acceptodds
Under review as a conference paper at ICLR 2027

MELD: Multi-Task Equilibrated Learning Detector for AI-Generated Text

Abstract

Large language models are widely used in everyday writing, making reliable AI-generated text detection crucial for academic integrity, content moderation, and provenance tracking. Yet high AUROC on clean, in-distribution benchmarks is not sufficient. Practical detectors must resist adversarial rewrites, generalize to unseen generators and writing domains, and maintain low false-positive rates (FPR). A pooled AI-versus-human objective does not explicitly require the model to distinguish among generator families, so it may fail to learn the generator-specific structure needed for generalization and attribution. We introduce MELD (Multi-Task Equilibrated Learning Detector), which instead trains on class-balanced, family-specific AI-versus-human tasks while sharing a common representation of human writing. MELD produces detection, generator-family, rewrite-task, and token-level predictions in a single forward pass. Format normalization further makes its scores invariant to the modeled reformatting operations. MELD ranks first among open-source submissions in the public RAID leaderboard and matches or exceeds supervised baselines on five of six held-out evaluation pools. To evaluate transfer to unseen generators, we introduce MELD-eval, a held-out test pool built from four frontier chat models. Without further fine-tuning, MELD achieves 99.7% TPR at 1% FPR on MELD-eval and 98% TPR at the same FPR on a held-out generator whose family is absent from training. Finally, in a case study of 5.9 million scientific texts from 2016–2026, MELD’s prediction scores remain stable through 2022 and increase from 2023 onward, coinciding with the widespread adoption of LLM-based writing tools. The model and MELD-eval pool (https://huggingface.co/anon-review-meld-2026/meld) and code (https://anonymous.4open.science/r/MELD-4D74/README.md) are available anonymously.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
AI-generated text detection, multi-task learning, adversarial robustness, out-of-distribution generalization, generator attribution

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_dh2qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MELD, an AI-text detector trained through class-balanced, generator-family-specific detection tasks sharing a common representation of human writing. An Ettin-1B encoder projects token representations into a low-dimensional style space containing multiple human anchors and generator-family centers. Family-specific token scores are aggregated using top-quantile pooling, and training combines detection, family attribution, rewrite-task prediction, and formatting consistency. The work also introduces a large training mixture containing content-matched human–AI rewrite pairs and a held-out evaluation pool, MELD-eval. The authors report strong results across several benchmarks, including 99.7% TPR at 1% FPR on MELD-eval and 91.9% seven-way family attribution accuracy. A further case study analyzes detector-score trends in approximately 5.9 million scientific texts and peer-review summaries。

Critical strengths
  1. The data construction directly addresses important content and formatting shortcuts. The content-matched human–AI rewrite pairs are a substantive contribution. Matching the source content across labels reduces obvious topic and document-content mismatches, while symmetric formatting treatment targets superficial cues that could otherwise dominate detection. The manuscript explicitly diagnoses line-break artifacts in scientific prose and introduces normalization and augmentation to reduce them. This is a meaningful improvement over simply collecting larger but poorly matched human and machine corpora.

  2. The reported performance is strong across multiple evaluation settings. The results are not confined to a single favorable benchmark. Under the same MELD-eval calibration protocol, MELD reports 99.7% TPR compared with 95.5% for the strongest listed baseline. The broader benchmark results also indicate substantial potential as a practical detection system. The authors appropriately disclose less decisive results, including overlapping confidence intervals with the strongest competitor on M4GT, rather than presenting every numerical improvement as conclusive superiority

Critical weaknesses
  1. The experiments do not directly establish the benefit of the central family-specific detection objective. The central methodological claim is that class-balanced, family-specific AI-versus-human tasks improve detection and generalization relative to pooled binary training. However, comparisons with existing public checkpoints do not control for encoder architecture, training data, compute budget, normalization, or pooling. Consequently, they establish the performance of the overall system more clearly than the contribution of the proposed objective. Crucially, the “no family CE” ablation removes the auxiliary family-attribution loss; it does not replace the core family-specific detection loss with a pooled binary objective. It therefore does not answer the paper’s principal methodological question. A matched pooled-binary baseline using the same encoder, corpus, sampling plan, normalization, and training budget is needed. The anchor analysis has a related limitation. Replacing eight anchors with one in an already trained checkpoint, without reoptimization, demonstrates dependence on the learned representation but does not establish superiority over a single-anchor model trained under comparable conditions.

