acceptodds
Under review as a conference paper at ICLR 2027

Resolving Score-Ordering Reversal in Watermarked Text: An Asymmetric Two-Sided Test for LLM Text Detection

Abstract

LLM watermarking and LLM-generated text detection are two complementary techniques for misinformation detection, intellectual property protection, and related provenance tasks. However, we found the two conflict: the watermarked LLM text can reverse the score-ordering of the statistical score by the state-of-the-art logits-based detectors, leading to failure in LLM text detection. We dive into this phenomenon to propose a new statistical measure for LLM-generated text and design a two-sided test accordingly. The test is enhanced threefold: 1) we design the upper and lower test branches separately with a pair of asymmetric statistical measures; 2) we remove the tail of the reference distribution due to the observation that both human and LLM rarely output low-probability tokens; 3) we use the context entropy to weigh each token position differently in the statistical measure for better discrimination. On a total of 660 configuration settings (LLMs watermarking methods datasets), our detector exceeds Fast-DetectGPT and AdaDetectGPT in terms of mean AUROC by 0.1929 and 0.1986, showing higher detection rates on a wider range of LLM text.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine-generated text detection, zero-shot detection, LLM watermarking, score-ordering reversal, asymmetric two-sided testing, likelihood-based detection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_V1U5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors investigate the issue of "score-ranking inversion" in machine-generated text detectors, which rely on language model probabilities—when watermarking techniques are applied. In such cases, some machine-generated texts receive lower detection scores than human-written texts, rendering detection rules based on a fixed, unidirectional threshold ineffective. The authors propose an asymmetric two-sided testing framework: the upper branch constructs a statistic using temperature adjustment, probability truncation, and context-entropy weighting, while the lower branch retains the original Fast-DetectGPT statistic. Both branches are calibrated using human-written text and then combined to reach a final decision. Experiments span five generative models, twelve datasets, and ten different generation scenarios (including both watermarked and non-watermarked settings).The proposed detector exceeds Fast-DetectGPT and AdaDetectGPT in terms of mean AUROC by 0.1929 and 0.1986.

Critical strengths

1.The research problem is practically important. Figure 1 and Table 3 show that different generation conditions can place machine-generated text on either side of the human-text distribution, exposing the limitations of fixed-direction detection. Addressing this failure mode without prior knowledge of the generation method is therefore valuable.

2.The method is concise and supported by a well-matched baseline. The authors construct a two-sided Fast-DetectGPT baseline using the same surrogate model, calibration sources, and budget allocation. Table 4 shows an AUROC gain of approximately 0.04–0.05 over this baseline, suggesting that the improvement goes beyond simply adopting two-sided detection.

3.The experiments are broad and well analyzed. The paper covers diverse generators, watermarking methods, and text domains, and reports AUROC, rank inversions, challenging-case performance, and low-FPR detection rates. Ablations on position weighting, reference distribution adjustment, and branch contributions further support the reported gains.

Critical weaknesses
  1. The score-ordering reversal phenomenon does not fully justify the specific asymmetric construction. Score-ordering reversal motivates considering deviations in both directions, but does not by itself determine which statistic should be used for each tail. The paper uses entropy weighting and reference shaping for the upper tail while retaining the original statistic for the lower tail. Section 5 provides empirical support, but does not isolate the gain from the asymmetric design itself. Comparing the proposed detector with a variant that applies the improved statistic to both tails would clarify whether the asymmetric construction provides additional benefits and under what conditions.

  2. The specific design of the reference distribution adjustment lacks sufficient motivation. Section 3 motivates temperature scaling and nucleus truncation by noting that low-probability tokens occur infrequently in both human and machine-generated text. However, low frequency does not imply negligible contributions to the reference mean and variance, and truncation changes both the center and scale of the standardized statistic. The paper shows empirical gains but does not sufficiently explain why this particular reference transformation improves human–machine separation. Ablating temperature scaling, nucleus truncation, and their combination would provide stronger evidence for the proposed design.

  3. The stated false-positive guarantee does not cover the complete parameter-selection and calibration procedure. Sections 4.1 and 5.3 indicate that the same human texts are used for hyperparameter selection and threshold calibration, while Proposition 5 requires the scoring rule to be independent of the calibration sample. Thus, the actual experimental pipeline does not fully satisfy the proposition's assumptions.

