acceptodds
Under review as a conference paper at ICLR 2027

Who's Behind the Text? Robust Model Authorship Attribution

Abstract

As large language models (LLMs) are increasingly used to generate text, their misuse has become a growing societal problem, with serious implications for academic misconduct, model distillation, and cyberattacks. For these reasons, Model Authorship Attribution (MAA), which identifies the source model of a given text, is important for promoting the responsible use of LLMs and improving AI safety. In practice, attribution systems must remain reliable beyond the distributions observed during training. They may encounter texts from unseen domains and writing settings, as well as generations from future versions of model families that were not available when the system was trained. However, existing MAA methods often exhibit substantial performance degradation under such distribution shifts. To address this challenge, we propose Authorship Tracing via Likelihood-based Attribution of Sources (ATLAS), which formulates MAA through source-conditioned text modeling rather than direct source prediction. ATLAS learns how text from each source family is generated and attributes a given text by comparing its likelihood under each learned source condition. We aggregate this likelihood evidence at the token level with clipping so that no single token dominates the decision. We further introduce MAABench, a novel and broad MAA benchmark that can evaluate more practical OOD settings such as domain and version. Our experiments show that ATLAS consistently outperforms existing baselines across MAABench and OpenTuringBench. In particular, when both the generation domain and the model version change on MAABench, ATLAS achieves 74.30% accuracy, 8.15pp higher than the strongest baseline. Beyond MAA, ATLAS can also help identify distillation source models by measuring which source model a target LLM most closely resembles. Our extended experiments support this claim by showing that ATLAS recovers the known teacher of every distilled model we evaluate.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model authorship attribution detection, Generative model, Likelihood, Benchmark

Reviews

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

This paper proposes a model-authorship attribution method that replaces direct source classification with source-conditioned language modeling. Each source family is represented using a learned prefix plus source-specific LoRA adaptation, and attribution is based on relative token likelihoods with clipping to prevent a few extreme tokens from dominating the decision. The paper also introduces a benchmark.

Critical strengths
  1. MAABench targets a weakness of existing attribution work: robustness to both domain shift and model-version evolution. The benchmark is reasonably broad. This makes the evaluation considerably more relevant to deployment than standard ID-only MAA experiments.
  2. The proposed ATLAS improves over the strongest baseline. It is also data-efficient.
  3. The token-level clipping mechanism has a clear robustness motivation rather than being an arbitrary architectural addition. The ablations support that motivation. The additional teacher-attribution experiment also suggests that the learned signal captures persistent family-level characteristics rather than only a benchmark-specific decision boundary.
Critical weaknesses
  1. ATLAS assumes that every test text comes from one of the predefined candidate families. This is a significant limitation for real deployment, where entirely unseen generators will continually appear. Attribution to previously unseen source families, particularly without retraining, remains unresolved. An open-set evaluation with an explicit unknown class or calibrated rejection criterion would strengthen the work.
  2. In the reported Domain shift, both the domain changes and the system prompting changes: held-out Patent/Stack Exchange domains use persona-based system prompts rather than the default training prompt. Thus, the experiment demonstrates robustness to a useful compound shift, but it is difficult to determine how much comes from domain versus prompt/style variation. Moreover, generations are relatively controlled, shared prompts, which leaves robustness to paraphrasing, human editing, arbitrary decoding configurations, very short texts, and adversarial obfuscation less established.
  3. Reported number comes from a single training run with a fixed seed. Given that ATLAS and several learned baselines involve stochastic optimization, reporting mean ± standard deviation across multiple seeds would make especially the smaller performance differences much more convincing.
  4. Code and data are not realased yet and thus reproducibility is limited.
Questions for authors

Please refer to the weaknesses

5 comments
AuthorsOpenReview ↗

Response to Reviewer tMGY (1/2)

Thank you for the careful reading of our paper and the constructive comments. Our responses to each weakness follow.


W1. Open-set attribution: ATLAS assumes that every test text comes from a predefined candidate family. Can it reject texts from unseen families with a calibrated rejection criterion?

Thank you for this suggestion. ATLAS can reject texts from unseen families: it rejects a held-out family with an average AUROC of 90.3 on ID data, using a score that its likelihood formulation already provides (new, Appendix C.3, Table 29).

