[-2,-1.5), with pointwise paired-bootstrap intervals.

White-box settingHigh vs. low entropy odds ratio [95% CI]Witness slope at P35Witness slope at P65
OPT-2.7B / WritingPrompts1.076 [1.024, 1.135]0.0222-0.0152
OPT-6.7B / WritingPrompts1.077 [1.019, 1.134]0.01550.0263
OPT-2.7B / C41.164 [1.088, 1.247]0.00370.0309

Within this log-probability window, the higher-entropy group has 7.6%–16.4% higher odds of machine authorship, and all three pointwise intervals exclude 1. For OPT-2.7B on WritingPrompts, the fitted witness slope switches from positive at P35 to negative at P65. Some lower-probability WritingPrompts windows also show the reverse odds pattern. The effect of entropy thus depends on where we look in the log-probability range. We will also visualize the difference in our essay.

Question 3: More recent generators

Thank you for recommending more recent generators. We evaluated Llama-3.1-8B and Qwen3-8B in white box. Each row averages the results across all five datasets.

AUROC

GeneratorOursFastAdaSurpMarkUncertaintyUncertainty++
Llama-3.1-8B0.80230.72300.78890.73260.68300.7676
Qwen3-8B0.94110.89660.90510.86220.74520.7896

Ours has the highest AUROC on both generators, exceeding Ada, the strongest baseline, by 0.0134 on Llama-3.1-8B and 0.0360 on Qwen3-8B.

TPR at 5% FPR (test ROC)

GeneratorOursFastAdaSurpMarkUncertaintyUncertainty++
Llama-3.1-8B36.64%21.40%32.72%25.08%14.92%28.08%
Qwen3-8B71.08%64.76%65.92%41.80%25.52%37.68%

Ours also has the highest TPR at 5% FPR: 3.92 percentage points above Ada on Llama-3.1-8B and 5.16 points above Ada on Qwen3-8B.

Question 4: SurpMark and Uncertainty++

Thank you for suggesting these baselines. Results average 20 settings in each regime.

AUROC

RegimeOursSurpMarkUncertaintyUncertainty++
White box0.83740.76980.70480.7992
Black box0.77010.69140.63320.6953

Uncertainty++ is the strongest baseline in both regimes. Ours leads it by 0.0382 AUROC in white box and 0.0748 in black box.

TPR at 5% FPR (test ROC)

RegimeOursSurpMarkUncertaintyUncertainty++
White box43.06%25.68%18.05%32.58%
Black box28.80%18.24%10.74%19.89%

At 5% FPR, Ours improves on Uncertainty++ by 10.48 percentage points in white box and 8.91 points in black box. The gains in overall ranking therefore also appear at this operating point.

Question 5: DetectRL and RAID

Thank you for recommending public benchmarks. We fit on the benchmark training splits and evaluate with fixed proxy models in black box, using 1,000 pairs per condition.

AUROC

Black-box subsetOursFastAdaSurpMarkUncertaintyUncertainty++
DetectRL main (4 generators)0.92430.87310.86890.92810.93100.8901
RAID subset (16 conditions)0.97500.90230.97370.91870.91880.9512
Equal-weight mean0.94960.88770.92130.92340.92490.9207

Uncertainty leads on DetectRL, where our mean AUROC is 0.0067 below the best result. On RAID conditions, our method has the highest mean AUROC (0.9750). Our method achieves the highest mean AUROC across the two benchmarks (0.9496) when each benchmark is weighted equally.

Reviewer_nd7ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies data-adaptive detection of LLM-generated text, with a goal of achieving an increase of the machine-text detection power with controlling false-positive rate on human writtens. It models token log probability and contextual entropy with a tensor-product spline, aggregates reference-centered token features into document features, and derives a discriminant direction under class-conditional Gaussian and shared-covariance assumptions. Independent human passages are provided for conformal calibration.

