acceptodds
Under review as a conference paper at ICLR 2027

Tide: Zero-Shot LLM-Generated Text Detection via Directional Generation Asymmetry

Abstract

With the rapid advance and wide use of large language models (LLMs), the text they generate is flooding into classrooms, publications, online forums and beyond. Telling it apart from human writing has become a practical necessity as concerns over misuse and attribution grow. Many existing detectors measure how predictable a text is to a surrogate LLM, reading it from left to right (i.e., in forward order), the order in which it was produced. However, we observe that adding a reading in the opposite direction (i.e., in reverse order) reveals a complementary and even stronger signal for detection. Specifically, reversal makes every text harder for the surrogate to predict, but LLM-generated text loses disproportionately more predictability than human text. We call this effect *directional generation asymmetry* and build our detector, Tide, on it. Tide computes the cross-entropy losses of a text and its reversal in one batched call to the frozen surrogate and combines the two losses into a single directional asymmetry score. On our new dataset of 225,000 documents spanning five domains, five frontier LLMs, six languages and six adversarial attacks, Tide outperforms nine state-of-the-art zero-shot baselines, achieving an average AUROC of 0.978, 4.7% higher than the best baseline. At a practical operating point of a 0.1% false-positive rate, Tide nearly doubles the true-positive rate of the best baseline, from 41% to 79%.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM-Generated Text Detection, Zero-Shot Detection, Large Language Models

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 4 reviews, above 85% of the others.
Reviewer_bdtDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work discusses how to create a zero-shot detector of LLM-generated text using a text's token sequence and its reversed copy with a frozen LLM. The hypothesis that the authors state is about how LLM text produces more unpredictability under reversal than human text.

Critical strengths
  • The method itself is zero-shot, very simple and cheap to run. The experiments conducted were also especially extensive and useful.
Critical weaknesses
  • I'm not entirely sure if reversal is the only thing that is useful here. According to Table 9, it seems that any other form of corruption is sufficient.
  • The paper's novelty is also limited. Several past works [1, 2] discuss using a reverse loss / cross-entropy for detection.
  • I also feel like the attacks are not realistic to evaluate this method. Maybe using RAID [3] as a benchmark could be useful here.

[1] Yu et al. Reverse Modeling in Large Language Models. NAACL 2025. [2] Guo et al. BiScope: AI-generated Text Detection by Checking Memorization of Preceding Tokens. NeurIPS 2024. [3] Dugan et al. RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors. ACL 2024.

Questions for authors
  • Does this hypothesis still hold for reverse-style LLMs (like LEDOM)?
  • What is the FPR on polished human prose (law articles, academic abstracts) and on texts under 150 words? Limited texts is also an interesting ablation to run.
  • How was D⋅rD \cdot r chosen in the first place (apart from the empirical study in Table 10)?
5 comments
AuthorsOpenReview ↗

Responses to Reviewer bdtD (1/2)

We thank the reviewer for the constructive suggestions and the pointers to related work. Here are our point-to-point responses:

W1 (is reversal the only useful corruption, Table 9).

We agree that on English with Qwen3.5-4B, the first column of Table 9, the existing corruptions seem sufficient. For example, the 4-token block shuffle reaches 0.994, compared to 0.995 for TIDE. However, the other three columns of Table 9 show that they fail once the surrogate or the language changes. For example, in the other languages, the best corruption reaches 0.960 with Qwen3.5-4B and 0.860 with Llama-2-7B-chat against 0.970 and 0.925 for TIDE, and the best corruption changes from column to column. The reason is that the existing corruptions either keep most of the forward order, changing the direction of reading only in part, or destroy the structure of the text, which raises the loss of both classes alike, whereas reverse reading in TIDE maintains every adjacency of the text and flips only the direction of reading. Therefore, the signal is the direction of reading rather than the amount of disruption, and TIDE's reverse reading is the only one that measures it in full.

W2 (novelty relative to Yu et al. [1] and BiScope [2]).

Yu et al. [1] train a model on both reading orders and use the gap between its forward and reverse losses, which tracks text quality, to select pretraining data. TIDE trains nothing and reads the reversed text with a frozen forward-only surrogate, so its reverse loss is the cost of reading against the direction of generation. The plain difference of the two losses is the DD row of Table 10, and the product with the ratio that TIDE uses improves on it, for example from 0.953 to 0.970 on non-English languages with Qwen3.5-4B.