  4. The evaluation lacks a truly held-out test of unseen generation conditions. The same five generators and ten watermarking methods are used for both parameter selection and testing. Although the text samples may be independent, the generation conditions themselves are not held out. Therefore, the experiments do not directly establish whether the selected parameters and statistical construction transfer to unseen generators or watermarking methods.

Questions for authors
  1. Does the asymmetric design provide additional benefits over applying the improved statistic to both tails?
  2. Why does the proposed combination of temperature scaling and nucleus truncation improve human–machine separation? Can you provide ablations for the two components?
  3. Does the current parameter-selection and calibration procedure satisfy the assumptions of Proposition 5? Can you verify the FPR using independent development and calibration sets?
  4. How does the method perform on a generator or watermarking method that is completely held out during parameter selection?
5 comments

Thank you for the detailed and constructive comments. We provide additional structural comparisons, reference-shaping ablations, independent calibration results, and held-out generation-condition evaluations.

Throughout, two-sided FD denotes Fast-DetectGPT using its original statistic in both tails.

Weakness 1 and Question 1 : Why use an asymmetric design rather than the modified statistic in both tails?

Thank you for highlighting this missing comparison. We have added a symmetric variant that applies the modified statistic to both tails to directly evaluate the additional benefit of our asymmetric design. Both configurations use the same 660 evaluation settings, proxy models, 800 human calibration texts, and allocation γ=0.5\gamma=0.5, with each configuration separately calibrated.

Two-sided detectorOPT AUROCFalcon AUROC
Original statistic in both tails (two-sided FD)0.77690.7762
Modified statistic in both tails0.77140.7721
Ours (asymmetric)0.81750.8299

Our asymmetric design improves mean AUROC by approximately 0.046/0.058 (OPT/Falcon), with paired 95% confidence intervals of [0.0329, 0.0587]/[0.0481, 0.0701]. The comparison isolates the benefit of our asymmetric design.

The gains are particularly pronounced on CGW (+0.1486/+0.1966) and HeavyWater (+0.1226/+0.1720), with additional improvements on SemStamp (+0.0838/+0.1136), while remaining negligible on SynthID. These results show that the asymmetric construction provides substantial additional benefits under conditions with informative lower-tail evidence.

We will include this comparison and the complete condition-level results in Section 5 and the appendix.

Weakness 2 and Question 2 : Ablations for temperature scaling and nucleus truncation components

We agree that low-probability tokens can contribute substantially to the reference mean and variance. Our motivation for reference shaping is to adjust the centering and normalization of the detection statistic to improve its discrimination relative to the human score distribution.

Specifically, reference shaping changes the mean and variance used for standardization while keeping the observed-token log probabilities unchanged. Proposition 3 shows that increasing temperature and reducing nucleus mass shift the reference center in opposite directions in the single-model setting, while both affect the normalization scale. Proposition 4 further characterizes when these changes improve detection relative to the human-calibrated rejection boundary. These properties motivate exploring the combined transformation, whose effectiveness we evaluate empirically.

Component ablations. We fixed the entropy weights, lower-tail statistic, and γ=0.5\gamma=0.5, and separately calibrated each configuration on the same independent set of 800 human texts, disjoint from development and test data. Results are macro-averaged over the original benchmark’s 660 test settings (12 datasets × 5 generators × 11 generation conditions), reported as OPT / Falcon.

Upper-tail referenceAUROCTPR@1% FPRTPR@5% FPR
Original (T=1,c=1T=1,c=1)0.8150 / 0.813138.07% / 35.61%51.93% / 50.84%
Temperature only (T=1.2,c=1T=1.2,c=1)0.6670 / 0.679517.17% / 17.78%24.19% / 27.11%
Truncation only (T=1,c=0.8T=1,c=0.8)0.7034 / 0.714423.07% / 20.94%37.85% / 37.80%
Combined (T=1.2,c=0.8T=1.2,c=0.8)0.8175 / 0.829938.02% / 34.20%51.23% / 52.56%

The combined transformation outperforms either component alone across all three metrics. Compared with the original reference, it improves AUROC by 0.0025/0.0167 (OPT/Falcon), with paired 95% confidence intervals of [−0.0037, 0.0110]/[0.0052, 0.0298]. The gain is clearer for Falcon, while the original reference retains higher TPR@1% FPR for both proxies and higher TPR@5% FPR for OPT.