Critical strengths
  1. Intuitive motivation. The tokens with the same probability may be given different detection evidence under different contextual uncertainty levels. Entropy conditioning is a natural extension, and the ablation result verifies its rationality.
  2. The analysis of the Jensen inequality error in AdaDetectGPT is concrete and convincing, helping clarify the limitations of its theoretical justification.
  3. Separating the score fitting and threshold calibration is convenient. Reporting the low-FPR performance and realized operating points after calibration also makes evaluation protocol quite transparent.
Critical weaknesses
  1. It is unclear how the methodological and theoretical contributions over AdaDetectGPT should be framed. AdaDetectGPT already uses a learned spline witness, a martingale central limit theorem, and linear-system fitting. The optimal direction we obtain here is equivalent to classical Fisher/LDA and false-positive control has a well-known conformal approach. In this sense the primary contributions seem to be the entropy conditioning, the document-level covariance modeling, and a correction to the previous derivation; the separate merits of these require more direct validation.

  2. The experiments hardly prove that the theoretical reformulation is necessary. Ablations seems to give most of the gain to entropy conditioning. However, there is no comparisons with logistic regression, the heteroscedastic discriminant analysis or low-FPR-oriented fitting method on the same features, training data and calibration. It's thus difficult to say where the gain mostly come from, either from introduced entropy information, or from the proposed fitting criterion.

  3. The experimental coverage is still quite limited. Despite that the evaluation encompasses four source models, five datasets, and cross-dataset testing, the evaluation with a more restricted generation protocol. The systematic evaluation on unseen generator, varying decoding strategy and text length, and paraphrased text are missing. Proxy-model black-box evaluation by itself does not demonstrate cross-generator generalization. In addition, each configuration employs 500 calibration passages and 500 human test passages, so estimates near 1% FPR are very sensitive to sampling variation. The amount of entropy-conditioning gain in calibrated power at this operating point is also not statistically significant.

Questions for authors
  1. How do alternative discriminative fits perform using exactly the same two-dimensional features and calibration protocol?
  2. Can the authors quantify class-covariance mismatch and evaluate its effect on performance?
  3. Can the evaluation include held-out generators, different decoding strategies, varying lengths, and paraphrased text?
3 comments

Thank you for your careful reading and your advices. Here we restate our contributions, test the covariance assumption and broaden the experimental coverage.

Weakness 1: Contributions relative to AdaDetectGPT

We agree that the previous experiments cannot show where the gains come from. Here we restate our contribution that the score design accounts for much of the gain. Previously we have done ablation over entropy and here we provide the ablation of document-level covariance modeling, or in other words, removing the passage-specific denominator. The motivation of this change is from the failed attempt of AdaDetectGPT for its theoretical claim. The previous difficulty is to deal with the denominator to form a Gaussian style score. Here we remove it and prove that the new score directly fall in the Gaussian distribution. The means and variances of this new score can therefore determine its Gaussian power. A regularized linear equation solve fits the Gaussian AUROC surrogate and supplies one direction for all FPR targets.

Let SθS_\theta denote the score that divides our linear score by the passage-specific reference standard deviation. We keep the fitted direction and the length-matched passages fixed and calibrate the two scores separately on the same human set. The table averages 20 white-box and black-box settings.

RegimeScoreAUROCFPR / TPR, nominal 1%FPR / TPR, nominal 5%
White boxOurs0.83740.84% / 17.72%4.40% / 40.02%
White boxSθS_\theta0.83400.68% / 14.06%4.21% / 37.63%
Black boxOurs0.77541.03% / 10.35%4.59% / 27.87%
Black boxSθS_\theta0.77210.86% / 8.42%4.34% / 25.35%

Removing the denominator adds about 0.003 to AUROC in both regimes. At nominal 5% FPR, it raises calibrated TPR by 2.39 percentage points in white box and 2.52 in black box. The TPR gains persist when we compare at exactly 5% FPR on the test ROC: from 41.91% to 43.06% in white box and from 28.91% to 29.61% in black box.

Weakness 2 and Question 1: What the fitting rule contributes

We agree that the optimal projection direction itself is not our main contribution. As we have mentioned above, we will restate our contributions that the gains mainly come from the new score design. This, especially the removal of the passage-level denominator, is our reformulation. Our main objective of providing the fitting criterion is to show how the Gaussian score provides a tractable fitting objective for a single score used at multiple FPR targets, and we can therefore solve this optimization problem correctly with theoretical support. Still we compared our covariance-sum fit with L2L_2-regularized logistic regression and Gaussian Neyman–Pearson (NP) fits using the same features and data splits.