As for BiScope, it does not change the reading direction when the surrogate reads the text. Its backward cross-entropy only looks back one position when the loss is computed. As suggested by the reviewer, we added BiScope [2] as a baseline on every cell of Table 1, trained once on the English training split and applied unchanged, as shown in the following table (the better one per column is in bold):

MethodClaudeDeepSeekGeminiGPTMistralzhdearesfrAttacksAvg.
Qwen3.5-4B
TIDE (zero-shot)0.9981.0000.9990.9801.0000.9410.9860.9630.9830.9770.9360.978
BiScope (trained on English)1.0001.0001.0001.0001.0000.5320.9610.9700.9850.9650.9630.943
Llama-2-7B-chat
TIDE (zero-shot)0.9950.9980.9950.9730.9990.8900.927–0.9520.9300.9540.961
BiScope (trained on English)0.9980.9990.9980.9970.9990.6880.862–0.8480.8890.8790.916

The results show that BiScope matches or exceeds TIDE on the English cells that share a similar distribution with its training set but falls behind on the other languages, and overall TIDE is 0.035 and 0.045 better than BiScope on the two surrogates. We have added BiScope to Table 1 of the revised paper, marked in blue.

AuthorsOpenReview ↗

Responses to Reviewer bdtD (2/2)

W3 (realism of the attacks, RAID [3]).

As suggested by the reviewer, we evaluated on RAID, which attacks human-written and machine-written text alike. We ran TIDE and the baselines of Table 1 with Qwen3.5-4B on the eight domains of RAID, 500 human-written documents per domain and 500 from its six chat generators, each document under each attack in the table, and report the AUROC per attack averaged over the domains in the following table (best per column in bold):

MethodNoneAlt. spell.Art. del.Homogl.Ins. par.NumberMisspell.SynonymUp/lowWhitesp.ZWSPAvg.
Entropy0.9160.9120.9030.8910.9130.9190.9120.8670.9100.8900.8670.900
Log-Rank0.9190.9160.9170.8890.9190.9220.9170.8650.9150.9040.8970.907
LRR0.9130.9110.9190.8220.9140.9150.9140.8590.9150.9090.8250.892
Fast-DetectGPT0.7800.7680.8260.7230.7450.7800.7720.7150.7320.7040.8040.759
Binoculars0.7930.7760.8270.7520.7620.7860.7770.6540.7130.7110.8410.763
DNA-DetectLLM0.7530.7360.8010.7730.6870.7510.7360.6900.6660.5670.8630.729
Lastde0.9010.8950.9010.7780.8960.8910.8980.6760.8830.9030.8590.862
SpecDetect0.9110.9000.9060.8540.9110.9110.9020.7740.8930.8950.8660.884
TIDE0.9350.9340.9190.9420.9280.9320.9330.8020.9280.8970.9140.915

The results show that TIDE is the best method on nine of the eleven attacks and on average. We have added this table to Appendix N of the revised paper, marked in blue.

Q1 (Does the hypothesis still hold for reverse-style LLMs, such as LEDOM).

Yes, with the two directions exchanged. We use the public preliminary checkpoint of LEDOM-7B on the 25 English cells of Table 1 with the framework of TIDE unchanged. The following table gives the class means of the two losses (human-written / LLM-generated) and the AUROC of TIDE, the last column with the two losses exchanged:

SurrogateLfwdL_{\mathrm{fwd}}LrevL_{\mathrm{rev}}TIDETIDE, losses exchanged
Qwen3.5-4B (forward)2.73 / 1.826.05 / 6.250.995–
Llama-2-7B-chat (forward)2.50 / 1.866.63 / 7.240.992–
LEDOM-7B (reverse-trained)6.27 / 6.792.47 / 2.170.5480.947

The results show that TIDE works with a surrogate trained in either direction, because it checks the difference between the two readings. For the reverse-trained model the reversed text is the native reading, LLM-generated text is again the class that costs more to read against the direction of generation, and TIDE with the two losses exchanged reaches 0.947 AUROC. The 0.548 of the original order is expected, since for a reverse-trained model the original text is the reading against its training direction, so the roles of the two losses are swapped.

Q2 (FPR on polished human prose and on texts under 150 words).

We test on the academic-abstracts domain of RAID and report the FPR of TIDE and the baselines of Table 1. The threshold of each method is set on the human-written documents of our English training split and applied unchanged, for TIDE and the average of the baselines, as shown in the following table:

MethodFPR, all abstractsFPR, under 150 words
Baselines0.0690.071
TIDE0.0230.031

The results show that TIDE falsely flags 2.3% of the abstracts and 3.1% of the short ones against 6.9% and 7.1% for the baselines on average, and at the same thresholds detects 87% of the machine-written abstracts of RAID against 75%. The reason is that TIDE scores the difference between the two readings of the same document, so what polished or short text changes in the forward loss it changes in the reverse loss as well, and the threshold holds. We have added this table to Appendix N of the revised paper, marked in blue.

Q3 (how D⋅rD \cdot r was chosen).