Benefits also vary across watermarking conditions. Averaged within each pair of methods, AUROC gains are 0.0298/0.0674 for ITSEdit/SemStamp, compared with 0.0047/0.0213 for CGW/HeavyWater (OPT/Falcon). In the OPT analysis (Appendix A.7), ITSEdit and SemStamp exhibit weaker score reversal, with original proxy-score medians closer to the human median. This suggests that reference shaping may provide greater benefits when machine-generated text is less clearly separated from human text. These condition-level differences are consistent with the smaller aggregate gain observed for OPT.

We selected T=1.2T=1.2 and c=0.8c=0.8 on the development set. We will refine the motivation in Section 3 and clarify the component ablations, condition-specific effects, and parameter sensitivity in Section 5.2.

Weakness 3 and Question 3 : Does the calibration procedure satisfy Proposition 5?

We agree that the original procedure did not fully satisfy the assumptions of Proposition 5, as the same 800 human texts were used for parameter selection and calibration. Proposition 5 establishes marginal FPR control when the scoring rules and allocation γ\gamma are fixed independently of the calibration and query samples, and these human samples are exchangeable.

To address this, we retained the original development set and fixed T=1.2T=1.2, c=0.8c=0.8, κ=1.5\kappa=1.5, and γ=0.5\gamma=0.5. We conducted two complementary evaluations: one to verify FPR under exchangeable sampling, and another to assess the corrected calibration procedure on the original benchmark.

1. FPR verification under exchangeable sampling. We constructed a new pool of 4,000 human texts, with 1,000 each from C4, SQuAD, Writing, and XSum, excluding identified overlaps, duplicates, and near-duplicates. Across 1,000 random splits, we uniformly partitioned the pool into 800 calibration and 3,200 test texts without per-corpus quotas. Conditional on the fixed pool and scoring rules, the calibration texts and a uniformly selected test query are exchangeable. Both tails were recalibrated for each split.

ProxyMean FPR at nominal 1%95% MC intervalMean FPR at nominal 5%95% MC interval
OPT0.992%[0.969%, 1.016%]4.953%[4.898%, 5.008%]
Falcon0.998%[0.974%, 1.023%]5.010%[4.956%, 5.065%]

The observed mean FPRs are consistent with the nominal levels under the exchangeable sampling design, as expected from Proposition 5. The intervals quantify Monte Carlo uncertainty across random splits.

2. Independent calibration on the original benchmark. We further evaluated the corrected calibration procedure while keeping the original detector parameters, proxy models, and 660 test settings unchanged. We replaced the original calibration bank with 800 additional human texts, comprising 200 from each of the four corpora and excluding identified overlaps with development and test sources. Both tails were recalibrated using the new bank.

CalibrationAUROC (OPT / Falcon)FPR@1% (OPT / Falcon)FPR@5% (OPT / Falcon)
Original0.8219 / 0.83171.31% / 2.29%6.69% / 6.53%
Independent0.8175 / 0.82991.06% / 1.31%6.14% / 6.19%

FPRs are measured on 2,360 unique human test texts at fixed nominal thresholds of 1% and 5%.

Independent calibration largely preserves detection performance, with only minor AUROC decreases of 0.0044/0.0018 (OPT/Falcon). The observed FPR decreases at both nominal levels, particularly for Falcon at 1%, where it drops from 2.29% to 1.31%. At the nominal 5% level, FPR remains slightly elevated (6.14%/6.19%), suggesting some sensitivity to differences between calibration and test-domain distributions.

We will update Sections 4.1 and 5.3 and Appendix C.3 to incorporate the independent-calibration protocol and results and clarify the scope of Proposition 5.

Weakness 4 and Question 4 : How does the method perform on generators or watermarking methods completely held out during parameter selection?

We conducted leave-one-generator-out (LOGO, 5 folds) and leave-one-watermark-out (LOWO, 10 folds) evaluations to assess performance on generation conditions excluded from parameter selection.