Each entry reports white-box / black-box results, averaged over 20 settings. TPR is measured at matched FPR on the test ROC. Time is the median for model selection and refitting on cached features.

Fitting methodAUROCTPR at 1% FPRTPR at 5% FPRFit time (s)
Ours0.8373 / 0.775321.13% / 11.33%42.97% / 29.79%0.35 / 0.64
L2L_2 logistic regression0.8371 / 0.774420.91% / 11.23%43.18% / 30.16%3.36 / 8.68
Gaussian NP, 1% target0.8226 / 0.765719.78% / 11.19%40.91% / 27.60%1.49 / 1.52
Gaussian NP, 5% target0.8302 / 0.760119.97% / 9.98%42.91% / 27.52%1.43 / 1.49

Ours and logistic regression yield nearly identical AUROC and similar TPR at both FPR levels. The clearest difference is fitting time: Ours is 9.6× faster in white box and 13.6× faster in black box than logistic regression. The comparison will Gaussian NP leads to similar conclusion. This is the practical benefit of the reformulation: the Gaussian AUROC objective gives an optimal linear direction in closed form, and a single linear solve per fit estimates that direction with regularization while retaining the detection performance.

Question 2: Unequal class covariances

Thank you for raising the shared-covariance assumption. We fit using the regularized sum of two separately estimated class covariances. Under the Gaussian model, AUROC is the probability that an independent machine score exceeds a human score. Their difference has mean θ⊤(μM−μH)\theta^\top(\mu_M-\mu_H) and variance θ⊤(ΣH+ΣM)θ\theta^\top(\Sigma_H+\Sigma_M)\theta. AUROC increases with the ratio of this mean to its standard deviation. Cauchy–Schwarz gives the maximizing direction θ0∝(ΣH+ΣM)−1(μM−μH)\theta_0\propto(\Sigma_H+\Sigma_M)^{-1}(\mu_M-\mu_H) even with unequal covariances.

At fixed FPR α\alpha, human variance sets the threshold’s offset from the human mean, and machine variance affects the fraction of machine scores above it. So the AUROC direction can lose TPR when the variance split changes with direction. We bound this loss here:

0≤sup⁡θπαG(θ)−παG(θ0)≤O(Δτ2).0\le \sup_\theta \pi_\alpha^G(\theta)-\pi_\alpha^G(\theta_0)\le O(\Delta_\tau^2).

Here παG\pi_\alpha^G is Gaussian TPR at FPR α\alpha, and Δτ\Delta_\tau is the range of the human share of total projected variance across directions. The bound holds when those shares stay away from 0 and 1 and the standardized mean gap along θ0\theta_0 is bounded above and away from zero. The loss will be zero as long as the covariances of the scores of human and machine are proportional.

To make things clearer we also do experiments on whether this fitting error will influence our test TPR. Here our setting fits one direction for both targets(human and machine) while Gaussian Neyman–Pearson (NP) uses both class covariances to fit separate linear directions on the training data by optimizing Gaussian power at 1% and 5% FPR. Both use the same cached passage features and data splits in 20 white-box and black-box settings. For each target, we compare Ours and the corresponding NP fit at the same FPR on the test ROC.

We measure covariance mismatch on held-out features. The residual in the table is the relative Frobenius distance after rescaling one covariance to best match the other (zero if proportional); it is distinct from Δτ\Delta_\tau. The human variance share is the human score variance divided by the sum of the two class score variances along our fitted direction. Each row averages the settings in that regime.

RegimeCovariance residualHuman variance shareOurs: TPR at 1% / 5% FPRSeparate-covariance NP: TPR at 1% / 5% FPR
White box14.4%55.1%21.13% / 42.97%19.78% / 42.91%
Black box14.2%50.0%11.33% / 29.79%11.19% / 27.52%

The covariance residual is about 14% in both regimes, but it does not directly measure a loss in power. At matched FPR, Ours leads the NP fits by 0.06–2.27 percentage points across the four comparisons. This means the shared-covariance fitting direction is reasonable.

Weakness 3 and Question 3: Held-out generation, decoding, and paraphrasing

Thank you for your advice that our evaluation coverage is still limited. Methods score the same paired passages within each condition. The results are macro averages over the five datasets. Fast and Ada denote Fast-DetectGPT and AdaDetectGPT.