The rejection score is the gain of the best source condition over the frozen backbone, max⁡ks~k(y)\max_k \tilde{s}_k(y), the quantity behind the rejection rate in Section 6. In a leave-one-family-out evaluation, ATLAS is trained on four families (10k texts per source), the fifth is the unknown class, and one threshold is calibrated on ID validation to accept 95% of known texts. Unknown families are separated well on average, best for Mistral and Llama and hardest for GPT, while known texts are falsely rejected at about the calibrated 5% rate:

Table R1. Leave-one-family-out open-set rejection.

Held-out familyAUROCUnknown rejectedKnown falsely rejected
GPT79.8827.345.18
Claude85.5044.975.16
Gemini91.1551.135.12
Mistral98.0190.195.71
Llama97.0784.385.42
Avg.90.3259.605.32

The same score already rejects models outside the candidate families: in Section 6, Falcon3-7B-Instruct and Granite-3.0-8B-Instruct are rejected for 68.2% and 74.6% of their texts. As stated in our conclusion, extending attribution to unseen families, particularly without full retraining, and rejection under domain and version shift are directions we leave to future work.


W2-1. Compound Domain shift: the held-out Patent and Stack Exchange domains also use persona-based system prompts. How much of the shift comes from the domain, and how much from the prompt?

For ATLAS, the whole drop comes from the domain change. For the baselines, the persona prompt causes a further drop on top of it.

The persona prompt reflects how each domain is written in practice, not an artificial perturbation: patent abstracts are generated as "a patent specialist" and Stack Exchange questions as "a member of an online Q&A community" (Appendix A.3). OpenTuringBench [2] likewise generates its out-of-domain essays with a different system prompt from its training domain (news). To separate the two factors, we added a domain-only control (new, Appendix C.1): texts generated from the same metadata by the same models, but with the default training prompt (means over three random seeds).

Table R2. Accuracy (%) with the domain and the prompt changed separately (mean over three random seeds).

SettingDomainPromptATLASBest baseline
IDsamedefault99.7299.58
Domain-only control (new)newdefault97.9297.19
Domain shift (paper)newpersona97.9795.31
  • Domain change: ATLAS drops by 1.8pp (99.72% to 97.92%), and the best baseline drops by 2.4pp.
  • Persona prompt: ATLAS stays almost unchanged (97.92% to 97.97%), while every baseline loses at least 1.7pp.

[2] La Cava, Lucio, and Andrea Tagarelli. "Openturingbench: An open-model-based benchmark and framework for machine-generated text detection and attribution." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

AuthorsOpenReview ↗

Response to Reviewer tMGY (2/2)

W2-2. Robustness beyond controlled generations: how robust is ATLAS to paraphrasing, human editing, arbitrary decoding configurations, very short texts and adversarial obfuscation?

ATLAS outperforms the strongest baseline in all five settings (Table 3 and new Appendix C.2). In each table, "Best baseline" is the strongest baseline in each row or column, and all models are trained on clean, full texts.

Paraphrasing, human editing and decoding (OpenTuringBench [2], Table 3, and RAID [1] reconstructed as 5-family attribution (GPT, Llama, Mistral, MPT, Cohere), new, Table 28). E3 lets each model rewrite its own text, E4 lets a human and a model write parts of the same text, and E1 changes the sampling temperature. ATLAS is best in every setting, with the largest gains under human editing and repetition penalty:

Table R3. Accuracy (%) under paraphrasing, human editing and decoding changes.

SettingATLASBest baseline
Paraphrasing (OpenTuringBench E3)98.7995.99
Human editing (OpenTuringBench E4)41.9232.36
Sampling temperature (OpenTuringBench E1)89.7184.43
Greedy decoding (RAID)72.8972.39
Repetition penalty (RAID)85.8866.37

Very short texts (new, Tables 26 and 27). Models are evaluated on only 16 or 32 words of each test text. To check that ATLAS does not rely on the opening of a text, we also took 32 consecutive words from the middle of each text. ATLAS stays far more accurate than the baselines at every length, including the middle words:

Table R4. Accuracy (%) on very short texts.

WordsATLAS (ID)Best baseline (ID)ATLAS (Version)Best baseline (Version)
First 1675.8226.3652.9422.81
First 3290.2636.3664.7428.77
Middle 3274.6736.4952.7129.35

Adversarial obfuscation (new, Table 25). We applied ten perturbations to the MAABench ID test set, following the existing attacks (i.e., the attack code and default attack rates of RAID [1]). ATLAS is best under nine of the ten attacks, with the largest gains under homoglyphs and zero-width spaces:

Table R5. Accuracy (%) under perturbation attacks on MAABench ID.