We chose the score in two steps. First, we observed that LLM-generated text loses more predictability than human writing when the surrogate reads it in reverse (Section 3). Second, to measure this directional asymmetry, we tried several candidates. Among them, the difference and the ratio are the most intuitive and the best-performing ones. We evaluated both and found them complementary, e.g., the ratio is stronger on news and the difference on peer reviews in Table 10, and their product combines them without a weight and adjusts itself to the scale of the losses (see W1 in our reply to Reviewer msKX). Therefore, D⋅rD \cdot r is the observed asymmetry, written in the form that the empirical comparison singled out.

AuthorsOpenReview ↗

Gentle Follow-Up

Dear Reviewer bdtD,

Thank you for your detailed and constructive review. We have posted our response and uploaded a revised PDF with the new experiments and clarifications marked in blue. As the author-reviewer discussion period is nearing its end, we would like to kindly check whether there are any remaining concerns that we can help address. We would be more than happy to provide any further results or discussion that would be useful.

Best regards,

Authors

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

This paper proposes Tide, a zero-shot LLM-generated text detector based on directional generation asymmetry. By combining the forward and reverse prediction losses of a frozen surrogate LLM, Tide achieves strong detection performance without additional training. Experiments across multiple domains, generators, languages, and adversarial attacks demonstrate improvements over existing methods.

Critical strengths
  1. The paper introduces directional generation asymmetry as a potentially novel signal for detecting LLM-generated text. Figure 1 shows that combining forward and reverse losses improves AUROC from 0.81 to 0.95, while Appendix K demonstrates consistent improvements over forward loss across 16 surrogate models.

  2. Tide outperforms nine zero-shot baselines, improving average AUROC from 0.934 to 0.978 on Qwen3.5-4B, with relatively low inference overhead. The authors also provide comprehensive ablations and analyses covering sequence perturbations, surrogate models, document lengths, adversarial attacks, and LLM-edited text.

Critical weaknesses
  1. The underlying mechanism and novelty relative to perturbation-based detection remain insufficiently established. Although Appendix I provides extensive perturbation controls, several generic perturbations achieve nearly identical performance to full token reversal on English (e.g., 0.994 for 4-token block shuffling versus 0.995 for Tide). This raises questions about whether the improvements specifically reflect generation direction rather than general sensitivity to sequence perturbations. Furthermore, closely related text-reordering detection methods discussed in §2.2 are not included in the direct baseline comparisons.

  2. The evaluation raises concerns about real-world generalizability. The LLM-generated texts follow a controlled generation pipeline, and all non-English test sets are translated from English rather than naturally written in their respective languages. Moreover, the low-FPR evaluation selects thresholds separately using human-written samples from each evaluation cell, leaving cross-domain threshold transfer untested. Evaluations on independent, naturally collected datasets with transferable thresholds would strengthen the practical claims.

Questions for authors

N/A

2 comments
AuthorsOpenReview ↗

Responses to Reviewer DJMC (1/2)

We thank the reviewer for the appreciation and the constructive suggestions. Here are our point-to-point responses:

W1 (mechanism and novelty relative to perturbation-based detection, and text-reordering baselines).

Table 9 in our paper shows the difference between TIDE's reverse reading and the perturbation techniques from two perspectives. First, none of the existing perturbation techniques has the same influence on the model loss as TIDE's reverse reading. Second, TIDE's reverse reading is more consistent across settings. For example, on the other languages it is 0.076 and 0.141 higher than the perturbation techniques on average on the two surrogates, respectively. Therefore, TIDE's reverse reading is different from and better than those perturbations.

As suggested by the reviewer, we added the three methods of Section 2.2 as baselines, as shown in the following table, with the same layout as Table 1. Luminol-AIDetect and Sun and Lv have no public code, so we reproduced both from their papers by ourselves. D&R runs with its official code and Claude Haiku 4.5 as the recovery model. The baseline that needs training uses the English training split.

MethodClaudeDeepSeekGeminiGPTMistralzhdearesfrAttacksAvg.
Qwen3.5-4B
TIDE0.9981.0000.9990.9801.0000.9410.9860.9630.9830.9770.9360.978
Luminol-AIDetect0.9920.9960.9890.9700.9940.8730.9100.8970.9250.8880.8890.938
Sun & Lv0.9120.9490.9740.7630.9510.5230.6060.5230.6840.6180.7390.749
Llama-2-7B-chat
TIDE0.9950.9980.9950.9730.9990.8900.927–0.9520.9300.9540.961
Luminol-AIDetect0.9560.9750.9580.8170.9740.6840.781–0.8310.8130.8280.862
Sun & Lv0.8150.8000.8830.6310.9090.5010.470–0.5590.4840.6550.671
Surrogate-independent
D&R (official classifier)0.9730.9900.9510.9780.9480.5200.7700.6680.7820.6960.9460.838

The results show that TIDE is the best on every column. Luminol-AIDetect is the closest at 0.938 on Qwen3.5-4B and 0.862 on Llama-2-7B-chat against 0.978 and 0.961 for TIDE. For Sun and Lv the reordering similarity carries less signal. D&R is strong on English and the attacks but falls to 0.52 to 0.78 on the languages other than English.

