The rapid advancement of large language models (LLMs) has made detecting AI-generated text an increasingly critical challenge. Traditional methods often fail to capture the nuanced semantic differences between human and machine-generated content. We therefore propose a novel approach based on semantic similarity analysis, leveraging a multi-layered architecture that combines a pre-trained DeBERTa-v3-large model, Bi-directional LSTMs, and linear attention pooling to capture both local and global semantic patterns. To enhance performance, we employ advanced input and output augmentation techniques such as sector-level context integration and wide output configurations. These techniques enable the model to learn more discriminative features and generalize across diverse domains. Experimental results show that this approach works better than traditional methods, proving its usefulness for AI-generated text detection and other text comparison tasks.
We present the GenAI Content Detection Task~1 -- a shared task on binary machine generated text detection, conducted as a part of the GenAI workshop at COLING 2025. The task consists of two subtasks: Monolingual (English) and Multilingual. The shared task attracted many participants: 36 teams made official submissions to the Monolingual subtask during the test phase and 26 teams -- to the Multilingual. We provide a comprehensive overview of the data, a summary of the results -- including system rankings and performance scores -- detailed descriptions of the participating systems, and an in-depth analysis of submissions. https://github.com/mbzuai-nlp/COLING-2025-Workshop-on-MGT-Detection-Task1
Recent work shows membership inference attacks (MIAs) on large language models (LLMs) produce inconclusive results, partly due to difficulties in creating non-member datasets without temporal shifts. While researchers have turned to synthetic data as an alternative, we show this approach can be fundamentally misleading. Our experiments indicate that MIAs function as machine-generated text detectors, incorrectly identifying synthetic data as training samples regardless of the data source. This behavior persists across different model architectures and sizes, from open-source models to commercial ones such as GPT-3.5. Even synthetic text generated by different, potentially larger models is classified as training data by the target model. Our findings highlight a serious concern: using synthetic data in membership evaluations may lead to false conclusions about model memorization and data leakage. We caution that this issue could affect other evaluations using model signals such as loss where synthetic or machine-generated translated data substitutes for real-world samples.
This paper presents a system developed for Task 1 of the COLING 2025 Workshop on Detecting AI-Generated Content, focusing on the binary classification of machine-generated versus human-written text. Our approach utilizes an ensemble of models, with weights assigned according to each model's inverse perplexity, to enhance classification accuracy. For the English text detection task, we combined RoBERTa-base, RoBERTa-base with the OpenAI detector, and BERT-base-cased, achieving a Macro F1-score of 0.7458, which ranked us 12th out of 35 teams. We ensembled RemBERT, XLM-RoBERTa-base, and BERT-base-multilingual-case for the multilingual text detection task, employing the same inverse perplexity weighting technique. This resulted in a Macro F1-score of 0.7513, positioning us 4th out of 25 teams. Our results demonstrate the effectiveness of inverse perplexity weighting in improving the robustness of machine-generated text detection across both monolingual and multilingual settings, highlighting the potential of ensemble methods for this challenging task.
This paper describes the approach of the Unibuc - NLP team in tackling the Coling 2025 GenAI Workshop, Task 1: Binary Multilingual Machine-Generated Text Detection. We explored both masked language models and causal models. For Subtask A, our best model achieved first-place out of 36 teams when looking at F1 Micro (Auxiliary Score) of 0.8333, and second-place when looking at F1 Macro (Main Score) of 0.8301
Purpose: This study aims to evaluate different setups of an AI-based solution to detect Protected Health Information (PHI) in medical images. Materials and Methods: Text from eight PHI and eight non-PHI categories are simulated and incorporated into a curated dataset comprising 1,000 medical images across four modalities: CT, X-ray, bone scan, and MRI. The proposed PHI detection pipeline comprises three key components: text localization, extraction, and analysis. Three vision and language models, YOLOv11, EasyOCR, and GPT-4o, are benchmarked in different setups corresponding to three key components. The performance is evaluated with classification metrics, including precision, recall, F1 score, and accuracy. Results: All four setups demonstrate strong performance in detecting PHI imprints, with all metrics exceeding 0.9. The setup that utilizes YOLOv11 for text localization and GPT-4o for text extraction and analysis achieves the highest performance in PHI detection. However, this setup incurs the highest cost due to the increased number of generated tokens associated with GPT-4o model. Conversely, the setup using solely GPT-4o for the end-to-end pipeline exhibits the lowest performance but showcases the feasibility of multi-modal models in solving complex tasks. Conclusion: For optimal text localization and extraction, it is recommended to fine-tune an object detection model and utilize built-in Optical Character Recognition (OCR) software. Large language models like GPT-4o can be effectively leveraged to reason about and semantically analyze the PHI content. Although the vision capability of GPT-4o is promising for reading image crops, it remains limited for end-to-end pipeline applications with whole images.