Held-out generators. With fixed black-box proxies, we fitted our method and two strongest baselines separately on OPT-2.7B, Falcon-7B, and LLaMA-2-7B, then tested each fit on all three generators and OPT-6.7B across five datasets. A cell is one fit generator, test generator, and dataset. Fast requires no generator-specific fit. We kept the fitted scores fixed across test generators. Entries are AUROC / empirical TPR at 5% test-ROC FPR.

Fitting generatorTest splitCellsOursFastAda
AllSame generator150.7073 / 18.48%0.6356 / 13.75%0.6855 / 16.96%
AllHeld out450.7079 / 18.22%0.6429 / 14.71%0.6886 / 16.99%
OPT-2.7BHeld out150.7081 / 19.94%0.6350 / 14.00%0.6838 / 16.75%
Falcon-7BHeld out150.7117 / 18.08%0.6455 / 14.87%0.6910 / 16.91%
LLaMA-2-7BHeld out150.7039 / 16.66%0.6481 / 15.27%0.6910 / 17.31%

Across 45 held-out cells, Ours leads in AUROC and matched-FPR TPR. It has the highest mean held-out AUROC for each fitting generator. Comparing fits on the same three test generators, Ours loses 0.0010 AUROC and 0.21 percentage points of matched-FPR TPR when the fitting generator changes.

Decoding. We generated OPT-2.7B continuations at top-k=50k=50 and temperatures 1.2 and 1.5. We report both matched-FPR TPR on the test ROC and the realized rates after human calibration.

TemperatureMethodAUROCTPR at 5% FPRFPR / TPR, nominal 5%
1.2Ours0.774230.60%4.32% / 27.96%
Fast0.47794.68%4.76% / 4.04%
Ada0.687620.16%5.56% / 18.72%
1.5Ours0.723025.32%4.32% / 22.84%
Fast0.18550.24%4.76% / 0.20%
Ada0.532310.28%5.56% / 9.68%

Paraphrasing. We score the clean OPT-2.7B passages and their Qwen3-8B full-passage paraphrases with the same proxies across five datasets. Entries are AUROC / empirical TPR at 5% FPR.

Test textOursFastAda
Clean OPT-2.7B controls0.8097 / 37.41%0.7180 / 22.21%0.7685 / 28.09%
Qwen3-8B paraphrases0.7221 / 24.17%0.6711 / 17.73%0.7005 / 19.37%

Paraphrasing lowers TPR for all three methods. Still, our AUROC and TPR are the highest of the three on the paraphrases.

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

The authors propose a LLM-generated text detection method that takes entropy as input in a learnable scorer.

Critical strengths

None

Critical weaknesses

The entirity of theoretical arguments are insubstantial in providing insight into the method. Close to no important ablations are performed. Methodology is not clearly discussed. No related works section.

Questions for authors

Overall assessment

In its current form, I do not believe the paper is ready for acceptance. My main reservations concern the theoretical development in Section 4 (M1–M4), which I find difficult to connect to the method as deployed. I also have concerns about how the method is presented and about missing controls for model capacity and cost (M5–M7). Several presentation issues (m1–m5) compound these. Addressing the major concerns would, in my view, require substantial rewriting of Sections 3–4 and additional experiments, likely beyond what the discussion period allows. I would therefore encourage the authors to consider a revised resubmission. That said, I will read the response carefully, and if the authors can fully address the major concerns, I remain open to revising my score.

Concerns are split into major (M) and minor (m).

Major concerns

M1. The centering steps in Section 4.2 appear to reduce to a CLT on UXU_X itself.

Section 4.2 introduces five successive centerings:

  • the reference-centered features ete_t (Eq. 8);
  • the drift/innovation split bc,t,ηc,tb_{c,t}, \eta_{c,t} (Eq. 9);
  • the mean/fluctuation split mc,T,Rc,Tm_{c,T}, R_{c,T} (Eq. 11);
  • the Doob increments ξc,t,T\xi_{c,t,T} (Eq. 12);
  • their sum ζc,t,T\zeta_{c,t,T} (Eq. 14).

