Breaking the Generator Barrier: Disentangled Representation for Generalizable AI-Text Detection

Xiao Pu; Zepeng Cheng; Lin Yuan; Yu Wu; Xiuli Bi

cs.CLApr 15, 2026

Breaking the Generator Barrier: Disentangled Representation for Generalizable AI-Text Detection

Authors: Xiao Pu, Zepeng Cheng, Lin Yuan, Yu Wu, Xiuli Bi

Organizations: Chongqing University of Posts and Telecommunications, China

Abstract

As large language models (LLMs) generate text that increasingly resembles human writing, the subtle cues that distinguish AI-generated content from human-written content become increasingly challenging to capture. Reliance on generator-specific artifacts is inherently unstable, since new models emerge rapidly and reduce the robustness of such shortcuts. This generalizes unseen generators as a central and challenging problem for AI-text detection. To tackle this challenge, we propose a progressively structured framework that disentangles AI-detection semantics from generator-aware artifacts. This is achieved through a compact latent encoding that encourages semantic minimality, followed by perturbation-based regularization to reduce residual entanglement, and finally a discriminative adaptation stage that aligns representations with task objectives. Experiments on MAGE benchmark, covering 20 representative LLMs across 7 categories, demonstrate consistent improvements over state-of-the-art methods, achieving up to 24.2% accuracy gain and 26.2% F1 improvement. Notably, performance continues to improve as the diversity of training generators increases, confirming strong scalability and generalization in open-set scenarios. Our source code will be publicly available at https://github.com/PuXiao06/DRGD.

Figures & tables

Figure 1: Comparison with competitive approaches. On the unseen OPT generators, our method achieves the best balance between accuracy and efficiency, and its advantage grows with training generator diversity.

Figure 2: Overview of the proposed framework. The model enhances generalization to unseen generators by enforcing semantic disentanglement between AI-detection and generator-aware features. It integrates compact dual-bottleneck (DB) encoding, cross-view (CV) regularization, and discriminator-guided (DG) adaptation into a cohesive disentanglement-enhancing pipeline.

Categories	MAGE Corpus Li et al. (2024)
Categories	FLAN-T5	GPT	LLaMA	OPT	GLM	BigScience	EleutherAI
LLMs included	small/base/large xl/xxl	davinci-002/003 gpt-turbo-3.5	6B/13B 30B/65B	2.7B/6.7B iml-1.3B/30B	130B	Bloom-7B	GPT-J-6B GPT-NeoX-20B
AI	9,000	13,800	9,000	9,000	9,100	8,900	9,000
Human	9,000	13,800	9,000	9,000	9,100	8,900	9,000

Table 1: Number of AI-generated and human-written samples across datasets. Balanced sampling is applied to categories with multiple LLMs so that each category contains a comparable number of samples, resulting in 20 representative models overall.

Algorithm 1: Structured Disentanglement Framework
Input: encoder $E$ , bottleneck encoders $E_{a}$ , $E_{g}$ , discriminators $D_{a}$ , $D_{g}$ , weight factor $\beta$ , training set $\mathcal{X}$
Output: optimized encoders and discriminators
1	for each training epoch do
2	for each mini-batch $\mathcal{B}=\{x_{i},y_{i},s_{i}\}_{i=1}^{N}$ do
3	Stage I: Encoding and Regularization
4	$\mathbf{h}_{i}=E(x_{i})$

Table 4

	Accuracy (%) / $F_{1}$ -Measure
LLM Categories	FLAN-T5	GPT	LLaMA	OPT	GLM	BigScience	EleutherAI
Zero-shot methods
$\text{GLTR}_{\text{ACL'19}}$	56.9 / 52.4	71.8 / 71.5	75.1 / 75.8	75.1 / 77.3	77.8 / 78.2	86.6 / 86.4	80.2 / 80.0
$\text{Fast-DetectGPT}_{\text{ICLR'24}}$	57.3 / 74.9	70.0 / 82.5	73.4 / 79.8	74.6 / 76.1	69.9 / 72.8	79.3 / 84.2	77.0 / 82.9
$\text{MCP}_{\text{ACL'25}}$	58.6 / 76.3	72.1 / 84.4	75.9 / 81.5	77.4 / 78.3	74.7 / 75.2	86.2 / 86.5	83.7 / 84.4
Training-based methods

Table 2: Cross-generator generalization results on seven LLM categories from the MAGE benchmark under the leave-one-generator-out protocol. Zero-shot and training-based methods are compared against our approach, which yields consistent gains across all held-out generators.