Recently there have been many shared tasks targeting the detection of generated text from Large Language Models (LLMs). However, these shared tasks tend to focus either on cases where text is limited to one particular domain or cases where text can be from many domains, some of which may not be seen during test time. In this shared task, using the newly released RAID benchmark, we aim to answer whether or not models can detect generated text from a large, yet fixed, number of domains and LLMs, all of which are seen during training. Over the course of three months, our task was attempted by 9 teams with 23 detector submissions. We find that multiple participants were able to obtain accuracies of over 99% on machine-generated text from RAID while maintaining a 5% False Positive Rate -- suggesting that detectors are able to robustly detect text from many domains and models simultaneously. We discuss potential interpretations of this result and provide directions for future research.
The rapid advancement in large language models (LLMs) has significantly enhanced their ability to generate coherent and contextually relevant text, raising concerns about the misuse of AI-generated content and making it critical to detect it. However, the task remains challenging, particularly in unseen domains or with unfamiliar LLMs. Leveraging LLM next-token distribution outputs offers a theoretically appealing approach for detection, as they encapsulate insights from the models' extensive pre-training on diverse corpora. Despite its promise, zero-shot methods that attempt to operationalize these outputs have met with limited success. We hypothesize that one of the problems is that they use the mean to aggregate next-token distribution metrics across tokens, when some tokens are naturally easier or harder to predict and should be weighted differently. Based on this idea, we propose the Perplexity Attention Weighted Network (PAWN), which uses the last hidden states of the LLM and positions to weight the sum of a series of features based on metrics from the next-token distribution across the sequence length. Although not zero-shot, our method allows us to cache the last hidden states and next-token distribution metrics on disk, greatly reducing the training resource requirements. PAWN shows competitive and even better performance in-distribution than the strongest baselines (fine-tuned LMs) with a fraction of their trainable parameters. Our model also generalizes better to unseen domains and source models, with smaller variability in the decision boundary across distribution shifts. It is also more robust to adversarial attacks, and if the backbone has multilingual capabilities, it presents decent generalization to languages not seen during supervised training, with LLaMA3-1B reaching a mean macro-averaged F1 score of 81.46% in cross-validation with nine languages.
Generating visual text in natural scene images is a challenging task with many unsolved problems. Different from generating text on artificially designed images (such as posters, covers, cartoons, etc.), the text in natural scene images needs to meet the following four key criteria: (1) Fidelity: the generated text should appear as realistic as a photograph and be completely accurate, with no errors in any of the strokes. (2) Reasonability: the text should be generated on reasonable carrier areas (such as boards, signs, walls, etc.), and the generated text content should also be relevant to the scene. (3) Utility: the generated text can facilitate to the training of natural scene OCR (Optical Character Recognition) tasks. (4) Controllability: The attribute of the text (such as font and color) should be controllable as needed. In this paper, we propose a two stage method, SceneVTG++, which simultaneously satisfies the four aspects mentioned above. SceneVTG++ consists of a Text Layout and Content Generator (TLCG) and a Controllable Local Text Diffusion (CLTD). The former utilizes the world knowledge of multi modal large language models to find reasonable text areas and recommend text content according to the nature scene background images, while the latter generates controllable multilingual text based on the diffusion model. Through extensive experiments, we respectively verified the effectiveness of TLCG and CLTD, and demonstrated the state-of-the-art text generation performance of SceneVTG++. In addition, the generated images have superior utility in OCR tasks like text detection and text recognition. Codes and datasets will be available.