Unless I am missing something, the combination is the Doob martingale of UXU_X, which exists for any integrable UXU_X without the intermediate decomposition. Apart from the exact-reference remark in Appendix A.2, Bc,TB_{c,T}, Mc,TM_{c,T}, and Rc,TR_{c,T} are not used downstream. The estimator (Appendix A.7) uses only sample means and covariances of UXU_X.

M2. The asymptotic regime is not stated.

Convergence statements (⇒\Rightarrow, →p\xrightarrow{p}) appear in Eqs. (5), (6), (15), (17), and (38)–(40), but the paper never states what tends to infinity. TT is the passage length; The triangular-array indexing (ξc,t,T\xi_{c,t,T}) suggests T→∞T \to \infty, but only implicitly. Dangerously, Lemma 2 presents the limit and a finite-TT approximation as equivalent ("Ac,T−1/2Zc,T⇒N(0,I)A_{c,T}^{-1/2} Z_{c,T} \Rightarrow \mathcal{N}(0, I), or equivalently Zc,T ∼˙ N(0,Ac,T)Z_{c,T} \,\dot\sim\, \mathcal{N}(0, A_{c,T})").

M3. Theorems 2–3 rest on exact Gaussianity and equal class covariance, under which the problem reduces to Fisher's discriminant.

Theorem 1 is an approximation (∼˙\dot\sim), yet Eqs. (19)–(22) are derived as equalities. The optimality statements therefore hold exactly only if Gaussianity holds exactly. The paper's own tests reject normality for 9/20 human-score distributions at the 5% level after Benjamini–Hochberg adjustment, versus 0/20 for machine scores (Table 7). This asymmetry indicates that the two class score distributions are not location shifts of one another.

The only stated rationale for equal covariance is convenience: it is assumed "for closed-form fitting" (Section 4.3). I could not find any comparison of ΣH\Sigma_H and ΣM\Sigma_M, or of class-wise variances of the projected score.

The assumption also sits in tension with the paper's own motivation: Section 1 argues that the reference-law variance may not match "the actual human or machine fluctuation scale", while Section 3.1 states that the method instead learns the "class-conditional fluctuation scale".

Worse, under these assumptions the result is classical, if not trivial. The paper identifies θ⋆\theta^\star as "the classical shared-covariance discriminant direction (Fisher, 1936)". The proof (Appendix A.6) is a single whitening and Cauchy–Schwarz step. The estimator follows regularized discriminant analysis (Friedman, 1989; Hastie et al., 1995).

Theorem 3's results contrast with the abstract's description of the fixed-FPR objective as "the central challenge", and with Section 2's characterization of the problem as "nontrivial" and "ill-posed".

M4. Section 4 is predominantly theoretical justification rather than methodology.

Section 4 runs to roughly three pages:

  • The score definition (Section 4.1) occupies under one page.
  • Sections 4.2–4.3 consist of two lemmas, three theorems, and their derivations.
  • The practical content (the finite-sample fitting and calibration paragraphs at the end of Section 4.3) runs to roughly 20 lines.

Section 4 does not state how μt=Ev∼qtϕt(v)\mu_t = \mathbb{E}_{v\sim q_t}\phi_t(v) is computed in the experiments. It points to the sampled-moment form in Appendix A.4, whereas Appendix B states that exact moments over the full vocabulary are used. Deployment details are absent from Section 4 and partly relegated to Appendix B.

This means important details such as discussions on whether the gains stem from additional parameters and reports on training and inference costs are not included in the main text.

Minor concerns

m1. No related-work section. Neither the main text nor Appendices A–D contains one. No concurrent work is discussed. A related-work section could be placed in the appendix.

m2. The term "witness" is not defined. The term appears throughout the abstract, the introduction, and Figure 1 before Eq. (3). Even in Eq. (3), it is introduced only as "a learned smooth one-dimensional witness ww". Its role (a function applied to token log probability before conditional centering) is left to be inferred. The term is later applied both to the function gθg_\theta (Section 4.1) and to the coefficient vector θ\theta ("learn witness direction", Algorithm 1).

m3. Section 3.1 assumes familiarity with AdaDetectGPT. rr, RR, hβh_\beta, and the TNR lower bound are used without definition in the main text. Several steps refer to quantities that never appear: "A machine-side martingale normal approximation provides the corresponding threshold" is given without a formula; "Further turned into maximizing rr" and "algebraically solving that objective" refer to an objective that is never written down.

m4. The separation between qq and rr is not motivated. Section 2 describes only the two roles vaguely. Both require full next-token distributions. Appendix B computes moments "over the shared GPT-J vocabulary", so the pair must also share a vocabulary.