AttackATLASBest baseline
Homoglyph69.3658.86
Zero-width space45.0425.50
Whitespace99.1999.20
Upper/lower case99.6399.19
Insert paragraphs99.6699.42
Article deletion99.6899.02
Number99.7399.48
Alternative spelling99.7899.51
Misspelling99.7899.40
Synonym90.6687.40
Avg. 10 attacks90.2586.30

[1] Dugan, Liam, et al. "Raid: A shared benchmark for robust evaluation of machine-generated text detectors." Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

[2] La Cava, Lucio, and Andrea Tagarelli. "Openturingbench: An open-model-based benchmark and framework for machine-generated text detection and attribution." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.


W3. Single seed: do the results, especially the smaller differences, hold as mean and standard deviation over multiple seeds?

The results hold over three random seeds: ATLAS still outperforms every baseline on all four splits, with large margins on Version shift (12.3pp) and Joint shift (9.0pp) that far exceed the seed variation. It also has the lowest standard deviation on every OOD split. We trained ATLAS and all four baselines with two additional random seeds. Table 2 now reports the three-seed means, with standard deviations in Appendix B.5 (Table 19):

Table R6. Accuracy (%) over three seeds (mean ± std).

MethodIDDomainVersionJoint
ATLAS99.72 ± 0.0597.97 ± 0.1784.30 ± 0.0874.27 ± 0.25
DeTeCtive99.51 ± 0.0594.69 ± 0.6070.59 ± 0.5863.34 ± 0.50
OTBDetector99.49 ± 0.0794.88 ± 0.6370.19 ± 1.3063.47 ± 2.29
DA-MTL99.58 ± 0.0295.31 ± 0.2771.08 ± 0.5664.99 ± 1.16
Idiosyncrasies99.57 ± 0.1095.03 ± 1.1871.97 ± 0.8365.27 ± 0.38

W4. Reproducibility: the code and data are not released yet.

We have uploaded the code and data as anonymized supplementary material. The code includes the training, scoring and evaluation scripts for ATLAS. The data includes the MAABench generated texts, seed identifiers and corpus revisions. Human-written texts are not redistributed, as described in the Ethics Statement, but can be recovered with the provided content-addressed identifiers.

We thank the reviewer again for the constructive comments, which helped us strengthen the paper.

Dear Reviewer tMGY,

We hope this message finds you well. We are writing to kindly follow up on our response to your comments, which includes three-seed results with standard deviations, open-set rejection, a domain-only control for Domain Shift, additional robustness results, and the released code and data. We want to make sure that our explanations adequately address the concerns you raised.

If any part of our response remains unclear, we would be very happy to provide further details. As the discussion period is coming to a close, we would greatly appreciate it if you could take a look at your convenience.

Thank you again for your time and thoughtful review.

Best regards, Authors

Reviewer_ShuuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on Model Authorship Attribution (MAA), namely the problem of identifying the specific source model (e.g., Large Language Model) that generated a given text. This is a widely studied area of machine-generated text detection and it is known to suffer when the deployment involves patterns that differ from the ones seeing during training (e.g., new domains, unseen generators). To this aim, this work proposes Authorship Tracing via Likelihood-based Attribution of Sources (ATLAS), a new framework that instead of modeling the MAA problem as a source-prediction task, shapes it as a source-conditioned text modeling one. Beyond this, the authors also propose a new benchmark dubbed MAABench, which evaluates OOD settings like domain and version variations. ATLAS is hence evaluated on the newly proposed benchmark and other authorship attribution benchmarks and compared with existing detection/attribution approaches.