Verifying the provenance of content is crucial to the function of many organizations, e.g., educational institutions, social media platforms, firms, etc. This problem is becoming increasingly challenging as text generated by Large Language Models (LLMs) becomes almost indistinguishable from human-generated content. In addition, many institutions utilize in-house LLMs and want to ensure that external, non-sanctioned LLMs do not produce content within the institution. In this paper, we answer the following question: Given a piece of text, can we identify whether it was produced by a particular LLM or not? We model LLM-generated text as a sequential stochastic process with complete dependence on history. We then design zero-shot statistical tests to (i) distinguish between text generated by two different known sets of LLMs $A$ (non-sanctioned) and $B$ (in-house), and (ii) identify whether text was generated by a known LLM or generated by any unknown model, e.g., a human or some other language generation process. We prove that the type I and type II errors of our test decrease exponentially with the length of the text. For that, we show that if $B$ generates the text, then except with an exponentially small probability in string length, the log-perplexity of the string under $A$ converges to the average cross-entropy of $B$ and $A$. We then present experiments using LLMs with white-box access to support our theoretical results and empirically examine the robustness of our results to black-box settings and adversarial attacks. In the black-box setting, our method achieves an average TPR of 82.5\% at a fixed FPR of 5\%. Under adversarial perturbations, our minimum TPR is 48.6\% at the same FPR threshold. Both results outperform all non-commercial baselines. See https://github.com/TaraRadvand74/llm-text-detection for code, data, and an online demo of the project.
The recent emergence of the Segment Anything Model (SAM) enables various domain-specific segmentation tasks to be tackled cost-effectively by using bounding boxes as prompts. However, in scene text segmentation, SAM can not achieve desirable performance. The word-level bounding box as prompts is too coarse for characters, while the character-level bounding box as prompts suffers from over-segmentation and under-segmentation issues. In this paper, we propose an automatic annotation pipeline named Char-SAM, that turns SAM into a low-cost segmentation annotator with a Character-level visual prompt. Specifically, leveraging some existing text detection datasets with word-level bounding box annotations, we first generate finer-grained character-level bounding box prompts using the Character Bounding-box Refinement CBR module. Next, we employ glyph information corresponding to text character categories as a new prompt in the Character Glyph Refinement (CGR) module to guide SAM in producing more accurate segmentation masks, addressing issues of over-segmentation and under-segmentation. These modules fully utilize the bbox-to-mask capability of SAM to generate high-quality text segmentation annotations automatically. Extensive experiments on TextSeg validate the effectiveness of Char-SAM. Its training-free nature also enables the generation of high-quality scene text segmentation datasets from real-world datasets like COCO-Text and MLT17.
This study demonstrates that the modern generation of Large Language Models (LLMs, such as GPT-4) suffers from the same out-of-domain (OOD) performance gap observed in prior research on pre-trained Language Models (PLMs, such as BERT). We demonstrate this across two non-topical classification tasks: 1) genre classification and 2) generated text detection. Our results show that when demonstration examples for In-Context Learning (ICL) come from one domain (e.g., travel) and the system is tested on another domain (e.g., history), classification performance declines significantly. To address this, we introduce a method that controls which predictive indicators are used and which are excluded during classification. For the two tasks studied here, this ensures that topical features are omitted, while the model is guided to focus on stylistic rather than content-based attributes. This approach reduces the OOD gap by up to 20 percentage points in a few-shot setup. Straightforward Chain-of-Thought (CoT) methods, used as the baseline, prove insufficient, while our approach consistently enhances domain transfer performance.
