acceptodds
Under review as a conference paper at ICLR 2027

Mixture-of-Experts Routing Responses Are Strong Signals for AI-Generated Text Detection

Abstract

Generalizing AI-Generated Text Detection to unseen domains, generators, and generation tasks remains challenging. We introduce a simple yet effective detector that uses the native routing responses of a frozen Mixture-of-Experts (MoE) language model. Although pretrained without any AI-Generated Text Detection objective, these routers produce responses that could be used for distinguishing authorship. Our method averages router logits over tokens, concatenates the resulting features across layers, and trains a single linear classifier while keeping the observer frozen. With GPT-OSS-20B, the classifier contains only trainable parameters and is trained on human-AI text pairs. On MIRAGE, it achieves AUROCs of , , and in cross-domain, cross-generator, and cross-task evaluations, respectively, outperforming all evaluated baselines. Controlled analyses show that detection performance is largely preserved when removing logit shifts shared across experts, but declines when expert identity is removed. Even discrete expert-selection frequencies retain a strong detection signal. Evaluations across nine frozen MoE observers further show that this routing footprint recurs across model families. These findings show that native MoE routing exposes an authorship footprint carried by relative, expert-specific preferences, and that a minimal linear readout of routing responses supports strong detection across diverse observers.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-Generated Text Detection, AIGC-Detection, Machine-Generated Text Detection, Mixture-of-Experts

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_9K4WOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the possibility to use native routing signals from pre-trained Mixture-of-Experts LLMs to detect machine-generated text. The core idea is that, despite not being trained for this kind of task, they inherently retain signals involving authorship-related information, and these can be extracted and fed into a linear classifier to perform detection. To this aim, the authors use GPT-OSS-20B as a routing model and, based on 500 human-AI pairs, they train a small linear classifier. The experimental setup based on the MIRAGE dataset shows high AUROC values, which outperform the considered baselines.

Critical strengths
  • The idea is interesting and involves a rather underexplored connection between MoE inner workings and detectability signals. This is also valuable due to the fact such models have not been instructed to perform detection, and the proposed method simply exploits representations learned for language modeling.
  • The overall architecture depends only on a frozen MoE model and a simple readout module, trained on just 500 training pairs.
  • The ablation of routing representations provides valuable insights into the inner workings of such a detection signal, and the fact the latter generalizes properly across MoE model implementations corroborates it.
Critical weaknesses
  • First, the proposed approach is evaluated on a single benchmark (MIRAGE). This limits the strengths of the claims about the proposed approach, and demands further evaluations on widely used approaches (e.g., RAID, OpenTuringBench, DetectRL). Furthermore, the human-edited AI text might also introduce some stylistic artifacts that trained readout could exploit rather than leverage proper authorship signal. Therefore, external benchmarks might strengthen generalizability and exclude any kind of leakage.
  • While the reported AUROC is high overall, this might hide some practical constraints for real-life deployment. For example, as suggested by Table 2, TRP at 1% FPR is rather low, which means leaving a non-negligible set of AI-generated text undetected. Similar patterns hold in Table 4, where Prob [email protected] for DeepSeek is remarkably low despite the high AUC. Furthermore, Figure 3(a) suggests a strong performance imbalance based on the prefix length. Taken together, such results suggest that strong ranking performance does not reflect onto detection effectiveness, and matter a lot for real deployment of the proposed approach.
  • Last but not least, this approach requires accessing MoE models’ internals, and such models are typically larger in size (with some exceptions e.g., OLMoE) than smaller proxy models used by current detectors (e.g., RoBERTa-based or < 7-8B decoders). Hence, the small cost of the projection module has to balance the high cost of serving such models to extract the signal.
Questions for authors
  • Can you evaluate the proposed approach on additional benchmarks like RAID, OpenTuringBench, and DetectRL (as they also cover attacks and out-of-domain data)? This is particularly important and I'll be happy to revise my score accordingly.
  • Can you better shed light on the imbalance between AUROC and operating point performance? This would matter a lot too for deployment.
  • Across the paper, authorship identity and information are widely used concepts. However, this might wrongly overlap with the idea of Model Authorship Attribution, i.e. identifying the specific generator that produced a given text to be detected. Please clarify accordingly.
  • How are the “operating-point” parameters calibrated for real-life deployment considering the reported TPR @ 0.1 FPR?
Reviewer_ubhcOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether the native routing responses of pretrained Mixture-of-Experts language models contain information useful for detecting AI-generated text. The proposed approach extracts routing logits or probabilities from a frozen MoE [the paper called this as observer], averages them over tokens, concatenates the resulting layer-wise features, and trains a linear binary classifier. Using GPT-OSS-20B [the observer] and 500 human-AI training pairs, and they evaluated on MIRAGE cross-domain, cross-generator, and cross-task splits. It further has logit centering, expert identity, discrete selections, contextual positions, layer depth, and interaction features, and evaluate multiple MoEs. The paper also compares several things like native routing with full hidden-state, PCA, and random-projection readouts to investigate what detection information survives routing compression.