  2. The training–evaluation relationship for several public benchmarks is not sufficiently auditable. Section 5.2 describes MAGE, M4GT, and DetectRL evaluation pools as excluded from training, while Table 5 explicitly lists these datasets among the training sources. These statements could be consistent if disjoint splits are used, but the manuscript needs to distinguish dataset-level holdout, split-level holdout, and deduplicated evaluation within the same source distribution. These settings support different generalization claims. The reported near-duplicate removal against an evaluation registry is a positive safeguard. Nevertheless, datasets containing multiple generations and attacks derived from a common source require clear source-document or prompt-level provenance. The arXiv corpus has an explicit paper-level split, but comparable documentation is needed for the public benchmark components. I am not asserting that leakage occurred; the concern is that the current description does not allow the reader to establish the intended independence of the evaluations.

Questions for authors
  1. For RAID, MAGE, M4GT, and DetectRL, which exact splits contribute to training, validation, threshold calibration, and final evaluation? Are all generations and attack variants from the same original document or prompt assigned together? What precisely does “excluded from training” mean for each benchmark?
  2. How does pooled binary training perform with the same Ettin-1B encoder, training corpus, sampling distribution, formatting normalization, pooling rule, and optimization budget? The current auxiliary-loss ablation does not isolate the contribution of LfamL_{\mathrm{fam}}.
  3. Are the human examples used to estimate thresholds independent of those used to measure realized FPR? For MELD-eval, does the bootstrap resample original seeds or prompts as clusters, keeping their generated and attacked variants together? Does it propagate uncertainty from threshold estimation?
  4. What direct quantitative evidence supports AI-span localization and training-free new-family enrollment? Mixed-document detection AUROC does not establish token- or span-level localization accuracy. Similarly, an enrollment rule and a conditional score-stability bound do not establish new-family recognition performance or preservation of the old population’s FPR. Please identify the corresponding evaluations, including before-and-after enrollment results.
2 comments
AuthorsOpenReview ↗

Response to Reviewer dh2q (1/2)

Thank you for your positive assessment of our work. We believe MELD has the potential to make a meaningful contribution to the field, both as a strong open-source detection tool and as a way to explore broader patterns of LLM usage. We address each point below. The new tables are added to the revised manuscript: Table R6 is Table 19 (Appendix J), Tables R7 and R8 are Tables 7 and 8 (Appendix B), Table R9 is Table 9 (Appendix C) and Table R10 is Table 11 (Appendix E).

1. Matched pooled-binary baseline and separately trained one-anchor model (W1 and Q2).

Table R1 (please also refer to the response to Reviewer J5SJ) provides the matched comparison. All configurations share the backbone, training corpus, sampling plan, format normalization, pooling rule and training budget. Another two ablations keep MELD’s architecture and auxiliary losses and only replace the family-specific detection loss with pooled binary cross-entropy, with and without equal weighting of the positive families. MELD is ahead of all three on the Table 17 subset, on MELD-eval and on both DeepSeek tests. The ordering of the four matched configurations also holds on Muse Glimmer and Nemotron (Appendix J, Table 16).

As for the anchor, we train a separate one-anchor model from the same pretrained encoder with the same corpus, losses and schedule, and compare it with the full model on the Table 17 subset. The results in Table R6 demonstrate the performance gain of having multiple human anchors.

Table R6. Separately trained one-anchor model. Values are %.

ConfigurationHuman anchorsAUROCTPR at 1% FPR
One-anchor MELD, separately trained199.6797.02
MELD899.8098.25

2. Compositions of the public benchmarks (W2 and Q1).

Thank you for your comment, our work indeed combines many different data sources. We want to stress that we have been very rigorous to prevent any data leakage from training to testing, additionally we have also employed the near-duplicate removal strategy to further prevent evaluation on “near-seen” or “similar” data. We depict our training and testing set curation in Table R7 in detail, i.e. listing the exact partitions of the four benchmarks that enter training, validation, threshold calibration and final evaluation.