AuthorsOpenReview ↗

Responses to Reviewer DJMC (2/2)

W2 (real-world generalizability: controlled generation, translated languages, cross-domain threshold transfer).

We want to clarify that the generation pipeline of our dataset follows existing benchmark papers, such as DetectRL. The constraints we add on top of it only make detection harder, since they remove the signal of plain text features such as word count. We did not make the generation more controlled to ease detection.

As suggested by the reviewer, we ran TIDE and the baselines of Table 1 on the multilingual subtask of M4GT-Bench, where the non-English texts are written natively in those languages without translation, as shown in the following table:

LanguageEntropyLog-RankLRRFast-DetectGPTBinocularsDNA-DetectLLMLastdeSpecDetectTIDE
Arabic0.8510.8810.8950.6080.6780.7000.8230.9000.921
Chinese0.8270.8490.8690.5690.5620.4520.8470.8950.907
Italian0.9980.9980.9980.4240.4150.4340.8980.9860.997
Average0.8920.9090.9210.5340.5520.5290.8560.9270.942

The results show that on natively written non-English text, TIDE is still the best method on average.

The TPR at fixed FPR in the submission is read from the ROC curve of each cell, the protocol commonly used in previous studies, such as Binoculars and Fast-DetectGPT. As suggested by the reviewer, we tested the transfer of thresholds. The following table sets the thresholds on the English training split of each domain, applies them unchanged to the language and attack cells of that domain, and averages the TPR values over the generators, domains and the non-English conditions. The best per column is in bold:

Methodqwen3.5-4b TPR@FPR=10%qwen3.5-4b TPR@FPR=5%qwen3.5-4b TPR@FPR=1%llama2-7b-chat TPR@FPR=10%llama2-7b-chat TPR@FPR=5%llama2-7b-chat TPR@FPR=1%
TIDE0.8080.7700.6690.9050.8530.705
Entropy0.7080.6260.3940.5710.4830.311
Log-Rank0.6130.5030.3110.6060.4860.253
LRR0.4150.3170.1880.4000.3030.166
Fast-DetectGPT0.2240.1460.0550.8300.7630.597
Binoculars0.1780.0960.0220.7360.6220.364
DNA-DetectLLM0.2280.1370.0460.7760.6820.456
Lastde0.1290.0880.0330.5660.4970.388
SpecDetect0.3700.2720.1370.3180.2390.128
DNA-GPT0.6800.5610.2970.3460.2070.057

The results show that TIDE keeps the highest TPR at every operating point on both surrogates, showing the consistent and robust performance of TIDE when the threshold is calibrated on one data split but tested on others.

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

The paper proposes TIDE, a zero-shot machine text detection method based on directional generation asymmetry in text. The paper observes that reversing the token order increases prediction loss more for LLM-generated text than for human-written text. Based on this observation, TIDE computes the forward and reverse losses using a surrogate model and combines them into a single detection score. The extensive experiments, covering diverse domains, generators, and baseline detectors, demonstrate that TIDE outperforms existing zero-shot detection methods.

Critical strengths
  • The paper is overall well written and easy to follow.
  • The paper proposes a new calibration technique for zero-shot machine text detection, motivated by the difference in how LLMs and humans write text. It mainly leverages a loss for the reversed text, which is expected to cancel text-specific factors such as topic and wording and lead to better classification performance.
  • The extensive experiments, covering diverse domains, generators, and baseline detectors, demonstrate the superiority of the proposed method in detection performance.
  • The broad analysis including surrogate model sensitivity, adversarial robustness, document length, score function design, and reversal controls, support the effectiveness of the proposed method in various cases.
Critical weaknesses
  • The paper does not sufficiently justify why reverse loss is necessary for improving detection performance. The paper claims that comparing forward and reverse losses cancels text-specific factors such as topic, wording, and intrinsic difficulty. However, similar calibration might be achieved using other reference losses, such as the forward loss of another surrogate model or the average loss of perturbed texts similarly in DetectGPT. While the paper explores different ways of corrupting the original text, comparing these alternative reference losses using the same scoring function would help clarify whether reverse loss is particularly effective.
  • The paper does not sufficiently explain why human-written and LLM-generated texts exhibit different directional asymmetries. While LLMs generate text autoregressively, the paper attributes this difference to the fact that humans plan and revise their writing. However, it is unclear why this difference would lead to a smaller increase in reverse loss for human-written text, especially when the surrogate model is trained only on forward text.
Questions for authors

See the weakness part

5 comments
AuthorsOpenReview ↗

Responses to Reviewer mXhB (1/2)

We thank the reviewer for the insightful suggestions. Here are our point-to-point responses:

W1 (is the reverse loss necessary, compared with other reference losses in the same scoring function).

