Exons-Detect: Identifying and Amplifying Exonic Tokens via Hidden-State Discrepancy for Robust AI-Generated Text Detection
Exons-Detect uses hidden-state discrepancies to identify exonic tokens, boosting AI text detection robustness without training.
Key Findings
Methodology
Exons-Detect employs a dual-proxy LLM framework to extract token-level hidden representations, computing cosine distances to measure discrepancies. Tokens with significant differences are classified as exonic, then assigned importance weights via a nonlinear function g(·). Combining these weights with probabilistic measures like log-perplexity and cross-perplexity, the method constructs an interpretable translation score. This process enhances detection robustness by focusing on source-relevant tokens without requiring model training. Core components include token feature extraction, discrepancy-based identification, importance weighting, and score calculation.
Key Results
- On the DetectRL dataset, Exons-Detect achieved an average AUROC of 92.14%, outperforming the previous best DNA-DetectLLM by 2.2%. It maintained high performance across multi-model and multi-domain settings, demonstrating strong generalization.
- Under adversarial attacks such as paraphrasing with DIPPER and polishing with GPT-4, Exons-Detect's AUROC decreased less than 2%, significantly better than training-based detectors like BiScope, confirming its robustness in real-world scenarios.
- In tests across varying input lengths (from 50 to 240 tokens), Exons-Detect consistently outperformed baselines, with an average improvement of 2.7%, especially excelling in short-text detection, highlighting its effectiveness in limited information environments.
Significance
This work advances AI-generated text detection by introducing a biologically inspired, source-sensitive mechanism that does not rely on training data. It addresses key limitations of existing methods, such as sensitivity to short texts and localized modifications, providing a scalable, robust solution suitable for large-scale deployment. Its interpretability and efficiency make it promising for real-time content moderation, helping mitigate misinformation, protect intellectual property, and ensure content integrity in the era of powerful language models.
Technical Contribution
The primary innovation lies in translating biological exon-intron concepts into NLP, leveraging hidden-state discrepancies to identify source-relevant tokens. The method combines nonlinear importance mapping with probabilistic ratio scores and a mutation-repair mechanism, creating a novel, training-free detection framework. This approach offers theoretical robustness, interpretability, and adaptability across diverse models and datasets, setting a new standard for source-aware AI detection.
Novelty
This is the first work to incorporate the exon-intron analogy for token importance in AI detection, using hidden-state discrepancies to identify source-specific tokens. Unlike prior methods that rely solely on probability statistics or averaging, this approach emphasizes fine-grained, interpretable source signals, providing a new paradigm for training-free detection with enhanced robustness and explainability.
Limitations
- The effectiveness depends on the choice of proxy models and their parameter settings; overly similar or dissimilar models may reduce discrepancy signals, impacting accuracy.
- In extremely short texts or highly complex contexts, hidden-state differences may be insufficient for reliable exonic token identification, leading to potential misclassification.
- Current validation is limited to English models; cross-lingual and multimodal scenarios require further testing and adaptation.
Future Work
Future research will explore multi-model fusion strategies to stabilize exonic token detection, extend the framework to multimodal content (images, videos), and refine the theoretical understanding of discrepancy measures. Additionally, efforts will focus on cross-lingual robustness and reducing computational overhead for real-time large-scale deployment.
AI Executive Summary
The rapid development of large language models (LLMs) like GPT-4 and LLaMA has blurred the line between human and AI-generated text, raising societal concerns such as misinformation, authorship ambiguity, and intellectual property risks. Existing detection methods, especially those relying on supervised training, struggle to generalize across domains, lengths, and adversarial manipulations. Unsupervised, training-free approaches like LogLikelihood and DetectGPT offer flexibility but often lack robustness, particularly against localized modifications and short texts.
In response, this paper introduces Exons-Detect, a novel training-free detection framework inspired by biological exon-intron structures. By employing a dual-proxy LLM setup, it measures hidden-state discrepancies at the token level to identify 'exonic' tokens—those carrying strong source signals. These tokens are then amplified through a nonlinear importance mapping, and combined with probabilistic scores like log-perplexity ratios and a mutation-repair mechanism to produce an interpretable translation score. This score effectively distinguishes AI-generated from human-written texts.
Extensive experiments on datasets such as DetectRL, RealDet, and M4 demonstrate that Exons-Detect achieves state-of-the-art performance, with an average AUROC of over 92%, surpassing existing methods by 1.4-2.2%. It maintains high robustness under adversarial attacks, including paraphrasing and polishing, and performs well across varying text lengths, especially in short-text scenarios. Its efficiency—processing each sample in under 0.8 seconds—makes it suitable for real-time deployment.
The approach's core innovation lies in leveraging hidden-state discrepancies to identify source-relevant tokens, providing a fine-grained, interpretable, and robust detection mechanism. This work paves the way for scalable, unsupervised AI content verification, addressing critical societal needs. Future directions include multi-model fusion, multimodal extension, and theoretical analysis to further enhance detection capabilities and applicability across languages and content types.