Table R7. Components of the public benchmarks used at each stage.

BenchmarkTrainingValidation and threshold calibrationFinal evaluation
RAIDOfficial train splitHeld out from the train splitOfficial test set (leaderboard)
MAGEAll MAGE rows outside the official test splitHeld out from the training rowsOfficial test split
M4GTOfficial train splitHeld out from the train splitOfficial test split
DetectRLOfficial train filesHeld out from the train files (validation only)Official test files

To reply to your individual questions: HC3 and Ghostbuster are held out as whole datasets (Appendix B, Table 6). RAID, MAGE, M4GT and DetectRL also contribute training data, and for them “excluded from training” refers to the evaluation set. Table 1 uses the official RAID test set scored through the public leaderboard (see also point 6 of our response to Reviewer wfQo), and Table 2 uses held-out partitions of MAGE, M4GT and DetectRL. Training examples that near-duplicate the evaluation registry were removed as well (104,660 in total, §4).

Regarding whether variants of one source stay together, the public benchmarks of Tables 1 and 2 are evaluated on official test partitions or on datasets held out as a whole, so we did not split them. Our own split exists only inside the training data, where it holds out rows for validation and for the fixed threshold. It is made per row rather than per source, and Table R5 shows that the results also hold when the threshold is set on held-out human texts.

To further ensure that the reported results do not rely on sources shared with training, we removed every test row that matches a training row exactly, shares at least 20% of its word 8-grams with the training data, or has the same opening as a training row, and evaluated the remaining rows again (Table R8). This check is stricter than the near-duplicate removal in §4, which drops a training example when at least half of its 8-word shingles appear in an evaluation set, so it still finds rows to remove. AUROC stays within 1 point on the three benchmarks.

Table R8. AUROC after removing all test rows that could potentially overlap the training data.

BenchmarkTest rows keptAUROC, all rowsAUROC, overlap removed
MAGE50,642 of 56,58999.0899.04
M4GT17,072 of 23,67277.4978.26
DetectRL33,142 / 34,76899.7799.76
AuthorsOpenReview ↗

Response to Reviewer dh2q (2/2)

3. Independent thresholds and bootstrap intervals (Q3).

For the main benchmark results, we use the same pool of human texts to set the threshold and measure FPR (Appendix C). To test independent calibration, we set the threshold on one half of each pool’s human texts and measure FPR on the other half. Table R5 compares this procedure with the fixed checkpoint threshold. On HC3, MAGE, and DetectRL, TPR remains close to the value obtained with a same-pool threshold.

Table R5. MELD with and without pool-specific calibration. The pool-calibrated threshold is set on one half of the pool’s human texts and evaluated on the other half. Values are %.

PoolMELD FPR, fixedMELD TPR, fixedMELD FPR, pool-calibratedMELD TPR, pool-calibratedModernBERT-Detect TPR, pool-calibrated
HC31.6799.201.0098.4674.13
MAGE1.8491.991.0089.5587.02
DetectRL1.6397.691.0196.9217.36
Short texts (≤192 words)5.4793.601.0789.9381.74

Even with a fixed threshold, MELD keeps the FPR at 1.6–1.8% on all three benchmarks. Pool-specific calibration brings it to 1.0% at a cost of only 0.7–2.4 TPR points, and MELD stays ahead of ModernBERT-Detect on every benchmark under the same protocol. For short texts, the same calibration gives 1.1% FPR at 89.9% TPR. In practice, most detectors would need calibration on human text from the target population to reach a low FPR (§3.4, §5.5).

For MELD-eval, we now further show results at the level of source seeds: calibration and evaluation use disjoint human seeds, and each seed stays together with all of its generated and attacked variants (Table R9). Appendix C did not state the resampling unit of the bootstrap. It now resamples seeds as clusters and re-estimates the threshold in every replicate, so the intervals also reflect the uncertainty of the threshold.

Table R9. MELD-eval with independent calibration seeds. Intervals are 95% percentile intervals from 1,000 seed-cluster bootstrap replicates, with the threshold re-estimated in each. Values are %.