In each fold, we excluded the target generator or watermarking method from the machine-generated development data. We selected TT, cc, and γ\gamma through five-fold cross-validation grouped by the original 800 development sources, with κ=1.5\kappa=1.5 fixed. The selected parameters were then frozen, and our detector was calibrated using the independent 800-human bank introduced in W3/Q3. Evaluation covered 12 test domains, with 660 LOGO and 600 LOWO settings.

The table reports mean AUROC and TPR at matched observed FPR operating points.

DetectorLOGO AUROCLOWO AUROCLOGO TPR@1% FPRLOWO TPR@1% FPRLOGO TPR@5% FPRLOWO TPR@5% FPR
Likelihood0.52950.509716.81%16.70%25.52%24.79%
Entropy0.54490.551710.57%11.16%20.06%20.95%
Rank0.57020.549013.28%12.33%24.30%22.64%
Log-Rank0.54520.524717.87%17.71%26.64%25.79%
Fast-DetectGPT0.63880.611329.94%28.05%38.90%36.02%
AdaDetectGPT0.63310.606528.30%26.72%36.33%33.70%
Binoculars0.61350.587922.68%21.72%32.27%30.17%
NTS0.73410.725129.22%29.13%39.08%38.45%
Ours (OPT)0.71940.697526.63%23.30%40.78%37.41%
Ours (Falcon)0.82980.821031.13%29.91%51.82%50.13%

The Falcon configuration achieves the highest performance across all reported metrics. It outperforms Fast-DetectGPT by 0.1910/0.2097 AUROC and NTS by 0.0957/0.0959 under LOGO/LOWO. At 5% FPR, its TPR reaches 51.82%/50.13%, exceeding Fast-DetectGPT by 12.92/14.11 percentage points. Falcon consistently selected T=1.2T=1.2, c=0.8c=0.8, and γ=0.75\gamma=0.75 across all folds.

The OPT configuration shows more limited transfer. It improves AUROC over Fast-DetectGPT by 0.0806/0.0862, but remains below NTS. At 5% FPR, its TPR exceeds Fast-DetectGPT under LOGO and LOWO, while its TPR at 1% FPR is lower. Most OPT folds selected T=1.0T=1.0, c=0.8c=0.8, which also showed weaker performance in our reference-shaping ablations. Reselecting parameters using all development conditions yielded 0.7034/0.6983 AUROC on the corresponding task sets, suggesting sensitivity to parameter selection.

we examined how text length affects its detection statistic. Human texts in the development set had a median length of 168 scored tokens, compared with 21 in HellaSwag, 26 in ROCStories, and 68 in TLDR. These differences matter because, for single-model OPT at T=1,c=0.8T=1,c=0.8, truncation introduces a negative centering term whose magnitude grows approximately with neff\sqrt{n_{\mathrm{eff}}} when local mean shifts and variances are comparable (Appendix F.2, Eq. 44).

We therefore fixed neffn_{\mathrm{eff}} in this term at 133.76, its median over the 800 development human texts. We left the texts, reference variances, entropy weights, and lower-tail statistic unchanged. Both versions used κ=1.5,γ=0.5\kappa=1.5,\gamma=0.5 and were separately calibrated on the same independent set of 800 human texts. Fixing the length factor lowered mean AUROC from 0.8061 to 0.7896 over the 220 development settings, but raised it from 0.7034 to 0.8081 over the 660 test settings.

These results support parameter transfer for the Falcon configuration across held-out benchmark conditions, while showing greater configuration sensitivity for OPT. We will incorporate the held-out evaluations and parameter-selection results into the updated version.

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

This paper identifies a score-reversal phenomenon in LLM text detection, where certain watermarking methods assign machine-generated text lower scores than human writing. To address this, the authors introduce an asymmetric two-sided statistical test combining entropy-weighted upper-tail detection with conventional lower-tail detection. Extensive evaluations across diverse language models, watermarking schemes, and datasets demonstrate superior performance over existing approaches.