Critical strengths
  • (Model) Authorship Attribution is a key issue for many applications involving machine-generated text (accountability, copyright infringements), and new methods that can also better generalize out-of-domain as well as to unseen generators (yet from the same family) are always useful, if properly working;
  • The idea of moving toward a generative formulation rather than a direct discriminative task might yield higher robustness to OOD settings, as the model itself might be less prone to learn shortcuts yet exploits a better likelihood modeling;
  • The reported empirical improvements under distribution shifts seem (see weaknesses) strong and this matter a lot for practical deployment, also considering the training efficiency of ATLAS.
Critical weaknesses
  • Overall, novelty is modest for both methodology and benchmark construction as, for instance, (i) LLMDet already exploits source-specific likelihoods (and it is never considered as a competing method), and (ii) domain and source changes as distribution shifts have already been considered in well-known benchmarks like RAID.
  • In MAABench, in-domain data is generated using the default system prompt, whereas the OOD data is generated via persona-based prompting. This introduces a potential confounder, i.e., a persona-related signal, that might explain the performance drop observed for OOD.
  • Source-conditioned inference comes with a non-negligible cost: ATLAS requires evaluating every given text under every candidate source, and this does not scale well if we account all existing families of models (even worse in the case of precise model checkpoints). The scalability to deployment settings remains hence underexplored in my opinion.
  • The experimental setup is a bit noisy. Encoder baselines run on LLM2Vec, despite their original implementations leverage RoBERTa-based and similar approaches, and all of them share the hyperparameters with ATLAS (Table 12), thus yielding potential under-optimization wrt their original implementations. Under this “unfair” setting, it is difficult to discern the proper informativeness of the source-conditioned signal used by ATLAS over existing approaches.
Questions for authors
  • ATLAS might be Claude-biased, as shown in Figure 4 under domain shifts, where the latter tends to attract errors. This is important for explaining the trends in Figure 3, and warrants further investigation. Is this a bias in the framework? If so, which is its source? Can it be fixed?
  • What is the robustness of ATLAS under adversarial attacks (e.g., those in RAID)? This would matter a lot for practical deployment.
  • Related to the previous point, how does ATLAS perform under (very) short texts?
  • Why using Qwen2.5 as a backbone model if Table 6 suggests Qwen3 would be better?
  • How do you separate and discern stylistic similarity from genuine teacher-specific transfer in the distillation analysis?
  • Last but not least, what happens to results when all baselines are implemented via their original settings? This point would matter a lot in showing that the signal used by this approach is more informative than competing ones.
6 comments
AuthorsOpenReview ↗

Response to Reviewer Shuu (1/5)

We thank the reviewer for the valuable comments. We address the weaknesses and questions below.


W1. Novelty: LLMDet already exploits source-specific likelihoods but is not compared, and RAID already considers domain and source shifts. What is new in ATLAS and MAABench?

Thank you for pointing out the relevant work. Unlike LLMDet, ATLAS can attribute texts to closed-weight models such as GPT, Claude and Gemini. Unlike RAID, MAABench is built for model authorship attribution, not for detection. In summary:

  1. LLMDet cannot be applied to closed-weight models. It needs white-box access, since it builds a dictionary for each source by querying that model's next-token probabilities. ATLAS learns a source-conditioned model of each family from text samples alone, so it applies to any model whose outputs can be collected. On OpenTuringBench, where all sources are open, ATLAS outperforms LLMDet on every split.
  2. The primary goal of MAABench is authorship attribution. It tests which model family wrote a text, including newer versions of that family. RAID, in contrast, is a machine-generated text (MGT) detection benchmark that only asks whether a text is human- or machine-written, and even when reconstructed for attribution, it cannot test newer versions of each family.

The detailed analysis is described below.

1. LLMDet. LLMDet [3] cannot be applied to MAABench (Table 2), since the most widely used families, GPT, Claude and Gemini [4], are closed-weight and form three of its five families. We therefore compared the two on OpenTuringBench [2], whose seven source models are all open, following the original LLMDet setup (salient n-grams from 16,000 texts per source, next-token probabilities from the (n-1)-token prefix, top 2,000 for 2-grams and top 100 for 3- and 4-grams, and a LightGBM classifier on held-out texts). LLMDet is now in Table 3, and ATLAS outperforms it by a wide margin on every split:

Table R1. Accuracy (%) of LLMDet and ATLAS on OpenTuringBench.

OpenTuringBenchIDE1E2E3E4OOD
LLMDet58.0337.1855.7553.598.4023.86
ATLAS99.8089.7199.7098.7941.9295.71

2. RAID. RAID [1] includes several generators, domains and attacks, but only to test whether a detector still flags machine text, not which model wrote it. Even when reconstructed for attribution (Q2), it leaves two gaps: only GPT has several releases, so version shift cannot be tested for the other families, and several of the most widely used families today are missing. MAABench is designed for authorship attribution: each family has three newer versions held out from training, so domain shift and version shift in attribution can be measured separately and jointly.