Deep Analysis
Background
近年来，随着GPT-4、LLaMA等大模型的广泛应用，AI生成文本的质量显著提升，逼近人类写作水平。早期方法如OpenAI的RoBERTa分类器、RADAR的对抗训练、Biscope的高维表示映射等，推动了检测技术的发展。然而，这些方法多依赖大量标注数据，面临泛化不足和成本高的问题。无训练检测方法如LogLikelihood、DetectGPT等，通过统计生成概率实现检测，具有灵活性，但在短文本和局部修改场景中表现欠佳。近年来，研究者开始关注模型间差异和隐状态信息，试图突破传统局限。本文提出的Exons-Detect，结合生物学中外显子-内含子概念，利用双模型差异识别关键Token，开启了无训练检测的新思路。
Core Problem
现有检测方法在短文本、局部修改和跨模型场景中表现不稳定，尤其在对抗性攻击和文本长度变化时效果下降明显。训练模型虽能提升准确率，但泛化能力有限，且成本高。无训练方法虽灵活，但难以捕获源信息的细粒度特征，导致检测效果受限。如何在无需训练的情况下，准确识别出源信息丰富的关键Token，成为亟待解决的问题。特别是在复杂、多样的文本环境中，利用模型间差异提取源相关特征，仍存在技术难题。这限制了检测技术的广泛应用。
Innovation
第一，提出“外显子”Token识别机制，将生物学中的外显子-内含子概念引入文本检测，强调Token贡献的差异性。第二，利用双模型隐藏状态差异，采用余弦距离作为指标，识别出源信息丰富的关键Token，增强判别能力。第三，设计非线性映射g(·)，动态调整外显子Token的权重，避免过度强调，提高鲁棒性。第四，结合生成概率比值（log-perplexity与cross-perplexity）和突变修复机制，构建可解释的翻译得分，显著提升检测准确率。这一体系突破了传统统计和单模型方法的局限，提供了新颖的无训练检测思路。
Methodology
- �� 输入：文本序列，通过两个Proxy LLM模型（M和˜M）提取每个Token的隐藏表示和生成概率。
- �� 特征提取：在每层提取隐藏状态，计算余弦距离，衡量模型间差异。
- �� 识别外显子Token：将差异超过阈值θ的Token标记为外显子。
- �� 权重映射：利用非线性函数g(·)将差异映射为额外权重，结合初始均匀权重，归一化。
- �� 得分计算：利用加权的log-perplexity与cross-perplexity比值，结合突变修复机制，生成可解释的检测得分。
- �� 判定：将得分与阈值比较，输出是否为AI生成文本。
Experiments
采用M4、RealDet和DetectRL三大公开数据集，比较多种训练和无训练检测方法。指标包括AUROC和F1。模型均使用Falcon-7B-Instruct作为Proxy模型，阈值设为0.15，映射参数α为10。对抗攻击包括DIPPER改写和GPT-4润色，测试不同文本长度（50-240Token），验证鲁棒性。通过消融实验分析非线性映射和突变修复机制的贡献。
Results
在DetectRL上，Exons-Detect实现92.14%的平均AUROC，优于最优基线DNA-DetectLLM的90.0%；在抗对抗攻击中，性能下降不到2%，表现出极强鲁棒性。短文本检测中，平均提升2.7%，显示其在信息有限环境下的优势。多场景、多模型验证表明，该方法具有良好的泛化能力和稳定性，显著优于传统统计方法和训练模型。
Applications
可应用于内容平台、新闻审核、学术诚信和版权保护等场景，实现大规模实时检测。无需大量标注数据，部署简便，适合多模型、多域环境。未来可结合多模态信息，扩展到图像、视频内容的源识别，提升整体内容安全水平。
Limitations & Outlook
依赖Proxy模型的选择和参数调节，模型差异过大或过小可能影响识别效果。对极端短文本或复杂语境下的识别仍存在误差。当前主要验证于英文模型，跨语言和多模态场景的适应性有待验证。未来需优化模型差异度量和多模态融合策略。
Plain Language Accessible to non-experts
想象你在厨房做菜，厨师用不同的刀切菜。普通刀切得快但不精细，关键的刀（像“外显子”）能切出特别漂亮的片。现在，厨师用两个不同的刀（模型）切菜，观察切出来的片子差异大不大。差异大说明这片菜很特别，可能是关键的部分。我们用一种特殊的“放大镜”把这些关键片放大，让厨师更容易找到重要的菜片。这样一来，无论菜多复杂，厨师都能准确找到最重要的部分，避免被假菜蒙骗。这个方法就像用两个不同的刀和放大镜，帮我们识别出真正的“宝贝菜片”。
ELI14 Explained like you're 14
想象你在玩拼图游戏，有些拼图块特别重要，决定拼图能不能成功。你用两个不同的拼图专家（模型）来看每个拼图块，发现有些块在两个专家手里差异很大，说明它们特别重要。你用一种特殊的放大镜，把这些重要的块放大，让拼图变得更清楚。这样一来，不管拼图有多复杂，你都能找到那些关键的块，避免被假块骗到。这个方法就像用两个不同的眼睛和放大镜，帮你找到拼图的核心部分，确保拼得又快又准。
Abstract
The rapid advancement of large language models has increasingly blurred the boundary between human-written and AI-generated text, raising societal risks such as misinformation dissemination, authorship ambiguity, and threats to intellectual property rights. These concerns highlight the urgent need for effective and reliable detection methods. While existing training-free approaches often achieve strong performance by aggregating token-level signals into a global score, they typically assume uniform token contributions, making them less robust under short sequences or localized token modifications. To address these limitations, we propose Exons-Detect, a training-free method for AI-generated text detection based on an exon-aware token reweighting perspective. Exons-Detect identifies and amplifies informative exonic tokens by measuring hidden-state discrepancy under a dual-model setting, and computes an interpretable translation score from the resulting importance-weighted token sequence. Empirical evaluations demonstrate that Exons-Detect achieves state-of-the-art detection performance and exhibits strong robustness to adversarial attacks and varying input lengths. In particular, it attains a 2.2\% relative improvement in average AUROC over the strongest prior baseline on DetectRL.