In the digital age of today, the internet has become an indispensable platform for people's lives, work, and information exchange. However, the problem of violent text proliferation in the network environment has arisen, which has brought about many negative effects. In view of this situation, it is particularly important to build an effective system for cutting off violent text. The study of violent text cutting off based on the BERT-fasttext model has significant meaning. BERT is a pre-trained language model with strong natural language understanding ability, which can deeply mine and analyze text semantic information; Fasttext itself is an efficient text classification tool with low complexity and good effect, which can quickly provide basic judgments for text processing. By combining the two and applying them to the system for cutting off violent text, on the one hand, it can accurately identify violent text, and on the other hand, it can efficiently and reasonably cut off the content, preventing harmful information from spreading freely on the network. Compared with the single BERT model and fasttext, the accuracy was improved by 0.7% and 0.8%, respectively. Through this model, it is helpful to purify the network environment, maintain the health of network information, and create a positive, civilized, and harmonious online communication space for netizens, driving the development of social networking, information dissemination, and other aspects in a more benign direction.
Recently, tampered text detection has attracted increasing attention due to its essential role in information security. Although existing methods can detect the tampered text region, the interpretation of such detection remains unclear, making the prediction unreliable. To address this problem, we propose to explain the basis of tampered text detection with natural language via large multimodal models. To fill the data gap for this task, we propose a large-scale, comprehensive dataset, ETTD, which contains both pixel-level annotations for tampered text region and natural language annotations describing the anomaly of the tampered text. Multiple methods are employed to improve the quality of the proposed data. For example, elaborate queries are introduced to generate high-quality anomaly descriptions with GPT4o. A fused mask prompt is proposed to reduce confusion when querying GPT4o to generate anomaly descriptions. To automatically filter out low-quality annotations, we also propose to prompt GPT4o to recognize tampered texts before describing the anomaly, and to filter out the responses with low OCR accuracy. To further improve explainable tampered text detection, we propose a simple yet effective model called TextSleuth, which achieves improved fine-grained perception and cross-domain generalization by focusing on the suspected region, with a two-stage analysis paradigm and an auxiliary grounding prompt. Extensive experiments on both the ETTD dataset and the public dataset have verified the effectiveness of the proposed methods. In-depth analysis is also provided to inspire further research. Our dataset and code will be open-source.
Connected component (CC) is a proper text shape representation that aligns with human reading intuition. However, CC-based text detection methods have recently faced a developmental bottleneck that their time-consuming post-processing is difficult to eliminate. To address this issue, we introduce an explicit relational reasoning network (ERRNet) to elegantly model the component relationships without post-processing. Concretely, we first represent each text instance as multiple ordered text components, and then treat these components as objects in sequential movement. In this way, scene text detection can be innovatively viewed as a tracking problem. From this perspective, we design an end-to-end tracking decoder to achieve a CC-based method dispensing with post-processing entirely. Additionally, we observe that there is an inconsistency between classification confidence and localization quality, so we propose a Polygon Monte-Carlo method to quickly and accurately evaluate the localization quality. Based on this, we introduce a position-supervised classification loss to guide the task-aligned learning of ERRNet. Experiments on challenging benchmarks demonstrate the effectiveness of our ERRNet. It consistently achieves state-of-the-art accuracy while holding highly competitive inference speed.
Large Language Models (LLMs) have revolutionized text generation, making detecting machine-generated text increasingly challenging. Although past methods have achieved good performance on detecting pure machine-generated text, those detectors have poor performance on distinguishing machine-revised text (rewriting, expansion, and polishing), which can have only minor changes from its original human prompt. As the content of text may originate from human prompts, detecting machine-revised text often involves identifying distinctive machine styles, e.g., worded favored by LLMs. However, existing methods struggle to detect machine-style phrasing hidden within the content contributed by humans. We propose the "Imitate Before Detect" (ImBD) approach, which first imitates the machine-style token distribution, and then compares the distribution of the text to be tested with the machine-style distribution to determine whether the text has been machine-revised. To this end, we introduce style preference optimization (SPO), which aligns a scoring LLM model to the preference of text styles generated by machines. The aligned scoring model is then used to calculate the style-conditional probability curvature (Style-CPC), quantifying the log probability difference between the original and conditionally sampled texts for effective detection. We conduct extensive comparisons across various scenarios, encompassing text revisions by six LLMs, four distinct text domains, and three machine revision types. Compared to existing state-of-the-art methods, our method yields a 13% increase in AUC for detecting text revised by open-source LLMs, and improves performance by 5% and 19% for detecting GPT-3.5 and GPT-4o revised text, respectively. Notably, our method surpasses the commercially trained GPT-Zero with just $1,000$ samples and five minutes of SPO, demonstrating its efficiency and effectiveness.