[1] Dugan, Liam, et al. "Raid: A shared benchmark for robust evaluation of machine-generated text detectors." Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

[2] La Cava, Lucio, and Andrea Tagarelli. "Openturingbench: An open-model-based benchmark and framework for machine-generated text detection and attribution." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

[3] Wu, Kangxi, et al. "Llmdet: A third party large language models generated text detection tool." Findings of the association for computational linguistics: emnlp 2023. 2023.

[4] Sajadieh, Sha, et al. "Artificial Intelligence Index Report 2026." arXiv preprint arXiv:2606.15708 (2026).

AuthorsOpenReview ↗

Response to Reviewer Shuu (2/5)

W2. Confounder: in-domain data uses the default system prompt, while OOD data uses persona-based prompts. Could a persona-related signal explain the OOD drop?

Thank you for pointing out this potential confounder. The persona prompt is not the main cause of the OOD drop. A new domain-only control (Appendix C.1) shows that the persona prompt leaves ATLAS unchanged but lowers every baseline by at least 1.7pp, and the largest drop occurs under Version shift, which uses no persona prompt. Below are the details.

The persona prompt reflects how each domain is written in practice, not an artificial perturbation: patent abstracts are generated as "a patent specialist" and Stack Exchange questions as "a member of an online Q&A community" (Appendix A.3). OpenTuringBench [2] likewise generates its out-of-domain essays with a different system prompt from its training domain (news). The domain-only control uses texts generated from the same metadata by the same models, but with the default prompt (means over three random seeds):

Table R2. Accuracy (%) with the domain and the prompt changed separately (mean over three random seeds).

SettingDomainPromptVersionATLASBest baseline
IDsamedefaultsame99.7299.58
Domain-only control (new)newdefaultsame97.9297.19
Domain shift (paper)newpersonasame97.9795.31
Version shift (paper)samedefaultnew84.3071.97

Each row changes one factor at a time. From ID to the domain-only control, only the domain changes, and ATLAS stays ahead of the best baseline.

[2] La Cava, Lucio, and Andrea Tagarelli. "Openturingbench: An open-model-based benchmark and framework for machine-generated text detection and attribution." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.


W3. Scalability: ATLAS evaluates every text under every candidate source. How does its cost scale with many model families, or with individual checkpoints?

ATLAS offers a better accuracy-cost trade-off than the baselines: at a comparable inference cost, it is more accurate on Version and Joint shift (new, Appendix C.4, Figure 6).

We compare ATLAS with Qwen2.5 backbones from 0.5B to 7B against the four baselines (DeTeCtive, OTBDetector, DA-MTL and Idiosyncrasies), which all run on LLM2Vec-Qwen2.5-7B and therefore share the same cost. Cost is measured in TFLOP per text of 512 tokens. Accuracy (%) is from a single run with one random seed on the two hardest MAABench splits: Version shift (newer versions of each family than those seen in training) and Joint shift (newer versions in unseen domains).

Table R3. Cost per text and accuracy (%).

ModelCost (TFLOP): Training/InferenceAccuracy: Version shift (%)Accuracy: Joint shift (%)
Best baseline13.7/6.871.7365.83
ATLAS, Qwen2.5-0.5B (10k texts/source)1.2/2.778.3364.98
ATLAS, Qwen2.5-1.5B (10k texts/source)3.6/8.281.0070.28
ATLAS, Qwen2.5-7B (full data)15.4/36.884.3674.30

With a 1.5B backbone, ATLAS has an inference cost comparable to the baselines and is 9.3pp (Version) and 4.5pp (Joint) more accurate.

As the number of candidates grows, three properties keep ATLAS practical. Training uses only the true-source condition, so its cost does not grow with the number of candidates. Adding a family adds one prefix embedding and one up-projection, under 0.1% of the backbone parameters. At inference, a smaller backbone keeps the cost low, and for many open models or individual checkpoints, hierarchical scoring (family first, then version or checkpoint) can avoid scoring every candidate, which we leave to future work.

AuthorsOpenReview ↗

Response to Reviewer Shuu (3/5)

W4 & Q6. Baseline setup: the encoder baselines run on LLM2Vec with the hyperparameters of ATLAS, not in their original RoBERTa-based settings. Are they under-optimized, and how do they perform in their original settings?