ThresholdFPRTPR95% interval of TPR
Same pool (Table 3)1.0099.70[99.16, 99.83]
Independent calibration seeds1.0999.50[99.04, 99.79]

4. Span localization and new-family enrollment (Q4).

To provide quantitative evidence that supports the AI-span localization and training-free new-family enrollment we have carried out additional experiments.

  • For localization, experiments in Table R2 score the predicted spans against word-level labels on the public DAMASHA benchmark (Sai Teja et al., 2026) and two held-out mixed-document sets. Zero-shot MELD exceeds the supervised model on the latter two sets, the fine-tuned version outperforms the supervised model in the paper on DAMASHA. MELD also offers much lower human-word FPR.

Table R2. Span localization on the same documents. Word-level F1 (%) of the AI class on documents containing both human and AI text. Human-word FPR is the percentage of words in fully human documents that are marked as AI. Binoculars is scored on 64-word windows.

MethodDAMASHAarXiv interleavedContinuationHuman-word FPR (lower is better)
DAMASHA model (supervised)94.7258.4577.538.47
Binoculars39.983.9416.380.99
MELD (zero-shot)73.4175.8994.870.06
MELD (fine-tuned on DAMASHA data)98.7174.0191.430.34
  • Regarding enrollment, we enroll DeepSeek as a new family using 64 labeled documents (32 Flash and 32 Pro) and report the results in Table R10. We report detection and attribution before enrollment, after enrollment with the threshold unchanged, and after enrollment with the threshold re-estimated from human texts only. The results show that the enrolled family is recognized with 92.4% recall, while the attribution accuracy on the existing families is nearly unchanged. With the threshold unchanged, the human FPR rises slightly from 1.09% to 1.25%, and recalibration on human texts brings it back to 1.14%.

Table R10. DeepSeek-family enrollment, before and after. Values are %.

MetricBefore enrollmentAfter enrollment, threshold unchangedAfter enrollment, threshold recalibrated
Human FPR, MELD-eval1.091.251.14
AI TPR, MELD-eval99.5099.5399.48
Existing-family top-1 accuracy91.9091.7691.76
DeepSeek-family recall—92.4192.41

We hope these clarifications and additional results address your questions. Your comments on the data splits, the independence of the thresholds and the enrollment evaluation helped us make the evaluation protocol much clearer, and we sincerely appreciate your careful review. We would be grateful if you would consider raising your score, and we are happy to answer any further questions.

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

This paper introduces MELD, a new supervised detector for AI-generated text which formulates detection as a set of generator-family-specific binary tasks sharing a common representation of human writing. The architecture is based on the Ettin-1B encoder and builds 48-dimensional representations serving as anchors in the latent space for subsequent classification. MELD is trained on a very large corpora of about 10M examples, and the authors also introduce MELD-eval, which serves as a robustness benchmark for perturbation and held-out generators detection. Finally, the paper investigates the representation geometry introduced by MELD, as well as the evolution of AI-writing in scientific texts.

Critical strengths
  • MELD achieves strong empirical performance across all evaluation scenarios (e.g., RAID, DetectRL, and MELD-eval), suggesting the proposed method produces a strong detection signal;
  • I appreciate the idea of having family-specific objectives, which ensures both (i) avoiding hiding weaknesses on specific generator families and (ii) better covering within-family variations;
  • All the main results involve low-FPR operating points rather than simply AUROC, and this is relevant for practical deployment of MELD.
Critical weaknesses
  • Despite being a potential source of novelty, the contribution of the family-specific perspective is never properly isolated. Specifically, there are some missing yet crucial parts like (i) a comparisons with a “pooled” baseline replacing L_fem, (ii) a comparison with baselines sharing the same underlying model so as to isolate the contribution of the family-specific objective. Related to the last point, it is unclear whether all baselines have been trained on the same data, which is anoter key point.
  • Several capabilities stated at the beginning of the paper about MELD are not actually evaluated, such as attribution and span localization.
  • MELD-eval lists among the potential attacks certain interventions that are actually filtered out by the MELD preprocessing, which might introduce an unfair comparison with baseline methods.
  • Generalizability to unseen generators is actually narrower than the stated one, as most of them are within-family generators. So, results show transferability to other, unseen models from trained families.
  • The reported low-FPR robustness requires further validation, as it relies on pool-specific threshold calibration, and this does not transfer properly to other domains and text lengths.
Questions for authors
  • Can you provide the missing, stated, evidence about attribution and span localization?
  • Can you isolate the contribution of family-specific training objective by retraining MELD on the same data, with the same recipe, but using a pooling strategy?
  • Given that Tab 5 lists some datasets like RAID as training sources, can you clarify whether this has an impact on e.g., Table 1? Similar considerations for other training sources also used during evaluation.
5 comments
AuthorsOpenReview ↗