Critical strengths

An interesting finding about MoE routing: The main strength of this paper is showing that the routing logits of pretrained MoE models, even though they were never trained for AI-text detection, already carry strong signals to distinguish human-written and AI-generated text. What makes this more interesting is that the same phenomenon is observed across nine different MoE models which the paper has evaluated, which is like saying this is not just one model family specif but all the MoEs exhibit the same phenomenon. The generalization experiments and ablations support this finding. More importantly, this finding opens an interesting research direction into what other latent properties of text might be implicitly encoded and exposed through pretrained expert-routing behavior.

Critical weaknesses
  1. The paper shows that routing captures authorship signals, but does not really explain why this happens or what exactly the experts are learning that makes this possible.
  2. Although the experiments cover different domains, generators, tasks, and MoE models, everything is evaluated on MIRAGE alone. Testing on an independent dataset would help confirm how general this finding is.
  3. [IMP] Random projections and PCA of hidden states actually perform better than the native routing logits [Appendix D]. So, it is not clear what makes routing particularly special compared to other projections.
Questions for authors
  1. What is the main advantage of using routing responses over Random projections and PCA of hidden states?
  2. Why the routing captures authorship signals?
  3. How confident that this phenomenon also holds on other independently constructed datasets other than MIRAGE?
2 comments

Thank you for your constructive feedback.

Comparison of Native Routing with PCA and Random Projections (W3, Q1) Our goal is to analyze how much authorship information a pretrained MoE router preserves from hidden states (§1, L068–084; §3.1, L165–194), rather than to outperform all dense activation probes. Projecting hidden states into a compact expert space (Eq. 1, Thm. 3.1) incurs information loss; comparing routing with PCA and hidden-state probes (§4.2, L452–458; App. D) quantifies what survives in this fixed pretraining subspace. Unlike PCA (which stores full dmodeld_{\text{model}} activations to fit a task-specific projection) and random projections (which sample random directions and train task-specific classifier weights), router scores are task-agnostic signals produced during standard inference. Moreover, discrete top-KK expert counts—a non-linear, piecewise-constant operation irreducible to a linear probe on mean hidden states—match continuous scores (§3.2, L211–215; Table 2), showing that expert allocation encodes text provenance.

Empirical Analysis of How Routing Captures Authorship Signals (W1, Q2) Controlled probes and ablations (§3.3, L216–297; §4.2, L365–451, Tables 2–3, Figs. 2–4) clarify how routing captures authorship:

  1. Signal carrier: Removing logit shifts shared across experts preserves performance, whereas using the shared shift alone fails. Destroying expert identities via layer-wise probability sorting or retaining only coarse shape statistics (entropy, HHI) degrades detection, while discrete top-KK frequencies retain strong signal. Thus, relative preferences over specialized experts—not global logit shifts or entropy—carry the footprint.
  2. Depth, context, and interactions: Single-layer readability peaks in middle-to-late layers and accumulates across depth to improve low-FPR detection. Readability scales with context length and is stronger at suffix and random positions than at prefixes. Adding cross-layer covariance or temporal-gap interactions yields no gain over token-mean readouts, confirming that global shifts in expert preferences dominate across all nine MoE observers (§4.3, L459–470, Table 4).