Thank you for raising this point. The baselines are not under-optimized, and in their original settings they also stay far below ATLAS (by at least 10.6pp on Version and 16.4pp on Joint shift). We run them on LLM2Vec-Qwen2.5-7B, which is also the original encoder of Idiosyncrasies, because it matches the 7B scale of ATLAS, so the comparison isolates the method. Their original encoders are much smaller, so a comparison in the original settings would change the method and the model size at once.

Original settings (new, Appendix B.2). As a reference, we trained DeTeCtive with SimCSE-RoBERTa-base, OTBDetector with Longformer-base and DA-MTL with XLM-R-base, each with its published hyperparameters, loss and inference rule.

Classifier on the ATLAS backbone (new). To test the value of the source-conditioned signal, we trained a classifier that keeps every component of ATLAS (Qwen2.5-7B, LoRA, data and schedule) but predicts the source directly with a linear classification head trained with cross-entropy (CE-head). CE-head thus differs from ATLAS only in this signal.

Table R4. Accuracy (%) in original settings and of CE-head.

MethodIDDomainVersionJoint
ATLAS99.7798.1184.3674.30
DeTeCtive, original99.3490.8273.7957.91
DeTeCtive, LLM2Vec99.4995.3870.0862.92
OTBDetector, original97.5188.8172.0956.53
OTBDetector, LLM2Vec99.4295.4471.2765.02
DA-MTL, original97.4385.1571.2955.73
DA-MTL, LLM2Vec99.5995.5471.7365.83
Idiosyncrasies (original = LLM2Vec)99.5294.3971.0265.66
CE-head (same backbone)99.6095.2771.1966.36

CE-head is 13.2pp (Version) and 7.9pp (Joint) below ATLAS, so the gain comes from the generative formulation, not from the backbone.


Q1. Claude bias: under domain shift, Claude tends to attract errors (Figure 4). Is this a bias of the framework? If so, what is its source, and can it be fixed?

We found no sign that ATLAS favors Claude by default. In Figure 3(b), MiniMax, Kimi and GLM are attributed to Claude, so we ran an additional experiment on five recent models from other families. All five are attributed mainly to GPT rather than Claude. Below are the details.

Unseen families go to GPT (new, Appendix D, Table 31). Like DeepSeek-V4-Pro in Figure 3 (90.6% GPT), all five go mainly to GPT:

Table R5. Share (%) of each model's texts attributed to each family (Section 6).

ModelGPTClaudeGeminiRejected
Grok-4.792.27.80.10.2
K-EXAONE-2.072.626.11.41.5
Motif-371.827.30.90.8
A.X-K262.637.20.20.2
Solar-Open259.240.00.81.0
Reviewer_aoxvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper examines model authorship attribution (i.e., identifying the source model of a given text). The contributions of this paper are as follows. First, the paper introduces a method (ATLAS), which learns a source-conditioned model of the text distribution for different model families. Second, the paper introduces MAABench to evaluate model authorship attribution under distribution shift (including shifts in generation domain and source-model version). Finally, the paper draws empirical insights from the proposed method and benchmark.

Critical strengths

The strengths include:

  1. This paper tackles a very relevant issue and makes thoughtful contributions on both the method and evaluation fronts. The experiments demonstrate the strengths of the proposed method, and the benchmark includes numerous baselines (e.g., other methods) for comparison.
  2. The paper is well written and easy to follow.
Critical weaknesses

The weaknesses of this paper include:

  1. Cost analysis: The paper would be stronger if it included a cost analysis (e.g., computational cost in FLOPs) of the proposed method compared with similar methods. Given that the authors have already shown that the proposed method performs well on MAABench, its efficiency should also be justified.
  2. Saturation on the benchmark: As the paper mentions, the performance of various methods (including the proposed method) is already saturated on the existing benchmark. It would be interesting to highlight some more challenging cases. In addition, it would also be good to see how the proposed method performs against other methods on other model authorship attribution benchmarks.
Questions for authors

None

5 comments
AuthorsOpenReview ↗

Response to Reviewer aoxv (1/2)

We thank the reviewer for the helpful comments on cost and evaluation. We address both points below.


W1. Cost analysis: what is the computational cost of ATLAS in FLOPs compared with similar methods, and is its efficiency justified?