Critical strengths
  1. The proposed asymmetric two-sided test effectively addresses score-ordering reversal. Separating upper- and lower-tail statistics alongside human-based calibration is both theoretically sound and well justified.
  2. The method is computationally efficient and does not require additional model training. Its ability to handle both watermarked and unwatermarked text across different watermarking mechanisms suggests good practical applicability and robustness.
Critical weaknesses
  1. The evaluation relies on outdated models such as OPT, Falcon-7B, and Llama-2-7B. Testing the proposed method on more recent, capable LLMs would better demonstrate its generalizability and practical applicability.
  2. Although effective, the proposed approach largely combines established techniques—two-sided hypothesis testing, entropy-based weighting, and distribution truncation. Furthermore, its modest AUROC improvements over a symmetric two-sided baseline (from 0.04 to 0.05) call its methodological novelty into question.
Questions for authors
  1. Could the authors evaluate the proposed method on more recent and capable LLMs, such as Llama-3, Qwen, or Gemma? It would be helpful to understand whether the observed score-ordering reversal and the proposed detector's effectiveness generalize to modern model families.
  2. Given that the proposed approach builds on existing statistical techniques, could the authors further clarify its methodological novelty beyond conventional two-sided testing? In particular, what are the key advantages of the asymmetric design, considering the relatively modest improvements over the symmetric two-sided baseline?
1 comment

We thank you for recognizing the practical importance of the problem and for the suggestions regarding generalizability and methodological novelty.

Weakness 1 and Question 1 : Generalizability to recent LLMs

We agree that evaluating modern LLM families such as Llama-3, Qwen, and Gemma will further strengthen the empirical scope of the paper. Because running the full grid of 10 watermarking schemes across 12 datasets for these additional models requires substantial generation and scoring compute, we are currently completing these runs (with all detector hyperparameters and calibration protocols strictly frozen) and will present the results after completing the experiment.

Weakness 2 and Question 2: Methodological novelty and advantages of the asymmetric design

Our contribution includes both a two-sided detection formulation and the asymmetric construction of its tail statistics.

For all comparisons below, we retained the original 800 development sources. All two-sided detectors were calibrated using the same independent set of 800 human texts, disjoint from the development and test sets.

1. From one-sided to two-sided detection. Score-ordering reversal motivates testing both tails of the human score distribution, allowing the detector to capture machine-generated text with either unusually high or unusually low scores. The symmetric two-sided Fast-DetectGPT and AdaDetectGPT are constructed by us, which are different from the one-sided tests in their original papers. Across 660 evaluation settings, the Falcon results are:

DetectorOne-sided AUROCSymmetric two-sided AUROC
Fast-DetectGPT0.63880.7762
AdaDetectGPT0.63310.7572

Two-sided testing improves mean AUROC by 0.1374 and 0.1241, respectively, demonstrating the benefit of incorporating evidence from both directions.

2. From symmetric to asymmetric two-sided detection. Our analysis in Section 5 shows that different generation conditions favor different detection tails. In particular, entropy weighting reduces the strength of lower-tail evidence on CGW and HeavyWater, while changing the lower-tail reference center impairs pairwise ranking. Our asymmetric design therefore enhances upper-tail discrimination while preserving the original lower-tail evidence, without requiring prior knowledge of the watermarking method.

Under matched evaluation settings, proxies, calibration sources, and error allocation γ=0.5\gamma=0.5, we obtain:

Two-sided constructionOPT AUROCFalcon AUROC
Original statistic in both tails (two-sided FD)0.77690.7762
Modified statistic in both tails0.77140.7721
Our asymmetric construction0.81750.8299

Our asymmetric design improves AUROC by 0.0406/0.0537 (OPT/Falcon) over symmetric two-sided FD (both using lower-tail evidence) and by 0.0461/0.0578 over applying the modified statistic (upper-tail evidence) to both tails. The improvement isolates the benefit of using asymmetric statistics for the two-sided test.

Together, these two components yield an overall Falcon AUROC of 0.8299, improving upon the original Fast-DetectGPT and AdaDetectGPT by 0.1911 and 0.1968, respectively. We will clarify these two contributions and their corresponding comparisons in the manuscript.