Response to Reviewer wfQo (1/4)

Thank you so much for the careful assessment and for identifying the comparisons and clarifications needed to better assess our work. We address each point below. The new results are added to the revised manuscript: Table R1 is Table 18 and Table R4 is Table 16 (Appendix J), Table R2 is Table 24 (new Appendix L), Table R3 is Table 14 (Appendix H) and Table R5 is Table 10 (Appendix C).

1. Contribution of the family-specific objective and comparability of training data (W1 and Q2).

(i) Pooled baseline replacing Lfam\mathcal{L}_{\mathrm{fam}}. We ran the requested comparison (Table R1 below). All configurations use the same backbone, training corpus and budget. In addition to the standard pooled binary classifier, we also add two ablations preserving MELD’s architecture and auxiliary objectives, while replacing the family-specific loss Lfam\mathcal{L}_{\mathrm{fam}} with pooled binary cross-entropy. One of them additionally gives equal weight to each positive generator family.

Table R1. Ablations on family-specific training. All values are TPR (%) at a pool-specific 1% FPR. The ablation subset contains the same examples used in Table 17; both DeepSeek tests contain 1,600 pairs.

ConfigurationAblation subsetMELD-evalDeepSeek-V4-FlashDeepSeek-V4-Pro
Standard pooled classifier95.6497.2794.6394.06
Pooled detection loss + auxiliary losses96.3598.4496.7596.69
Family-reweighted pooled loss + auxiliary losses97.2398.9897.3897.56
MELD98.2599.7098.0098.31

On the same subset as Table 17, MELD achieves 98.25% TPR at 1% FPR, compared with 95.64% for the standard pooled classifier, 96.35% for pooled detection with auxiliary losses, and 97.23% for its family-reweighted variant. The results show the same trend on MELD-eval and unseen tests with DeepSeek text. These comparisons further validate the contribution of the family-specific detection alongside other designs.

(ii) Baselines with the same backbone and training data.

The public baselines use their released checkpoints or official implementations and were not retrained on our corpus (§5.2). Retraining them all is not possible: GLTR, Fast-DetectGPT and Binoculars are training-free, the commercial detectors on the RAID leaderboard are closed, they do not disclose their training data or architecture details, and the trained open-source baselines were released with their own data and calibration choices (for example, ModernBERT-Detect is also trained on RAID and MAGE). In Table 1, we therefore compare them as complete detection systems, which is how they are used in practice.

Further we have designed a multitude of experiments, each targeted to evaluate and isolate the different components such as the underlying model, training data and family-specific objective:

  • Underlying model. RAID provides a public comparison of complete detectors on clean and attacked text (Table 1). Our standard pooled classifier uses the same backbone, training corpus, and budget as MELD (Table 18). We also compare separately trained one- and eight-anchor models (Table 19 and R6 in our response to Reviewer dh2q) and test mean versus top-quantile pooling on the same checkpoint (Appendix J).
  • Training data. At matched training volume, keeping complete human–AI rewrite pairs yields 96.71% TPR at 1% FPR, compared with 94.62% when the pairs are broken (Table 17). We also evaluate transfer across five public benchmarks, including HC3 and Ghostbuster, which are held out entirely during training (Table 2), and on recent generators and generator families not used for training (Table 3 and R4).
  • Family-specific objective. Two ablations keep MELD’s architecture and auxiliary losses and replace only the family-specific detection objective with pooled binary cross-entropy, with or without equal weighting across positive families (Table 18 and R1). MELD leads both on DeepSeek-V4-Flash and -Pro, Muse Glimmer, and Nemotron (§5.1 and Table 16 and R4).