Following your suggestion, we newly added a cost analysis (Appendix C.4). ATLAS achieves a better accuracy-cost trade-off than the baselines: at a comparable inference cost, it is more accurate on both Version and Joint shift (Figure 6).

Specifically, we compare ATLAS with Qwen2.5 backbones from 0.5B to 7B against the four baselines (DeTeCtive, OTBDetector, DA-MTL and Idiosyncrasies), which all run on LLM2Vec-Qwen2.5-7B and therefore share the same cost. Cost is measured in TFLOP per text of 512 tokens. Accuracy (%) is from a single run with one random seed on the two hardest MAABench splits: Version shift (newer versions of each family than those seen in training) and Joint shift (newer versions in unseen domains).

Table R1. Cost per text and accuracy (%) of ATLAS and the best baseline.

ModelCost (TFLOP): Training/InferenceAccuracy: Version shift (%)Accuracy: Joint shift (%)
Best baseline13.7/6.871.7365.83
ATLAS, Qwen2.5-0.5B (10k texts/source)1.2/2.778.3364.98
ATLAS, Qwen2.5-1.5B (10k texts/source)3.6/8.281.0070.28
ATLAS, Qwen2.5-7B (full data)15.4/36.884.3674.30
  • With a 1.5B backbone, ATLAS needs inference compute comparable to the baselines and is 9.3pp (Version) and 4.5pp (Joint) more accurate, even with only 10k training texts per source.
  • With a 0.5B backbone, ATLAS needs less inference compute than the baselines and still leads on Version shift.
  • With the 7B backbone, inference needs 5.4x the compute of the baselines, since ATLAS scores all five source conditions, while training needs only 1.13x, since ATLAS learns only the true-source condition.

W2. Saturation: can you highlight more challenging cases, and how does ATLAS compare with other methods on other MAA benchmarks?

We would like to clarify that the saturated results are in-domain results, while we also evaluate ATLAS under various OOD settings, and these are examples of the harder cases. As shown in the paper and in the new experiments below, ATLAS outperforms the strongest baseline on all OOD settings. We explain these settings in more detail below.

  1. MAABench OOD splits: Domain shift (unseen domains, with a system prompt of a different style), Version shift (newer versions of each family than those seen in training) and Joint shift (both). ATLAS leads most under Version shift (+12.3pp) and Joint shift (+9.0pp) (means over three random seeds). In each table below, "Best baseline" is the strongest baseline in each column or row:

Table R2. Accuracy (%) on MAABench (mean over three random seeds).

MAABenchIDDomainVersionJoint
ATLAS99.7297.9784.3074.27
Best baseline99.5895.3171.9765.27
  1. OpenTuringBench [2] (Table 3), a second MAA benchmark. E1 changes the sampling temperature, E2 uses about 70B versions of the training models, E3 lets each model rewrite its own text, E4 mixes human and machine text, and OOD uses essays generated with a different system prompt from the training domain (news). ATLAS is best in every setting, with the largest gain on the OOD split:

Table R3. Accuracy (%) on OpenTuringBench.

OpenTuringBenchE1E2E3E4OOD
ATLAS89.7199.7098.7941.9295.71
Best baseline84.4399.4895.9932.3662.19
AuthorsOpenReview ↗

Response to Reviewer aoxv (2/2)

  1. Perturbed texts (new, Appendix C.2, Table 25). We applied ten perturbations to the MAABench ID test set, following the existing attacks (i.e., the attack code and default attack rates of RAID [1]). All models are trained on clean texts. In this new experiment, ATLAS is more robust than the baselines: it is best under nine of the ten attacks, with the largest gains under homoglyphs and zero-width spaces:

Table R4. Accuracy (%) under perturbation attacks on MAABench ID.

AttackATLASBest baseline
Homoglyph69.3658.86
Zero-width space45.0425.50
Whitespace99.1999.20
Upper/lower case99.6399.19
Insert paragraphs99.6699.42
Article deletion99.6899.02
Number99.7399.48
Alternative spelling99.7899.51
Misspelling99.7899.40
Synonym90.6687.40
Avg. 10 attacks90.2586.30
  1. Very short texts (new, Appendix C.2, Tables 26 and 27). Models trained on full texts are evaluated on only 16 or 32 words of each test text. To check that ATLAS does not rely on the opening of a text, we also took 32 consecutive words from the middle of each text. ATLAS stays far more accurate than the baselines at every length, including the middle words:

Table R5. Accuracy (%) on very short texts.