Evaluation on Independent Datasets Beyond MIRAGE (W2, Q3) We evaluated the frozen GPT-OSS-20B observer (24×32=76824\times 32=768 router logits, 769769-parameter linear head) on RAID and DetectRL, training on 500500 clean human–AI pairs (3 seeds) and testing on source-disjoint splits:

BenchmarkConditionAUROC (mean ± SD)
RAIDNo attack0.9784 ± 0.0001
RAIDParaphrase0.9557 ± 0.0017
RAIDPerplexity misspelling0.9719 ± 0.0004
RAIDWhitespace addition0.9224 ± 0.0004
DetectRLDirect prompting0.9985 ± 0.0001
DetectRLDIPPER paraphrase0.9956 ± 0.0008
DetectRLBack translation0.9528 ± 0.0059
DetectRLCharacter perturbation0.9591 ± 0.0045

Nine of RAID's 11 attacks maintain strong readability (except homoglyph and zero-width-space spoofing), confirming generalization beyond MIRAGE. Comparison with RAID baselines (Tables 15–16) and official repository records:

MethodTPR@5% CleanTPR@5% ParaphraseAUROC CleanAUROC All
Fast-DetectGPT73.671.8——
Binoculars79.680.30.8440—
RADAR70.967.30.81890.8192
GLTR62.647.20.72680.7090
RoBERTa-base (GPT2)59.168.90.74950.7229
RoBERTa-large (GPT2)56.772.90.71600.6928
RoBERTa-base (ChatGPT)44.849.20.77290.6026
LLMDet35.028.50.65920.6290
GPTZero66.564.0——
Originality85.096.7——
Winston71.052.6——
ZeroGPT65.5*46.7——
Ours (GPT-OSS-20B)†91.180.20.97900.8331

“—”: unavailable; *missed target FPR; †source-disjoint public holdout using RAID per-domain thresholding and official pooling weights.

Thank you for the response and additional experiments. I appreciate the authors' efforts, the new evaluations on RAID and DetectRL, which address my concern about evaluating only on MIRAGE and further strengthen the finding that routing captures authorship information.

Regarding my question about why this phenomenon occurs, I agree that the ablations help identify which parts of routing carry the signal, such as relative expert preferences and expert identity. However, my concern was more about why pretrained routing captures this information in the first place. Is it actually related to learned expert specialization, or is the router simply preserving authorship information that is already present in the hidden states?

Overall, I appreciate the additional experiments and clarifications. They strengthen my positive assessment of the paper, but I think the underlying explanation of the phenomenon remains an open question. I maintain my original score.

Reviewer_NPv9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores the use of routing responses from pre-trained Mixture-of-Experts (MoE) models for AI-generated text detection. The authors propose a simple approach that extracts router logits from a frozen MoE model, averages them across tokens, and trains a linear classifier to distinguish human-written and AI-generated text. Using only 500 human-AI text pairs and 769 trainable parameters, the method achieves AUROCs of 0.9663, 0.9909, and 0.9937 in cross-domain, cross-generator, and cross-task evaluations on MIRAGE. The paper also analyzes different routing features and evaluates the method across nine MoE models, showing that routing responses contain useful information for authorship detection.

Critical strengths

The proposed method achieves strong results in cross-domain, cross-generator, and cross-task evaluations, outperforming the evaluated baselines. The results suggest that routing responses can provide useful signals for detecting AI-generated text, even on new data. The evaluation across nine MoE models further supports the generality of this finding.