We would also like to stress that part of our contribution is also the conception and creation of the data corpus: the content-matched rewrite corpus (§4), whose effect the volume-matched ablation in Appendix J separates from corpus size. In sum, we aimed for experiments that are comprehensive and as fair as possible, evaluating the overall performance of the full system while also isolating the individual contributions. We appreciate this comment and are happy to add further controlled comparisons that the reviewer finds useful.

AuthorsOpenReview ↗

Response to Reviewer wfQo (2/4)

2. Evidence for attribution and span localization (W2 and Q1).

Regarding the attribution, we would like to note that because of the page limit, the main text provided only the brief headline number (§5.1), and that the full evaluation was available in Appendix J (Table 11 of the original submission). The main finding is that on 8,400 AI validation examples, MELD identifies the correct generator family (among seven: OpenAI GPT, Claude, Gemini/Gemma, Llama, Mistral, Qwen and Other) in 91.9% of cases. On a separate test sample of 15,000 examples (2,500 per task), the rewrite task (among six: generate, polish, paraphrase, humanize, summarize and expand) is identified with 95.60% accuracy. Restricting to confident predictions raises family accuracy to 98.3% while still covering 81.4% of examples. In the revision we moved this table from Appendix J into the main text (Table 5 on page 8 of the revised manuscript, discussed in §5.1) and added the per-task breakdown, so that the attribution results appear next to the detection results.

Regarding the localization task, we agree that the manuscript only coarsely showed that MELD’s score responds to localized AI content. We showed that top-quantile pooling is designed to keep a short piece of AI text from being diluted by the surrounding human text (§3.2), and replacing it with mean pooling lowers AUROC by 1.3 points on documents containing 10% AI text (Appendix J).

Now to strengthen the localization analysis we carried out additional experiments. To evaluate the predicted spans, we evaluate MELD (with the decoder specified in Appendix A) in two settings. First, we evaluate against word-level labels on DAMASHA (Sai Teja et al., 2026), a public benchmark of mixed human/AI documents. Second, we evaluate on two held-out mixed sets: human and AI passages from arXiv interleaved at sentence boundaries, and human openings followed by an AI continuation.

Our experiments show that even without any span supervision, MELD achieves higher word-level F1 than the supervised DAMASHA model on two of the three sets (arXiv interleaved and continuation). After fine-tuning, MELD exceeds the supervised DAMASHA model on all three sets. The FPR of MELD is also lower.

Table R2. Span localization on the same documents. Word-level F1 (%) of the AI class on documents containing both human and AI text. Human-word FPR is the percentage of words in fully human documents that are marked as AI. Binoculars is scored on 64-word windows.

MethodDAMASHAarXiv interleavedContinuationHuman-word FPR (lower is better)
DAMASHA model (supervised)94.7258.4577.538.47
Binoculars39.983.9416.380.99
MELD (zero-shot)73.4175.8994.870.06
MELD (fine-tuned on DAMASHA data)98.7174.0191.430.34

We appreciate the comment and now report the main localization result in §5.1 of the main text (page 6), with the full analysis and Table R2 in the new Appendix L (Table 24).

3. Fairness of the formatting preprocessing (W3).

MELD’s format normalization is part of its core design, which effectively absorbs homoglyph and zero-width attacks. Yet we agree that the contribution of this preprocessing could be further evaluated. We therefore carry out additional experiments where we apply the same format normalization to MELD and other baselines (Table R3). Here we compare MELD to ModernBERT-Detect, the strongest baseline in Table 3, and Fast-DetectGPT, the strongest training-free detector in Table 3. For each detector, we calibrate the threshold separately with and without normalization, using human scores from the corresponding condition to set the 1% FPR operating point.

Table R3. Format normalization comparison on MELD-eval. All entries are TPR (%) at a pool-specific 1% FPR before and after.

DetectorGPT-5.4-MiniGemini-3-FlashClaude-Haiku-4.5Qwen-3.6-PlusOverall
Fast-DetectGPT9.2 → 13.619.8 → 27.314.3 → 19.924.7 → 33.317.0 → 23.5
ModernBERT-Detect97.2 → 98.693.0 → 97.598.2 → 98.993.7 → 97.895.5 → 98.2
MELD98.9 → 99.397.6 → 99.898.8 → 99.797.8 → 100.098.3 → 99.7

With this experiment we show that format normalization, which is part of our proposed method, brings consistent performance improvements to baseline detectors, too, while MELD remains ahead for every generator. This shows that preprocessing accounts for only part of the original gap. The contribution of the family-specific objective is assessed separately in Table R1.