As suggested by the reviewer, we kept the scoring function S=(Lrev−Lfwd)⋅Lrev/LfwdS=(L_{\mathrm{rev}}-L_{\mathrm{fwd}})\cdot L_{\mathrm{rev}}/L_{\mathrm{fwd}} and replaced the reverse loss LrevL_{\mathrm{rev}} by the two other reference losses the reviewer recommends: (a) the forward loss of another surrogate model, for which we use the base model of the same family as in Binoculars, and (b) the average loss of perturbed texts as in DetectGPT, for which we use the expected forward loss of the text with each token resampled from the base model, computed as in Fast-DetectGPT, which is faster and more accurate than the perturbation of DetectGPT. We ran both on every cell of Table 1 with the same two surrogates, as shown in the following table with the same layout as Table 1:

MethodClaudeDeepSeekGeminiGPTMistralzhdearesfrAttacksAvg.
Qwen3.5-4B
TIDE (reverse loss LrevL_{\mathrm{rev}})0.9981.0000.9990.9801.0000.9410.9860.9630.9830.9770.9360.978
(a) forward loss of another surrogate model0.7390.3350.9270.2800.5340.6580.5720.7290.6100.5510.5450.589
(b) average loss of perturbed texts0.6790.5180.7860.5010.7040.5460.4290.4950.4940.4780.5710.564
Llama-2-7B-chat
TIDE (reverse loss LrevL_{\mathrm{rev}})0.9950.9980.9950.9730.9990.8900.927–0.9520.9300.9540.961
(a) forward loss of another surrogate model0.9390.9640.9570.9230.9780.6580.795–0.8080.8250.9480.879
(b) average loss of perturbed texts0.9730.9880.9650.9300.9890.7710.844–0.9010.9150.9300.921

The results show that neither new reference loss matches the reverse loss. On Qwen3.5-4B the two references average 0.589 and 0.564, and which class receives the higher score changes between generators, so no fixed direction of the score holds across cells. On Llama-2-7B-chat they average 0.879 and 0.921 and fall to 0.658 to 0.844 on the other languages, against 0.961 and 0.890 to 0.952 with the reverse loss that TIDE uses. We think the reason is that both references are still forward losses, which read the text from left to right. They may add information that the forward loss alone lacks, but not as much as the reverse loss, which reads the text in the opposite direction and so complements the forward loss. We have added this table next to Table 10 (Appendix J) of the revised paper, marked in blue.

AuthorsOpenReview ↗

Responses to Reviewer mXhB (2/2)

W2 (why human-written and LLM-generated text differ in directional asymmetry under a forward-only surrogate).

To clarify, in our submission, the asymmetry mainly comes from the direction in which each class of text is produced. An LLM is trained and run to predict each token from the tokens before it, so the text it generates is aligned with a left-to-right reading, the direction that a causal surrogate shares. Read in reverse, this alignment is lost, since nothing in the generation made a token likely given the tokens after it, so the reverse loss of LLM-generated text rises far above its forward loss. A human writer is not bound to one direction. They often have the whole text in mind before writing and choose each word with a view to what precedes and what follows it, so human writing carries dependence in both directions and loses less when reversed. Existing work shares this view of human writing. M-RangeDetector (Jiao et al., Findings of ACL 2025) builds its detector on the observation that LLMs attend to the preceding tokens when generating a token, whereas humans consider bidirectional, global and local contexts. The following table shows the two effects with the mean forward and reverse loss per class on the English test cells of Table 1:

SurrogateClassLfwdL_{\mathrm{fwd}}LrevL_{\mathrm{rev}}
Qwen3.5-4Bhuman-written2.736.05
Qwen3.5-4BLLM-generated1.826.25
Llama-2-7B-chathuman-written2.506.63
Llama-2-7B-chatLLM-generated1.867.24

Forward, LLM-generated text is easier for a causal surrogate than human writing, with a mean forward loss of 1.82 against 2.73 on Qwen3.5-4B and 1.86 against 2.50 on Llama-2-7B-chat. Backward, this advantage is gone and LLM-generated text becomes the harder class, with a mean reverse loss of 6.25 against 6.05 on Qwen3.5-4B and 7.24 against 6.63 on Llama-2-7B-chat. Reversal therefore costs LLM-generated text more than human writing, which shows that the predictability of LLM-generated text depends more on the reading direction. The surrogate does not need to have seen reversed text during its training. The reversed pass is the same measurement applied to both classes, and what it measures is how much of a text's predictability depends on the reading direction. All 16 surrogates of Table 11, none trained on reversed text, show the asymmetry. We have added this explanation, with the class means above, to Section 3 of the revised paper, marked in blue. We also ran TIDE with a reverse-trained LLM as the surrogate, in response to Q1 of Reviewer bdtD, and the asymmetry flips with the training direction, so LLM-generated text is again the class that costs more to read against the direction of generation, and TIDE with the two losses exchanged reaches 0.947 AUROC.