The method only requires a small linear classifier while keeping the pre-trained MoE model frozen. The authors provide several studies to examine the effects of different routing representations, expert identities, and model layers. These experiments help explain which routing features contribute most to detection performance.

Critical weaknesses

Although the classifier has only 769 trainable parameters, the method still requires a large MoE model and access to its internal routing responses. This may limit its use with closed-source models or in environments with limited computational resources. The reported inference efficiency is measured using eight H200 GPUs, and a more detailed comparison of end-to-end computational costs would help clarify the practical advantages of the method.

The main experiments are conducted on the MIRAGE dataset, which may not fully represent real-world AI-generated text. The paper does not evaluate how the method performs on human-edited, paraphrased, or mixed human-AI text. These scenarios are common in practice, and additional experiments would help demonstrate whether the method can generalize to more realistic settings.

Although the ablation studies show that expert identities and relative routing preferences are important for detection, the paper does not sufficiently explain why these patterns differ between human-written and AI-generated text. It remains unclear whether these differences are mainly related to writing style, semantic structure, or other linguistic features. Further analysis would help explain the underlying mechanism and provide stronger support for the proposed routing footprint.

Questions for authors

How does the proposed detector perform on human-edited or paraphrased AI-generated text? Could the authors provide results on an additional benchmark to further validate its robustness?

Can the authors provide a more detailed comparison of end-to-end inference time and computational costs against existing detection methods, especially under more practical hardware settings?

Figure 3 shows that detection performance drops considerably for shorter texts. How reliable is the method for short inputs, particularly when a low false-positive rate is required?

2 comments

Thank you for the feedback on text transformations, efficiency, short texts, and routing mechanisms.

Robustness to Text Transformations (W2, Q1) We conducted new experiments on two independently constructed benchmarks—RAID and DetectRL—using the exact same frozen GPT-OSS-20B observer (pooling router logits into features with a -parameter linear head). For each benchmark, the readout is trained on clean human–AI pairs across three seeds and evaluated on source-disjoint test splits across diverse paraphrasing, perturbation, mixing, and polishing conditions. See Table A1 (provided in our follow-up comment) for details.

Across 9 of RAID's 11 attacks and across DetectRL's paraphrasing (DIPPER, back translation), character perturbation, LLM-centered mixing, and AI-polishing-human settings, the frozen routing readout maintains strong detection performance (with boundary cases only under character-level homoglyph and zero-width-space spoofing). See Table A2 (provided in our follow-up comment) for details. Regarding human-rewritten AI text, standardized evaluation is challenging: substantial human edits blur authorship labels, and rewriting protocols vary. We therefore evaluate related but non-equivalent settings—LLM-polished human and human–AI mixed text.

We conducted two sets of experiments: (1) on a single RTX 4090 under a 24 GiB VRAM budget (BF16), evaluating three lightweight MoE observers (Granite-3.1-1B-A400M, Granite-3.1-3B-A800M, and OLMoE-1B-7B) alongside a matched-backbone SV-Detect baseline; and (2) on a single H100 GPU with a unified GPT-OSS-20B backbone (native MXFP4 expert weights), comparing Ours, SV-Detect, and Fast-DetectGPT (SR, analytic self-reference) across five benchmarks/settings. End-to-end latency covers raw text to synchronized CPU score (512-token cap; 1,000 stratified texts/setting; 20 warmups, 3 repeats). Throughput (texts/s) is computed as the total number of processed texts divided by the total elapsed time, with ms/text reported using the same timing measurements. Both metrics are calculated from unrounded raw timings. Peak GPU memory is measured with NVML (Table A3, follow-up comment).