WordsATLAS (ID)Best baseline (ID)ATLAS (Version)Best baseline (Version)
First 1675.8226.3652.9422.81
First 3290.2636.3664.7428.77
Middle 3274.6736.4952.7129.35
  1. RAID [1] as a third MAA benchmark (new, Appendix C.2, Table 28). RAID is a binary detection benchmark, so we reconstructed it as 5-family attribution and trained every method on 10k texts per family. ATLAS is best in every column, with the largest gain under repetition penalty:

Table R6. Accuracy (%) on RAID reconstructed as 5-family attribution.

RAIDIDGreedy decodingRepetition penaltyAvg. 11 attacks
ATLAS93.4272.8985.8876.32
Best baseline91.7272.3966.3770.21

[1] Dugan, Liam, et al. "Raid: A shared benchmark for robust evaluation of machine-generated text detectors." Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

[2] La Cava, Lucio, and Andrea Tagarelli. "Openturingbench: An open-model-based benchmark and framework for machine-generated text detection and attribution." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

We thank the reviewer again for the helpful comments, which led to new analyses of cost and robustness.

Dear Reviewer aoxv,

We hope this message finds you well. We are writing to kindly follow up on our response to your comments, which includes a new FLOPs analysis and additional robustness results. We want to make sure that our explanations adequately address the concerns you raised.

If any part of our response remains unclear, we would be very happy to provide further details. As the discussion period is coming to a close, we would greatly appreciate it if you could take a look at your convenience.

Thank you again for your time and helpful review.

Best regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of reviewers feedback and authors response

We thank all reviewers for their constructive comments and careful reading of our work.

Strengths. Reviewers found the problem relevant, as model authorship attribution matters for accountability and copyright (Shuu, aoxv). They valued the generative formulation, which may be less prone to shortcuts under distribution shift (Shuu), and the token-level clipping as a principled design supported by the ablations (tMGY). They recognized MAABench as a broad benchmark that tests both domain shift and model-version evolution, which makes the evaluation more relevant to deployment than ID-only experiments (tMGY). They also noted the strong improvements under distribution shift and the data efficiency of ATLAS (Shuu, tMGY), the teacher-attribution experiment (tMGY), and the clear writing (aoxv).

Main concerns and how we addressed them.

  • Robustness and additional benchmarks (aoxv, Shuu, tMGY). We evaluated ATLAS under ten perturbation attacks, on very short texts, and on RAID reconstructed as an attribution benchmark, and compared it with LLMDet on OpenTuringBench. ATLAS outperforms the strongest baseline in every setting.
  • Computational cost (aoxv, Shuu). A new FLOPs analysis shows that, at a comparable inference cost with a smaller backbone, ATLAS remains more accurate than every baseline.
  • Domain shift and the system prompt (Shuu, tMGY). A new domain-only control (Appendix C.1) separates the two factors. The persona prompt leaves ATLAS unchanged but lowers every baseline by at least 1.7pp.
  • Baseline setup (Shuu). We run the encoder baselines on an encoder of the same scale as ATLAS, so that only the method differs. As an additional check, we also tested them in their original settings. ATLAS stays far ahead in every case.
  • Single seed (tMGY). The main results are now three-seed means, with standard deviations.
  • Unseen families (tMGY). ATLAS can reject texts from unseen families with its likelihood score. Full open-set attribution remains future work, as stated in our conclusion.
  • Reproducibility (tMGY). Code and data are provided as anonymized supplementary material.

Summary of revisions (changes in blue):

  • Table 2, abstract, and Sections 1 and 5: three-seed means.
  • Section 5.1 and Table 3: LLMDet on OpenTuringBench.
  • Appendix B.2: baselines in their original settings.
  • Appendix B.5: variance over seeds.
  • Appendix C (new section, Further Analysis): error analysis and a domain-only control (C.1), robustness to perturbation attacks, very short texts and RAID (C.2), open-set attribution (C.3), and computational cost (C.4).
  • Appendix D: resemblance analysis of five additional deployed models from other families.
  • Supplementary material: anonymized code (training, scoring and evaluation scripts) and MAABench data.

We believe these revisions address the reviewers' concerns and make the claims of the paper better supported.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.