AuthorsOpenReview ↗

Gentle Follow-Up

Dear Reviewer mXhB,

Thank you for your detailed and constructive review. We have posted our response and uploaded a revised PDF with the new experiments and clarifications marked in blue. As the author-reviewer discussion period is nearing its end, we would like to kindly check whether there are any remaining concerns that we can help address. We would be more than happy to provide any further results or discussion that would be useful.

Best regards,

Authors

Reviewer_msKXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TIDE, a novel zero-shot method for detecting LLM-generated text. Specifically, TIDE computes the forward and reverse cross-entropy losses of a given text using a surrogate language model and combines these two quantities into a detection score. Empirically, TIDE outperforms nine state-of-the-art zero-shot baselines in terms of AUROC across various domains, languages, and adversarial settings.

Critical strengths
  1. The paper is well organized, and the main idea is clearly presented and easy to follow.
  2. The empirical results are promising, demonstrating substantial improvements over existing baselines.
Critical weaknesses
  1. Although Section 4.4 provides some intuition for combining the absolute and relative differences between the forward and reverse losses, it remains unclear why the authors choose the specific multiplicative form S=(Lrev−Lfwd)(Lrev/Lfwd)S=(L_{\mathrm{rev}}-L_{\mathrm{fwd}})(L_{\mathrm{rev}}/L_{\mathrm{fwd}}). Could the authors provide a more compelling intuitive or theoretical justification for this particular choice?

Additionally, I wonder whether the authors have considered training a small neural network that takes the two losses as inputs and outputs a detection score. Although this approach would require labeled training data and therefore depart from the zero-shot setting, it could serve as an informative benchmark for assessing how much information is captured by the proposed hand-crafted score.

  1. Further analysis of the ablation results would be valuable. The authors have conducted several insightful ablation studies, but I believe the analysis could be strengthened. In particular, it would be helpful to quantify how much TIDE improves upon detectors that rely solely on hand-crafted statistics derived from reverse-direction token-level losses, and to provide an intuitive interpretation of the observed performance differences.
Questions for authors

Please see the critical weaknesses above.

4 comments
AuthorsOpenReview ↗

Responses to Reviewer msKX (1/2)

We thank the reviewer for the constructive suggestions. Here are our point-to-point responses:

W1 (justification of the multiplicative form, and a trained detector as a benchmark).

Why the multiplicative form. The difference D=Lrev−LfwdD=L_{\mathrm{rev}}-L_{\mathrm{fwd}} and the ratio r=Lrev/Lfwdr=L_{\mathrm{rev}}/L_{\mathrm{fwd}} measure the same excess of the reverse loss over the forward loss, and each lacks one thing the other has. The difference keeps the magnitude of the excess but ignores the base level, e.g., (Lfwd,Lrev)=(2,6)(L_{\mathrm{fwd}},L_{\mathrm{rev}})=(2,6) and (4,8)(4,8) have the same difference of 4 but ratios of 3 and 2. The ratio keeps the relative change against the base level but ignores the magnitude, e.g., (0.5,1.5)(0.5,1.5) and (2,6)(2,6) have the same ratio of 3 but differences of 1 and 4. Table 10 in our paper shows that the two are complementary, e.g., the ratio is the stronger factor on news while the difference is stronger on peer reviews, so a score that generalizes across domains should combine both. TIDE therefore multiplies them, for two reasons. First, the product rises fastest when both factors agree that the text is LLM-generated, and when one factor is weak it holds the other down, so a large difference on a text that is hard to read in both directions, or a large ratio on a tiny difference, does not score as high as a text that shows both. This self-adjustment pays off at low false-positive rates. For example, at a 1% false-positive rate on English with Qwen3.5-4B as the surrogate, the product detects 95.2% of the LLM-generated texts against 93.6% for the ratio and 82.6% for the difference. Second, the multiplicative form needs no hand-set weight for either factor, because the product is invariant to the scale of each factor: multiplying DD or rr by a constant multiplies every score by that constant and leaves the ranking unchanged. An additive form, e.g., aD+braD+br, adds a quantity on the loss scale of the surrogate to a ratio at a different scale, so the weights aa and bb have to be calibrated on labelled data to balance the influence of the two factors. We have added this justification to Section 4.4 of the revised paper, marked in blue.