The results show that: (1) lightweight MoE observers run comfortably within 4.8–15.5 GiB on a single RTX 4090 at 66–102 texts/s (batch 8) while maintaining strong cross-generator and cross-task detection ( AUROC), and even GPT-OSS-20B peaks at 18.09 GiB (within a 24 GiB card's memory capacity), whereas Fast-DetectGPT SR requires 27.03 GiB due to retaining vocabulary-level logit statistics; (2) under an identical backbone, our routing readout achieves higher batch-8 throughput than SV-Detect (62.46 vs. 59.11 texts/s), while maintaining higher AUROC on CG and CT; and (3) offline training is lightweight, requiring only (RTX 4090, unique texts across 3 settings) or (1× H100, unique texts across 5 settings) for training feature extraction, plus (4090) or (H100) for full 3-seed cross-validation and linear head fitting across all settings.

Reliability on Short Inputs (Q3) Our method already achieves strong performance on relatively short texts. On DetectRL's short-text subsets, our frozen readout reaches an AUROC of at just 20 tokens (cross_length_20, roughly a single sentence) and rises to at 40 tokens (cross_length_40, about 1–2 sentences). Similarly, on MIRAGE (Section 4.2, Lines 407–421, Figure 3), using only 32 token responses already yields AUROCs of (prefix) and (suffix/random), demonstrating that the routing footprint is already effective at the sentence-to-short-paragraph scale. Below a dozen tokens, sparse stylistic and structural evidence makes reliable authorship attribution substantially harder for any detector.

Why Routing Patterns Differ Between Human and AI Text (W3) While our probes are designed to localize where the authorship signal resides in the routing trace rather than isolate a single linguistic cause, our experimental design and ablations in Sections 3.3 and 4.2 (Lines 216–297, 365–451) rule out simple confounds. MIRAGE pairs each human text with an AI text rewritten or regenerated from the same source (Section 4.1, Lines 321–323), controlling for topic and semantic content so that routing differences primarily reflect stylistic and structural generation patterns. Ablations (Table 2, Figures 2–3) show that the signal is carried by specific expert identities rather than routing entropy, peaks in middle-to-late layers (where contextual and syntactic representations form) rather than early lexical layers. This is consistent with higher-level contextual and stylistic differences activating distinct expert coalitions.

Additional Results for Reviewer NPv9 (Tables A1–A3)

Table A1. Robustness on RAID and DetectRL. The same frozen GPT-OSS-20B routing observer; AUROC is mean ± SD across three seeds. Selected conditions are shown.

BenchmarkEvaluation conditionAUROC (mean ± SD)TPR@5% FPR
RAIDNo attack0.9784 ± 0.00010.9124
RAIDParaphrase0.9557 ± 0.00170.8080
RAIDPerplexity misspelling0.9719 ± 0.00040.8885
RAIDSynonym replacement0.9773 ± 0.00010.9111
RAIDWhitespace addition0.9224 ± 0.00040.7305
RAIDHomoglyph0.0615 ± 0.01660.0068
RAIDZero-width-space addition0.0184 ± 0.00290.0012
RAIDAll conditions, including no attack0.8110 ± 0.00130.7239
DetectRLDirect prompting0.9985 ± 0.00010.9964
DetectRLDIPPER paraphrase0.9956 ± 0.00080.9854
DetectRLBack translation0.9528 ± 0.00590.8242
DetectRLCharacter perturbation0.9591 ± 0.00450.8710
DetectRLLLM-centered mixing0.9876 ± 0.00190.9487
DetectRLPolishing human-origin text0.8956 ± 0.00170.5830

Table A2. RAID baseline reference comparison. Paper values come from RAID Tables 15–16; official AUROCs are historical evaluations in RAID's repository. Our row uses a separate source-disjoint public holdout, so these are reference results rather than strictly matched comparisons.