AuthorsOpenReview ↗

Response to Reviewer wfQo (3/4)

4. Generalization to unseen generators (W4).

We would like to acknowledge that the submission evaluated only one generator family that is absent from training: DeepSeek-V4-Flash and DeepSeek-V4-Pro (§5.1, with details in Appendices G and J). We originally did not include more because, at the time of our main experiments, few recent open-weight models were both representative and clearly outside our seven training families; many are fine-tuned or distilled from models of these families or share much of their pretraining data, so they would not have been a clean held-out test.

However, the field is very dynamic and new model families have been released since. We now provide results on more recent LLM families not used in training in Table R4, namely Muse Glimmer 30B (Meta) and Nemotron 3 Nano 30B (NVIDIA). They follow the Appendix G protocol on the held-out scientific-text source pool, with 1,600 human–AI pairs per model and the same settings.

Table R4. Extended evaluation on held-out families. All entries are TPR (%) at a pool-specific 1% FPR.

GeneratorBinocularsStandard pooled classifierFamily-reweighted pooled loss + auxiliary lossesMELD
DeepSeek-V4-Flash11.5694.6397.3898.00
DeepSeek-V4-Pro6.4494.0697.5698.31
Muse Glimmer 30B7.1981.3185.3888.81
Nemotron 3 Nano 30B19.4494.5695.3196.63

MELD has the highest observed TPR on each (unseen) generator. It reaches 98.00% on DeepSeek-V4-Flash, 98.31% on DeepSeek-V4-Pro, 88.81% on Muse Glimmer, and 96.63% on Nemotron. On Muse Glimmer, MELD exceeds the family-reweighted ablation by about 3.4 percentage points. Appendix J, Table 16 provides the full comparison.

5. Pool-specific thresholds and calibration (W5).

TPR at 1% FPR compares detectors at a matched operating point (Appendix C). It does not assume that one threshold fits every domain. The manuscript also already shows results with a fixed threshold (§5.1). To further deepen the analysis we now show additional results comparing MELD with and without pool-specific calibration on the three largest public benchmarks of Table 2 and on short texts (Table R5).

Table R5. MELD with and without pool-specific calibration. The pool-calibrated threshold is set on one half of the pool’s human texts and evaluated on the other half. Values are %.

PoolMELD FPR, fixedMELD TPR, fixedMELD FPR, pool-calibratedMELD TPR, pool-calibratedModernBERT-Detect TPR, pool-calibrated
HC31.6799.201.0098.4674.13
MAGE1.8491.991.0089.5587.02
DetectRL1.6397.691.0196.9217.36
Short texts (≤192 words)5.4793.601.0789.9381.74

Even with a fixed threshold, MELD keeps the FPR at 1.6–1.8% on all three benchmarks. Pool-specific calibration brings it to 1.0% at a cost of only 0.7–2.4 TPR points, and MELD stays ahead of ModernBERT-Detect on every benchmark under the same protocol. For short texts, the same calibration gives 1.1% FPR at 89.9% TPR. In practice, most detectors would need calibration on human text from the target population to reach a low FPR (§3.4, §5.5).

Reviewer_J5SJOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes MELD, a multi-task framework for detecting AI-generated text. Instead of using a single pooled AI-versus-human objective, MELD trains family-specific detection tasks with a shared representation of human writing. The model uses multiple human style anchors and generator-family centers to support AI detection, generator attribution, rewrite-task prediction, and token-level localization. The authors also construct a large corpus of paired human-written and AI-rewritten scientific passages and introduce format normalization to improve robustness.

The paper evaluates MELD on several public benchmarks and a newly constructed MELD-eval benchmark covering four recent LLMs and multiple attacks. MELD ranks first among open-source detectors on RAID and achieves 99.7% TPR at 1% FPR on MELD-eval. The authors further provide unseen-family experiments, ablation studies, and representation analyses. Finally, they apply MELD to 5.9 million scientific texts to study changes in AI-like writing patterns in academic papers and peer reviews over time.