Trained detectors on the two losses. As suggested by the reviewer, we trained a logistic regression, an MLP and a random forest on the two mean losses, and also on eight statistics (mean, max, min and std of the per-token loss in each direction), on the English training split, and tested them on the English test cells of Table 1, as shown in the following table. TIDE's score is within 0.001 of the best supervised detector trained on the two losses, and only 0.002 below the best detector trained on the eight statistics, on both surrogates, which means that TIDE's score extracts nearly all of the information the two losses carry.

FeaturesDetectorQwen3.5-4BLlama-2-7B-chat
Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}TIDE (zero-shot, no training)0.9950.992
Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}Logistic regression0.9960.992
Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}MLP0.9960.989
Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}Random forest0.9950.977
mean, max, min, std of Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}Logistic regression0.9970.989
mean, max, min, std of Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}MLP0.9970.993
mean, max, min, std of Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}Random forest0.9960.994

We then tested the detectors trained on the two losses on the other languages and the six attacks of Table 1, which they never saw in training, as shown in the following table (Languages and Attacks are the averages of the corresponding columns of Table 1):

Detector on Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}Qwen3.5-4B LanguagesQwen3.5-4B AttacksLlama-2-7B-chat LanguagesLlama-2-7B-chat Attacks
TIDE (zero-shot, no training)0.9700.9360.9250.954
Logistic regression0.9650.9370.9230.948
MLP0.9610.9360.9020.946
Random forest0.8650.9160.8360.924

TIDE matches or outperforms every trained detector on the other languages and the attacks on both surrogates, within 0.001 of logistic regression on the attacks with Qwen3.5-4B and otherwise ahead, and the gap grows with the capacity of the detector, to 0.105 and 0.089 on the other languages for the random forest. The fixed score therefore carries over to unseen languages and attacks, where a detector fitted on English drifts with the training distribution.

AuthorsOpenReview ↗

Responses to Reviewer msKX (2/2)

W2 (gain over detectors built only on reverse-direction token-level statistics, and interpretation).

In Table 10 of our paper (Appendix J), TIDE's score outperforms the LrevL_{\mathrm{rev}}-only score by 0.39 and 0.16 on average over the five domains. During the rebuttal, we ran new experiments on more reverse-direction token-level features, the max, min and std of the per-token LrevL_{\mathrm{rev}} next to its mean, each used as a zero-shot score in its better reading direction on every cell, and as the input of supervised detectors trained on the English training split, all evaluated on the English test cells of Table 1, as shown in the following table:

FeaturesDetectorQwen3.5-4BLlama-2-7B-chat
LrevL_{\mathrm{rev}} meanzero-shot0.7110.834
LrevL_{\mathrm{rev}} maxzero-shot0.5380.589
LrevL_{\mathrm{rev}} minzero-shot0.5790.656
LrevL_{\mathrm{rev}} stdzero-shot0.6580.604
mean, max, min, std of LrevL_{\mathrm{rev}}Logistic regression0.6410.870
mean, max, min, std of LrevL_{\mathrm{rev}}MLP0.6790.887
mean, max, min, std of LrevL_{\mathrm{rev}}Random forest0.6630.875
Lfwd,LrevL_{\mathrm{fwd}}, L_{\mathrm{rev}}TIDE (zero-shot)0.9950.992

The results show that no reverse-direction-only statistic separates the classes better than TIDE's score, and detectors trained on all four reach 0.68 on Qwen3.5-4B and 0.89 on Llama-2-7B-chat against 0.995 and 0.992 for TIDE. As an explanation, what identifies LLM-generated text is LrevL_{\mathrm{rev}} in excess of LfwdL_{\mathrm{fwd}} of the same document, a reference that an LrevL_{\mathrm{rev}}-only detector cannot learn. A high LrevL_{\mathrm{rev}} has two possible causes, a text that is hard to predict in either direction because of its topic, vocabulary or style; and a text that reads easily forward but not backward, which is the mark of LLM generation. LrevL_{\mathrm{rev}} alone cannot tell the two apart, so it contains only a weak signal. A text that is hard for the first reason is hard forward as well, so its LfwdL_{\mathrm{fwd}} is high too, whereas a text that is hard for the second reason has a low LfwdL_{\mathrm{fwd}}. Reading LrevL_{\mathrm{rev}} against LfwdL_{\mathrm{fwd}} of the same document therefore cancels the first cause and keeps the second, which is what TIDE's score does. We have added this table next to Table 10 (Appendix J) and this interpretation to Section 4.4 of the revised paper, marked in blue.