Model	Accuracy (%) / $F_{1}$ -Measure
Model	$N=2$	$N=3$	$N=4$	$N=5$
OPT Test
BERT ${}_{\text{NAACL'19}}$	72.8/72.2	73.1/69.9	76.7/74.9	78.1/76.1
$\text{IT:CLUST}_{\text{ACL\textquoteright 22}}$	69.4/67.7	68.4/69.8	70.1/68.1	68.4/68.1
PECOLA ${}_{\text{ACL'24}}$	70.1/68.1	76.8/75.5	80.5/80.0	81.1/80.9
SCRN ${}_{\text{ACL'24}}$	69.2/68.3	70.7/67.4	72.5/69.9	76.1/75.9

Table 3: Cross-generator generalization on OPT and FLAN-T5. Models are trained with

N

categories under a fixed training size of 12,000 samples and evaluated on an unseen LLM category.

Figure 3: Progressive t-SNE visualization of disentangled feature spaces. The BERT baseline (a) collapses, with unseen samples (red) mixed into training clusters (blue/purple). After decoupling, the AI-detection (b) and generator-aware (c) branches begin to separate (b1/c1), but generalization remains unstable and generator-sensitive. With DB encoding (b2/c2), semantic noise is reduced, yielding more compact clusters. Cross-view regularization (b3/c3) smooths boundaries and enhances robustness. Finally, discriminator-guided adaptation (b4/c4) yields clear human–AI separation in the detection branch and generator-invariant alignment in the auxiliary branch.

Word-level Perturbations
Metric	Accuracy (%) / $F_{1}$ -Measure
Model	BERT	SCRN	Ours
Original	82.1 / 84.2	81.3 / 79.2	89.1 / 88.3
Delete	68.8 / 72.6	67.6 / 64.9	80.0 / 79.9
Swap	66.9 / 64.2	62.6 / 59.9	66.5 / 69.9
Insert	66.9 / 65.1	62.6 / 60.9	67.5 / 70.6

Table 4: Robustness comparison under two types of attacks on the unseen OPT category. Word-level perturbation are evaluated using accuracy and

F_{1}

scores, while adversarial attacks are reported by attack success rate (ASR). Lower ASR indicates stronger robustness.

Figure 4: T-SNE visualizations of AI-detection features $\mathbf{a}$ from MLP and DB encoders. With human-AI coloring (top), the encoders separate authenticity. With generator coloring (bottom), MLP features split into distinct generator clusters, while DB suppresses generator bias, yielding compact, generator-invariant representations.

Figure 5: Robustness under generator-aware and Gaussian perturbations. Perturbation-based regularization enhances our method’s resilience, maintaining superior accuracy over the BERT baseline in both structured and random noise settings.

Figure 6: Intra-class compactness with vs. without adaptation. Adaptation consistently improves feature compactness across tasks, yielding lower dispersion metrics and more task-aligned representations.

Components					Accuracy (%) / $F_{1}$ -Measure
BERT	Disen.	DB	Regular.	Adapt.	FLAN-T5	OPT
✓					60.9 / 52.6	82.1 / 84.2
✓	✓				62.0 / 53.3	83.7 / 83.5
✓	✓	✓			65.2 / 60.2	87.1 / 86.0
✓	✓	✓	✓		66.8 / 63.8	88.3 / 87.4
✓	✓	✓	✓	✓	68.9 / 64.1	89.1 / 88.3

Table 5: Incremental ablation results on two unseen LLM categories.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

	Accuracy (%) / $F_{1}$ -Measure
LLM Categories	FLAN-T5	GPT	LLaMA	OPT	GLM	BigScience	EleutherAI
$\text{BERT}_{\text{NAACL'19}}$	70.5 / 66.5	83.9 / 82.7	87.4 / 88.5	86.5 / 87.3	93.1 / 93.9	98.1 / 97.3	98.0 / 97.1
$\text{PECOLA}_{\text{ACL'24}}$	75.5 / 75.8	85.6 / 85.1	88.9 / 88.7	88.5 / 88.3	80.9 / 80.2	90.4 / 89.6	86.7 / 87.4
$\text{SCRN}_{\text{ACL'24}}$	69.1 / 68.5	74.6 / 73.7	79.7 / 78.3	85.6 / 83.9	73.4 / 70.2	89.5 / 89.1	92.1 / 91.8
Ours	77.4 / 77.1	86.9 / 87.4	91.7 / 90.1	93.8 / 92.4	95.7 / 94.6	98.7 / 96.1	98.3 / 98.9

Appendix

Table 6: Performance comparison across seven LLM generator families on the full original MEGA dataset.