Critical strengths
  1. Strong and consistent empirical performance across diverse benchmarks.

I find the experimental results particularly impressive. MELD achieves the best performance among open-source detectors on RAID, including under adversarial attacks, and performs very well across multiple held-out benchmarks. In particular, it achieves 99.7% TPR at 1% FPR on MELD-eval, compared with 95.5% for the strongest baseline. I also appreciate the evaluation on the completely unseen DeepSeek family, where MELD still achieves around 98% TPR at 1% FPR. These results provide convincing evidence that the proposed framework is effective beyond the generators seen during training. The emphasis on low false-positive rates is especially valuable given the potential consequences of incorrectly labeling human-written text as AI-generated.

  1. Strong data construction and comprehensive evaluation of robustness.

Another strength is the careful design of the training data and evaluation methodology. I particularly like the construction of paired human-written and AI-rewritten scientific passages, which helps control for differences in topic and content when learning writing-style features. The authors also provide a useful ablation comparing unpaired data with a volume-matched subset that retains the pairing structure, showing that the matched data contribute beyond simply increasing the training set size. Beyond standard detection benchmarks, I appreciate the explicit evaluation of sensitivity to AI rewriting versus formatting changes, which addresses an important practical issue for AI-text detectors. The large-scale analysis of 5.9 million scientific texts is also an interesting application that demonstrates the potential usefulness of the framework beyond benchmark evaluation. Together, these efforts make the empirical contribution of the paper quite substantial.

Critical weaknesses

None.

Questions for authors

Could the authors provide a comparison with a standard pooled AI-versus-human classifier trained using the same backbone, training data, and computational budget? Since the family-specific detection objective is a central part of MELD, this comparison would help clarify how much of the improvement comes from the proposed learning objective rather than the data construction or other components.

2 comments
AuthorsOpenReview ↗

Response to Reviewer J5SJ

Thank you for the positive assessment of our methods and results. Providing the community with a free, strong, open-source detector that can also support meaningful studies of LLM use is our main motivation for this work.

As suggested, we compare MELD to a standard pooled AI-versus-human classifier, with a single binary detection head trained with binary cross-entropy using the same backbone, training corpus and computational budget. As an additional experiment to show that the gain comes from the family-specific loss itself and not from MELD’s architecture or auxiliary losses, we also include two controls that keep both but replace the family-specific detection loss with pooled binary cross-entropy, using either ordinary example weighting or equal positive-family weighting.

Table R1. Ablations on family-specific training. All values are TPR (%) at a pool-specific 1% FPR. The ablation subset contains the same examples used in Table 17; both DeepSeek tests contain 1,600 pairs.

ConfigurationAblation subsetMELD-evalDeepSeek-V4-FlashDeepSeek-V4-Pro
Standard pooled classifier95.6497.2794.6394.06
Pooled detection loss + auxiliary losses96.3598.4496.7596.69
Family-reweighted pooled loss + auxiliary losses97.2398.9897.3897.56
MELD98.2599.7098.0098.31

In this comparison, MELD improves TPR over the standard pooled classifier by 2.61 percentage points on the ablation subset, 2.43 points on MELD-eval, about 3.4 points on DeepSeek-V4-Flash, and 4.25 points on DeepSeek-V4-Pro. Its advantage over the two pooled-loss controls also indicates the contribution of family-specific detection, beyond the shared backbone, training data, auxiliary supervision, or positive-family weighting.

We appreciate this comment and have included this comparison and its experimental details in the revised manuscript (§5.1 and Appendix J, Table 18). We hope they resolve your question, and we would be grateful if you would consider raising your score.

Thank you for the detailed response and the additional experiments. I appreciate that the authors not only included the standard pooled binary classifier I requested, but also provided two additional controlled comparisons to isolate the contribution of the family-specific detection objective.

The results are convincing, particularly because MELD consistently outperforms the pooled alternatives across the ablation subset, MELD-eval, and the unseen DeepSeek family. These experiments directly address my main question and strengthen the empirical justification for the proposed approach.

I was already positive about the paper's strong empirical performance, careful data construction, and comprehensive evaluation. With this additional evidence, my main concern has been resolved, and I have increased my rating accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.