Thank you for the detailed responses, which have addressed most of my concerns, particularly regarding the motivation for the multiplicative form and its comparison with trained detectors. I suggest incorporating the intuitive justification for the construction of the test statistic into the main text, as it would help readers better understand the proposed method.

Overall, I am satisfied with the response and am willing to raise my score.

AuthorsOpenReview ↗

Thanks for your reply!

Thank you very much for your supportive reply! As suggested, we have added the intuitive justification for the construction of the test statistic to Section 4.4 of the revised paper in green, and we will make sure it is incorporated in the final version.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

We thank all the reviewers for their time and constructive feedback, and the Area Chair for handling our paper throughout the entire process. We appreciate the reviewers' positive comments on the clear presentation, the novelty and efficiency of the method, and the extensive experiments. We have responded to every comment, and we are glad that reviewers have confirmed that our rebuttal addressed most or all of their concerns. The revised paper includes new experiments and clarifications the reviewers suggested. Below we summarize our contributions, our responses to each reviewer, and the changes in the revised paper.

Novelty and Contributions

  1. We observe that LLM-generated text loses disproportionately more predictability than human writing when a causal LLM reads both in reverse order. We call this effect directional generation asymmetry and show that it complements the forward-order signals used by existing detectors.
  2. We propose TIDE, a zero-shot detector that reads a text in both orders in one batched call to a frozen surrogate LLM and combines the forward and reverse losses into a directional asymmetry score. TIDE achieves a higher AUROC than all nine zero-shot baselines on every generator, in every language and on the average over the six adversarial attacks, with the results averaged over the five domains.
  3. We build a dataset of 225,000 documents that spans five domains, five frontier LLMs, six languages and six adversarial attacks. With the default surrogate Qwen3.5-4B, TIDE improves the average AUROC from the best baseline's 0.934 to 0.978.

Responses to Each Reviewer

Reviewer msKX

  • W1: We explained the intuition behind the multiplicative form of the score, and compared TIDE with supervised detectors trained on the forward and reverse losses.
  • W2: We compared TIDE with statistics of the reverse loss used alone and with detectors trained on these statistics, and explained why the reverse loss is informative only together with the forward loss.

Reviewer mXhB

  • W1: We replaced the reverse loss in the score with the two reference losses the reviewer suggested, and compared both with TIDE on every cell of Table 1.
  • W2: We clarified where the asymmetry comes from and why it appears with a surrogate trained only on forward text, with the mean forward and reverse losses of each class of text.

Reviewer DJMC

  • W1: We explained with Table 9 how reverse reading differs from perturbations of the token sequence, and compared TIDE with the three suggested text-reordering detectors on every column of Table 1.
  • W2: We clarified that our generation pipeline follows existing benchmarks, evaluated TIDE on natively written non-English text from M4GT-Bench, and tested thresholds set on the English training split on the other languages and the attacks.

Reviewer bdtD

  • W1: We explained with Table 9 that the other corruptions of the token sequence fall behind reverse reading once the surrogate or the language changes.
  • W2: We clarified how TIDE differs from Yu et al. and BiScope, and compared TIDE with BiScope on every cell of Table 1.
  • W3: We evaluated TIDE and eight zero-shot baselines on the RAID benchmark under its more diverse attacks.
  • Q1: We ran TIDE with the reverse-trained LLM LEDOM-7B as the surrogate and showed that the asymmetry holds with the two directions exchanged, as expected.
  • Q2: We reported the false-positive rates of TIDE and the baselines on the human-written abstracts of RAID and on those under 150 words.
  • Q3: We explained that the difference and the ratio were the most direct and best-performing candidates for measuring the asymmetry, and that TIDE multiplies them because they are complementary.

Changes in the Revised Paper

The changes are marked in blue. Those made after Reviewer msKX's follow-up, in Sections 4.4 and 5.5, are marked in green.

  • Section 3: Added an explanation of where the asymmetry comes from, with the mean forward and reverse losses of each class (mXhB, W2).
  • Section 4.4: Added the intuition behind the multiplicative form of the score (msKX, W1) and the explanation of why the reverse loss is informative only together with the forward loss (msKX, W2).
  • Table 1 and Sections 5.1 and 5.2: Added the supervised detector BiScope (bdtD, W2).
  • Section 5.5: Added the true-positive rates of the product, the ratio and the difference at a 1% false-positive rate (msKX, W1).
  • Appendix J: Added Table 11 on statistics of the reverse loss used alone (msKX, W2) and Table 12 on the score with two other reference losses in place of the reverse loss (mXhB, W1).
  • Appendix N: Added Table 15 on the RAID benchmark (bdtD, W3) and Table 16 on the false-positive rate on its human-written abstracts (bdtD, Q2).

We thank all the reviewers and the Area Chair again for their time and effort.

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.