Metric	Accuracy (%) / $F_{1}$ -Measure
Model	BERT	SCRN	Ours
GPT-4	52.8 / 51.0	51.6 / 50.9	61.8 / 65.4
GPT-5	61.0 / 61.6	53.8 / 52.2	68.7 / 70.1
Claude-sonnet-4-5	51.6 / 56.6	52.9 / 52.1	57.1 / 60.5
DeepSeek-R1	63.3 / 69.7	55.4 / 53.2	71.0 / 69.8

Appendix

Table 7: Performance comparison among BERT, SCRN, and our method on the GPT-4 OOD split in MAGE (AI: 762, Human: 800), along with

\sim

800 manually collected samples from GPT-5, Claude, and DeepSeek.

Section	Train	Test
4.2.1 Main Results (Table 2 )	Leave-One-Generator-Out
4.2.2 Diversity (Table 3 )	BigScience GPT LLaMA GLM EleutherAI	OPT FLAN-T5
4.3 Progress Gen. (Fig. 3 )	BigScience GPT	FLAN-T5
4.4.1 Bias Filter (Fig. 4 ) A.5 Bias Filter (Fig. 8 )	BigScience GPT FLAN-T5	-
4.4.2 Robustness (Fig. 5 )	BigScience GPT	-
4.4.3 Compactness (Fig. 6 )	GPT	-

Appendix

Table 8: Overview of training and testing setups across different experiments.

Figure 7: T-SNE visualizations across LLaMA, OPT, and FLAN-T5. Training (blue/purple) and held-out (red) generators are shown, with dark/light shades for AI/human texts. (a) Baseline BERT entangles features on unseen generators. Conversely, our framework yields (b) tighter human–AI separation with mitigated generator bias in the AI-detection branch, and (c) explicit isolation of generator-specific features in the generator-aware branch.

Figure 8: MLP vs. DB encoder representations of generator-aware branch $\mathbf{g}$ . With Human–AI coloring (top), both separate distinct generators, but MLP retains human-AI bias while DB filters it. With generator coloring (bottom), MLP forms divided human and AI samples, whereas DB suppresses bias and yields compact, generator-aware clusters.

AI-Generated Text Detection

AI-generated text detection develops methods to distinguish human-written, machine-generated, and AI-edited text.

See all 167 papers

Detecting the Machine: A Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions
Mar 18, 2026Madhav S. Baidya, S. S. Baidya, Chirag ChawlaAI-Generated Text Detection Robustness of Machine-Generated Text Detection
Indian Institute of Technology (BHU), Varanasi, India · Indian Institute of Technology Guwahati, India
MELD: Multi-Task Equilibrated Learning Detector for AI-Generated Text
May 7, 2026Chenjun Li, Cheng Wan, Johannes C. PaetzoldAI-Generated Text Detection Domain Generalization
Cornell University, Ithaca, NY 14853, USA · Weill Cornell Medicine, New York, NY 10021, USA · Cornell Tech, New York, NY 10044, USA
READER: Reasoning-Enhanced AI-Generated Text Detection
May 24, 2026Pingfan Su, Kai Ye, Shijin Gong +4AI-Generated Text Detection
Department of Statistics, London School of Economics and Political Science · School of Management, University of Science and Technology of China · School of Mathematics, University of Birmingham

Mar 18, 2026J/K move · Enter open · S save

Detecting the Machine: A Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions

Madhav S. Baidya, S. S. Baidya, Chirag Chawla

Indian Institute of Technology (BHU), Varanasi, India · Indian Institute of Technology Guwahati, India

AI-Generated Text Detection Robustness of Machine-Generated Text Detection

Open paper AskPDF arXiv

The rapid proliferation of large language models (LLMs) has created an urgent need for robust and generalizable detectors of machine-generated text. Existing benchmarks typically evaluate a single detector on a single dataset under ideal conditions, leaving open questions about cross-domain transfer, cross-LLM generalization, and adversarial robustness. We present a comprehensive benchmark evaluating diverse detection approaches across two corpora: HC3 (23,363 human-ChatGPT pairs) and ELI5 (15,000 human-Mistral-7B pairs). Methods include classical classifiers, fine-tuned transformer encoders (BERT, RoBERTa, ELECTRA, DistilBERT, DeBERTa-v3), a CNN, an XGBoost stylometric model, perplexity-based detectors, and LLM-as-detector prompting. Results show that transformer models achieve near-perfect in-distribution performance but degrade under domain shift. The XGBoost stylometric model matches performance while remaining interpretable. LLM-based detectors underperform and are affected by generator-detector identity bias. Perplexity-based methods exhibit polarity inversion, with modern LLM outputs showing lower perplexity than human text, but remain effective when corrected. No method generalizes robustly across domains and LLM sources.

Abstract

AI-Generated Text Detection

Explore similar work

Detecting the Machine: A Comprehensive Benchmark of AI-Generated Text Detectors Across Architectures, Domains, and Adversarial Conditions

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

READER: Reasoning-Enhanced AI-Generated Text Detection