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

The paper starts with the observation that watermarking LLMs' output conflicts with the AI generated text detectors. Second part of the paper is about designing a better detection method for AI generated text that takes into account the fact that the text might be watermarked. Their idea is basically a two sided test where they design the two branches separately and improve the overall effectiveness of the detection.

Critical strengths
  • First of all, the analysis of the phenomenon (that watermarking can hinder AI generated text detectors) has enough depth. The authors considered many watermarking methods, LLMs, and datasets.

  • Second of all, the results of their proposed detector are relatively good compared to the baselines on watermarked text.

  • The paper is clearly-written.

Critical weaknesses
  • The main weakness of the paper is in the first contribution, because the observation that watermarking can interact with AI generated text detectors has been made before [1]. On the bright side, I believe the authors have a more comprehensive study of the impact of watermarking in this submission, so they can rephrase the contribution accordingly in the updated version.

  • There is a small performance drop when using the two-sided test on unwatermarked data.

Minor issue: "Evaluation across [...] show -> "Evaluation across [...] shows"

References

[1] "When Detection Tools Collide: On the Impact of Watermarking on AI-Generated Text Detectors" (https://openreview.net/forum?id=dlYNARXBfZ)

Questions for authors
  • Do the authors have any thoughts on a way to improve the two-sided test on unwatermarked data?
1 comment

Thank you for the positive assessment and for highlighting the relevant prior work.

Weakness 1: Relation to prior work and the first contribution

We appreciate the pointer to [1], which observes watermark–detector interference and below-chance AUROC on KGW and Unigram. While our work starts from a related observation, our first contribution goes further in both scope and analytical depth:

  1. Broader taxonomy: We systematically evaluate ten watermarking schemes spanning four mechanism families: logit modification, probability reweighting, keyed token sampling, and sentence-level rejection sampling.
  2. Stage-wise theoretical and empirical diagnosis: Rather than only observing end-to-end detection degradation, we disentangle local proxy-score shifts at fixed generation contexts from document-level human–machine ordering.

We will cite [1] in both the Introduction and Related Work, and rephrase our first contribution as follows:

Characterizing detector-score shifts and score-ordering reversal across generation stages. Building on initial observations of watermark–detector interference [1], We conduct a comprehensive study of ten watermarking methods, theoretically and empirically analyzing their effects on local score shifts and document-level score-ordering reversal.

Weakness 2 and Question 1: Performance trade-off on unwatermarked text

The performance gap on unwatermarked text can be reduced by assigning a larger fraction of the false-positive budget to the upper tail through γ\gamma.

We retained the original 800 development sources and fixed T=1.2T=1.2, c=0.8c=0.8, and κ=1.5\kappa=1.5 for our detector. All two-sided configurations were calibrated using the same independent set of 800 human texts, disjoint from the development and test sets. We compared γ=0.5\gamma=0.5 with γ=0.75\gamma=0.75, keeping the branch statistics and calibration texts fixed. The table reports macro-averaged test AUROC over 60 unwatermarked settings and all 660 benchmark settings. Entries report OPT / Falcon.

Detector configurationUnwatermarked AUROCOverall AUROC (660 settings)
One-sided Fast-DetectGPT0.9155 / 0.91410.6376 / 0.6388
Symmetric two-sided Fast-DetectGPT (γ=0.5\gamma=0.5)0.8409 / 0.87270.7769 / 0.7762
Ours (γ=0.5\gamma=0.5, default)0.8697 / 0.89750.8175 / 0.8299
Ours (γ=0.75\gamma=0.75, upper-weighted)0.9076 / 0.91760.8299 / 0.8298

At γ=0.5\gamma=0.5, our asymmetric construction improves unwatermarked AUROC over symmetric two-sided Fast-DetectGPT by 0.0288/0.0248 (OPT/Falcon). Increasing γ\gamma to 0.75 further improves unwatermarked detection: OPT comes within 0.0080 AUROC of one-sided Fast-DetectGPT, while Falcon exceeds it by 0.0035. Overall AUROC at γ=0.75\gamma=0.75 is 0.8299/0.8298. We will include this allocation sensitivity analysis and guidance on γ\gamma selection in the revised manuscript.

Minor issue

Thank you for catching this, we will correct "Evaluation across [...] show" to "Evaluation across [...] shows".

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.