Without a clear justification, the new entropy feature utu_t is computed from qq, whereas ztz_t comes from rr (Section 4.1). Only one black-box pair is used (GPT-J-6B as qq, GPT-Neo-2.7B as rr), with no rationale for the assignment.

4 comments

Thank you for the careful review. We will simplify the derivation and clarify which assumptions are needed for fitting and for calibration.

Response to Major Problem 1 (M1)

We agree that the proof can start directly from UXU_X. For class cc, let Ft\mathcal{F}_t contain the first tt tokens, with F0\mathcal{F}_0 trivial. The Doob increments are

Dc,t,T=Ec[UX∣Ft]−Ec[UX∣Ft−1].D_{c,t,T}=\mathbb{E}_c[U_X\mid\mathcal{F}_t]-\mathbb{E}_c[U_X\mid\mathcal{F}_{t-1}].

They have conditional mean zero and sum to UX−EcUXU_X-\mathbb{E}_c U_X. We will use this shorter proof in the main text. Fitting and scoring use only UXU_X. We will keep Bc,TB_{c,T} and Mc,TM_{c,T} in the appendix to explain reference-model mismatch.

Response to Major Problem 2 (M2)

The limit is T→∞T\to\infty, where TT is the number of scored tokens and the feature map is fixed. We will state the conditional covariance stabilization and Lindeberg conditions and distinguish the limit from the approximation at finite lengths.

For a fixed direction θ\theta with positive score variance, define

\qquad V_T=\sum_{t=1}^T\mathbb{E}_c[Y_t^2\mid\mathcal{F}_{t-1}].$$ With finite fourth moments, Haeusler's bound on the Kolmogorov distance to a standard normal is $$C\left(\sum_{t=1}^T\mathbb{E}_c[Y_t^4]+\mathbb{E}_c[(V_T-1)^2]\right)^{1/5}.$$ We will include the bound and explain how large increments and variation in $V_T$ affect the error. [Haeusler's bound](https://arxiv.org/pdf/1103.5050) ## Response to Major Problem 3 (M3) **Human scores and calibration.** We agree that human score tails can depart from Gaussianity. We ran supplementary experiments to assess how well the Gaussian approximation predicts detection power on held-out data. In repeated two-fold splits with a fixed score, we compared shared-variance Gaussian power predictions with held-out detection rates at the same threshold. Mean absolute errors ranged from 1.95 to 3.32 percentage points across the white-box and black-box results at nominal 1% and 5% FPR. This supports using Gaussianity as a working approximation at the tested lengths. At nominal 5% in white box, mean realized FPR was 4.37% with a Gaussian threshold and 5.14% with an empirical threshold. In deployment, ranks from independent human calibration control marginal finite-sample FPR under exchangeability. **Class covariance.** Under the Gaussian working model, an AUC-maximizing linear direction is $W^{-1}\delta$, where $W$ is the invertible sum of the two class covariances and $\delta$ is the machine-minus-human mean difference. Our fit already uses a regularized estimate of $W$. At a fixed FPR, power also depends on the ratio of the projected class variances. We will bound the potential power loss from using the covariance-sum direction and analyze it separately from the Gaussian approximation error. To test whether accounting for the observed covariance mismatch improves detection, we used the same features to fit a Gaussian Neyman–Pearson detector with separate class covariances, targeting 1% and 5% FPR. The covariance-sum fit achieved higher mean empirical power at both targets. At 5% FPR, it achieved 42.97% power in white box and 29.79% in black box, compared with 42.91% and 27.52% for the separate-covariance fit. **Why reformulate?** We agree that “ill-posed” was inaccurate. The Neyman–Pearson objective is well defined. The challenge is fitting a witness when both score distributions are unknown. We will change our statements for the contribution. AdaDetectGPT's reversed Jensen step invalidates its claimed population TNR bound as stated. Therefore we need to rebuild a correct and practical model. We remove the passage-specific denominator to make the score linear in the document feature. This gives a closed-form fit under the working model and recovers Fisher's direction under shared covariance. Furthermore we find entropy useful in this framework and form our final score. [AdaDetectGPT, Lemma S2](https://arxiv.org/abs/2510.01268) ## Response to Major Problem 4 (M4) We will detail the tensor-spline construction, the vocabulary-weighted reference mean, the direction fit, and independent human calibration. We will also report the basis sizes (16, 36, and 48 coefficients) and how we select regularization on training data. **m1–m3.** We will clarify the related work, define the witness as the map from log probability and contextual entropy to token evidence, and separate AdaDetectGPT's pointwise inequality from its population claim. **m4.** The reference model $q$ weights alternative tokens and supplies prefix entropy. The scoring model $r$ supplies their log probabilities. We use the generator for both roles in white box and GPT-J-6B for $q$ and GPT-Neo-2.7B for $r$ in black box. Sharing this proxy pair across methods keeps the comparison focused on their scores and fits.