MethodPaper TPR@5% no attack (%)Paper TPR@5% paraphrase (%)Official AUROC unattacked AIOfficial AUROC all AI
Fast-DetectGPT73.671.8——
Binoculars79.680.30.8440—
RADAR70.967.30.81890.8192
GLTR62.647.20.72680.7090
RoBERTa-base, GPT259.168.90.74950.7229
RoBERTa-large, GPT256.772.90.71600.6928
RoBERTa-base, ChatGPT44.849.20.77290.6026
LLMDet35.028.50.65920.6290
GPTZero66.564.0——
Originality85.096.7——
Winston71.052.6——
ZeroGPT65.5*46.7——
Ours, GPT-OSS-20B, separate public holdout†91.180.20.97900.8331

“—” = unavailable historical aggregate. The asterisk indicates that ZeroGPT could not attain the target FPR in non-adversarial evaluation. †Our source-disjoint public holdout uses RAID's per-domain threshold search and official all-human-variant pooling weights.

Table A3. End-to-end single-GPU efficiency and detection. Batch-1/8 times are ms/text; throughput is texts/s; AUROC is listed as MIRAGE CD/CG/CT and DetectRL/RAID clean.

DeviceObserverMethodVRAM (GiB)B1 ms/textB8 ms/textB8 texts/sMIRAGE CD/CG/CTDetectRL/RAID clean
RTX 4090 (24 GiB)Granite-3.1-1B-A400MOurs4.7768.819.79102.140.8974 / 0.9706 / 0.9796—
RTX 4090 (24 GiB)Granite-3.1-3B-A800MOurs9.15101.5715.0666.390.9209 / 0.9777 / 0.9857—
RTX 4090 (24 GiB)OLMoE-1B-7B-0924Ours15.5071.3511.1389.850.9285 / 0.9700 / 0.9843—
RTX 4090 (24 GiB)OLMoE-1B-7B-0924SV-Detect15.4870.4611.4487.450.9464 / 0.9604 / 0.9619—
H100 (matched 20B)GPT-OSS-20B (MXFP4)Ours18.0960.0362.4674.190.9744 / 0.9906 / 0.99430.9838 / 0.9778
H100 (matched 20B)GPT-OSS-20B (MXFP4)SV-Detect17.8860.5059.1168.750.9818 / 0.9885 / 0.99350.9861 / 0.9863
H100 (matched 20B)GPT-OSS-20B (MXFP4)Fast-DetectGPT SR27.0359.6054.0162.690.6545 / 0.6763 / 0.69520.6866 / 0.5199

Batch-8 throughput is computed as the total number of processed texts divided by the total measured elapsed time across runs. Per-text latency is derived from the same aggregate measurements. Minor discrepancies in reported reciprocals arise from rounding.

Reviewer_KzoFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores using MoE routing signatures to train AI-generated text detectors. In an MoE model, each layer has a router that scores which expert subnetworks should handle each token. The paper records router scores from gpt-oss-20b, average over tokens, and concatenate across layers. That produces 24 layers × 32 experts = 768 dimensional vector. Then a logistic regression is trained on only 500 human / AI examples to do AI-generated text detection.

Main findings:

  • On MIRAGE, this detector achieves AUROC of 0.966 for unseen domains, 0.991 for unseen generators, and 0.994 for unseen tasks. It beats all baselines. Accuracy rises with text length. The method works across 9 MoE models
  • Ablation finds that expert identity matters in addition to distribution of score. Discrete top-K expert choice counts work well.
  • However, replacing router scores with full hidden states is better (0.990 vs 0.975), even after PCA reduction to the same 768 dimension (0.987).
Critical strengths

None

Critical weaknesses
  • Using the router scores is actually not special. Router scores (more accurately, logits) are a linear projection of hidden states. The average of router scores over tokens is a projection of the average of hidden state over tokens. The detector is essentially a 768 dimensional linear probe on hidden states averaged over tokens.
  • This is corroborated by the control experiments using hidden states directly. Averaged hidden states after PCA dimensionality reduction to the same size gives a more accurate detector. So really the finding is just “linear probes on averaged hidden states detect AI text”. which RepreGuard and others already established. The MoE framing adds no demonstrated value.
Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.