An AI-text detector designed to spot machine-written passages across generated, polished and rewritten cases reported high scores on the MIRAGE-SIG benchmark. Using AUROC, a score for how well a detector separates human-written and machine-generated text, NeuroStat recorded 0.9954 for Generate, 0.9739 for Polish and 0.9665 for Rewrite. On MOSAIC, a benchmark covering eight attack categories and six frontier language models, overall AUROC was 83.4 for the Qwen2-0.5B version and 78.1 for Qwen3.5-0.8B.
Two signals from one language model
The central idea is to combine signals that are often treated separately. NeuroStat extracts token-level statistical trajectories and deep semantic representations from one CausalLM backbone. A component called Macro-State Residual Modulation, or MSRM, conditions local features on mean entropy and mean log-rank.
The framework was trained end to end on only 500 pairs. The authors say that budget was chosen as a fairness constraint, not as a threshold showing that 500 pairs are sufficient.
A broader attack mix
To build MOSAIC, the authors' purification pipeline reduced 36,753 raw samples to 16,000 high-quality seeds. Texts had to reach a score of at least 85.0, and 3,101 samples were removed at the final scoring stage.
Each of those 16,000 human seeds was assigned one of 36 attack sub-methods and one of six frontier LLMs through length-stratified sampling rather than a full cross-product. The design yielded roughly 2,667 pairs per generator and between 1,777 and 2,667 pairs per attack category.
On MOSAIC, the two versions did not have the same profile at the study's reported five-per-cent operating point. TPR@5%, the reported detection rate at a setting fixed to 5% false positives, was 57.7 for Qwen2-0.5B and 58.1 for Qwen3.5-0.8B, even though the Qwen2 version had the higher overall AUROC.
Character/Encoding attacks produced another separation from the named baseline. NeuroStat reported AUROC of 71.2 with Qwen2-0.5B and 72.4 with Qwen3.5-0.8B, compared with 52.7 for DetectAnyLLM.
Where the score changed
Component comparisons on MIRAGE-DIG also favored the full configuration on the three listed tasks. For Polish, Rewrite and Generate, respectively, it reported AUROCs of 0.9576, 0.9490 and 0.9955. The corresponding figures were 0.9280, 0.9145 and 0.9934 for TF+TB concatenation, and 0.8990, 0.8918 and 0.9787 for TB-Only.
A separate loss comparison showed higher listed AUROCs for the full loss configuration than for cross-entropy alone. The full setup scored 0.9576 for Polish, 0.9490 for Rewrite and 0.9955 for Generate, while cross-entropy alone scored 0.8859, 0.8798 and 0.9850.
The detector's score also varied by target generator. Using Qwen2-0.5B as a surrogate, AUROC was 79.6 for GPT-5.4, 81.3 for Claude-Opus-4-6, 82.7 for Gemini-3.1-Pro, 84.1 for MiniMax-M2.7, 86.2 for GLM-5.0 and 86.5 for Kimi-K2.5. The micro-average was 83.4.
On the standard ImBD test for GPT-4o-polished text, the Qwen3.5-0.8B version reported AUROCs of 0.9972 on XSum, 0.9845 on Writing and 0.9679 on PubMed.
What the tests leave open
Those results remain benchmark evidence within a narrow test frame. The evaluation was limited to English-language texts and selected domains; multilingual and low-resource settings, code-generation text and legal-document data remain unexplored.
The paper also reports a computational cost difference. Relative to DetectAnyLLM, NeuroStat with Qwen2-0.5B used 6.4GB of peak memory and 31.1 milliseconds per sample, compared with 5.8GB and 28.4 milliseconds. The paper describes the gap as 0.6GB, 2.7 milliseconds and roughly 10 per cent overhead.
The work is a preprint identified as arXiv:2608.28009v1 and dated 28 August 2026. Its authors state that the code and MOSAIC benchmark are available at the listed repository.
Paper data and sources
Original title: Beyond Global Scalars: Synergizing Token-Level Statistics and Deep Semantics for Adversarial AIGC Text Detection
Authors: Peiming Li, Yifan Wang, Zhiyuan Hu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text