M3: Reference variance versus class covariance.

Thank you for pointing out that our earlier wording blurred two different quantities. The reference-law variance in Eq. (2) is computed for a given passage by evaluating alternative next tokens under qtq_t at each observed prefix. It determines that passage's score divisor. The class covariance Σc=Cov⁡c(UX)\Sigma_c = \operatorname{Cov}_c(U_X), by contrast, measures how the observed feature vector varies across passages within class cc. The original shared-covariance assumption concerned only whether ΣH=ΣM\Sigma_H = \Sigma_M; it did not equate either covariance with the reference-law variance. The reference divisor can therefore be a poor measure of score variability even when the two classes have equal passage-level covariances.

We clarified this distinction in the revised manuscript:

  • Sections 1 and 3 explain that we retain reference centering, remove the passage-specific variance divisor, and estimate variability from observed passages in each class.
  • Sections 4.2 and 4.3 define and estimate ΣH\Sigma_H and ΣM\Sigma_M separately. The population direction based on their sum maximizes Gaussian AUROC over linear scores without assuming equal class covariances.
  • Section 4.4 explains where covariance differences matter: the AUROC direction can lose power at a fixed FPR when the human share of projected variance changes with direction. We give a bound on this loss.
Reviewer_whFYOpenReview ↗

Response to Rebuttals and Revisions

I thank the authors for their detailed response and substantial revision. Many of my specific points were addressed: the martingale argument is simplified, the limiting regime is stated, key terms are defined, costs are reported, and the experiments are considerably broader.

However, the revision does not address my central concern: the paper's stated contribution does not match what its evidence supports. The authors revised their paper by shrinking their original theoretical section, collapsing derivations and theorems, while new methodological subsections were added with considerable effort, all without raising my confidence in the paper's stated "reformulation of data-adaptive witness fitting" as its main theoretical contribution.

What's more, the rebuttal describes entropy as something the authors "find useful in this framework", but entropy conditioning is what carries the result, as it is the dominant factorial effect in both settings (Table 13). Without it, the method essentially ties AdaDetectGPT. The fitting rule, on the other hand, is interchangeable: in Table 16, L2 logistic regression matches the covariance-sum fit, while the Gaussian Neyman–Pearson fit, which targets the stated objective, does worse.

Therefore, I see the paper's authentic finding to be that conditioning the witness on prefix entropy improves detection. This is a worthwhile result, yet it is the least developed part of the paper. Meanwhile, the main text has grown with material that does not bear on this finding. I would advice the authors recenter the paper on this finding in a hypothetical resubmission.

Since there is no opportunity for a further author response, I am raising my score to 2 in recognition of the authors' responsiveness. I also acknowledge that I did not write "unmatching contribution" as my central concern in the original review, though it was implied with my major concerns on theory: without the theoretical support, the paper's story collapses.

Thank you for your reply and score raising. Previously we have restated our contribution that the score design accounts for much of the gain, which can be divided into two parts, entropy conditioning and removing the denominator. We have done ablation over entropy and here we provide the ablation of document-level covariance modeling, or in other words, removing the passage-specific denominator. The motivation of this change is from the failed attempt of AdaDetectGPT for its theoretical claim. The previous difficulty is to deal with the denominator to form a Gaussian style score. Here we remove it and prove that the new score directly fall in the Gaussian distribution. The means and variances of this new score can therefore determine its Gaussian power. A regularized linear equation solve fits the Gaussian AUROC surrogate and supplies one direction for all FPR targets.

Let SθS_\theta denote the score that divides our linear score by the passage-specific reference standard deviation. We keep the fitted direction and the length-matched passages fixed and calibrate the two scores separately on the same human set. The table averages 20 white-box and black-box settings.

RegimeScoreAUROCFPR / TPR, nominal 1%FPR / TPR, nominal 5%
White boxOurs0.83740.84% / 17.72%4.40% / 40.02%
White boxSθS_\theta0.83400.68% / 14.06%4.21% / 37.63%
Black boxOurs0.77541.03% / 10.35%4.59% / 27.87%
Black boxSθS_\theta0.77210.86% / 8.42%4.34% / 25.35%

Removing the denominator adds about 0.003 to AUROC in both regimes. At nominal 5% FPR, it raises calibrated TPR by 2.39 percentage points in white box and 2.52 in black box. The TPR gains persist when we compare at exactly 5% FPR on the test ROC: from 41.91% to 43.06% in white box and from 28.91% to 29.61% in black box.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair

Dear Area Chair,

We sincerely appreciate your time and effort in handling our submission. We would like to summarize how we have addressed the reviewers’ concerns through additional experiments, theoretical clarifications, and substantial manuscript revisions. Our revised paper now centers on entropy-conditioned witness design and removal of the passage-specific variance denominator, supported by a correction to AdaDetectGPT’s theoretical argument and an explicit analysis of the revised score.

Reviewer whFY — Alignment between the paper’s contributions and its presentation. The reviewer acknowledged the broader experiments and clarified derivations, but remained concerned that the paper emphasized a reformulation of witness fitting while entropy conditioning carried the empirical improvement. We have substantially reorganized the manuscript in response. We changed the title to Entropy-Conditioned Data-Adaptive Detection of LLM-Generated Text, rewrote the abstract and introduction, and reorganized the methodology around entropy conditioning and denominator removal. The main-text ablations now separately examine these two choices and compare simple entropy weighting with a jointly learned witness. The fitting theory provides mathematical support for the revised score, with its assumptions stated explicitly. Section 3 also explains the gap in AdaDetectGPT’s claimed TNR bound and our corrected fitting rationale. We believe this revision implements the substantive refocusing that the reviewer recommended for a hypothetical resubmission within the present review process.

Reviewer nd7Z — Contributions, broader evaluation, and covariance modeling. We clarified the contributions relative to AdaDetectGPT and added controlled comparisons that distinguish score design from direction fitting. Comparisons with L2 logistic regression and separate-covariance Gaussian Neyman–Pearson fits are now reported on identical features. We explicitly acknowledge their results and present the covariance-sum fit as an efficient, theoretically characterized fitting procedure. We also expanded the evaluation to held-out generators, alternative decoding temperatures, paraphrased passages, and mixed lengths. For unequal class covariances, we distinguish Gaussian AUROC optimality from fixed-FPR power optimality, state the conditions required for the latter, and provide covariance diagnostics and power-loss bounds. These responses and their supporting evidence have been incorporated into the revised manuscript.

Reviewer WbCX — Sources of improvement, modeling assumptions, and additional validation. We clarified how entropy conditioning, denominator removal, and the theoretical correction contribute to the method. We added entropy diagnostics at comparable token log probabilities, mixed-length evaluations under a shared threshold, recent generators including Llama-3.1-8B and Qwen3-8B, recent baselines including SurpMark and Uncertainty++, and evaluations on DetectRL and RAID. We also expanded the Gaussian and covariance diagnostics and compared fits using separate class covariances. The manuscript now explicitly states that the Gaussian approximation guides fitting, while marginal FPR control follows from independent human calibration under exchangeability. The experimental findings and their limitations are reported in the final revision.

We believe these changes substantially address the concerns about contribution, experimental coverage, and covariance assumptions, and have incorporated the corresponding explanations and evidence into the manuscript. We respectfully hope that Reviewers 2 and 3 will reconsider their assessments in light of these revisions, and that the substantive refocusing requested by Reviewer 1 will be considered in the final decision. We would welcome further discussion of any remaining concerns.

Thank you very much for your time and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.