Advanced large language models (LLMs) can generate text almost indistinguishable from human-written text, highlighting the importance of LLM-generated text detection. However, current zero-shot techniques face challenges as white-box methods are restricted to use weaker open-source LLMs, and black-box methods are limited by partial observation from stronger proprietary LLMs. It seems impossible to enable white-box methods to use proprietary models because API-level access to the models neither provides full predictive distributions nor inner embeddings. To traverse the divide, we propose Glimpse, a probability distribution estimation approach, predicting the full distributions from partial observations. Despite the simplicity of Glimpse, we successfully extend white-box methods like Entropy, Rank, Log-Rank, and Fast-DetectGPT to latest proprietary models. Experiments show that Glimpse with Fast-DetectGPT and GPT-3.5 achieves an average AUROC of about 0.95 in five latest source models, improving the score by 51% relative to the remaining space of the open source baseline (Table 1). It demonstrates that the latest LLMs can effectively detect their own outputs, suggesting that advanced LLMs may be the best shield against themselves.
The ability of large language models to generate complex texts allows them to be widely integrated into many aspects of life, and their output can quickly fill all network resources. As the impact of LLMs grows, it becomes increasingly important to develop powerful detectors for the generated text. This detector is essential to prevent the potential misuse of these technologies and to protect areas such as social media from the negative effects of false content generated by LLMS. The main goal of LLM-generated text detection is to determine whether text is generated by an LLM, which is a basic binary classification task. In our work, we mainly use three different classification methods based on open source datasets: traditional machine learning techniques such as logistic regression, k-means clustering, Gaussian Naive Bayes, support vector machines, and methods based on converters such as BERT, and finally algorithms that use LLMs to detect LLM-generated text. We focus on model generalization, potential adversarial attacks, and accuracy of model evaluation. Finally, the possible research direction in the future is proposed, and the current experimental results are summarized.
Natural Language Generation has been rapidly developing with the advent of large language models (LLMs). While their usage has sparked significant attention from the general public, it is important for readers to be aware when a piece of text is LLM-generated. This has brought about the need for building models that enable automated LLM-generated text detection, with the aim of mitigating potential negative outcomes of such content. Existing LLM-generated detectors show competitive performances in telling apart LLM-generated and human-written text, but this performance is likely to deteriorate when paraphrased texts are considered. In this study, we devise a new data collection strategy to collect Human & LLM Paraphrase Collection (HLPC), a first-of-its-kind dataset that incorporates human-written texts and paraphrases, as well as LLM-generated texts and paraphrases. With the aim of understanding the effects of human-written paraphrases on the performance of state-of-the-art LLM-generated text detectors OpenAI RoBERTa and watermark detectors, we perform classification experiments that incorporate human-written paraphrases, watermarked and non-watermarked LLM-generated documents from GPT and OPT, and LLM-generated paraphrases from DIPPER and BART. The results show that the inclusion of human-written paraphrases has a significant impact of LLM-generated detector performance, promoting TPR@1%FPR with a possible trade-off of AUROC and accuracy.
Texts on the intelligent transportation scene include mass information. Fully harnessing this information is one of the critical drivers for advancing intelligent transportation. Unlike the general scene, detecting text in transportation has extra demand, such as a fast inference speed, except for high accuracy. Most existing real-time text detection methods are based on the shrink mask, which loses some geometry semantic information and needs complex post-processing. In addition, the previous method usually focuses on correct output, which ignores feature correction and lacks guidance during the intermediate process. To this end, we propose an efficient multi-scene text detector that contains an effective text representation similar mask (SM) and a feature correction module (FCM). Unlike previous methods, the former aims to preserve the geometric information of the instances as much as possible. Its post-progressing saves 50$\%$ of the time, accurately and efficiently reconstructing text contours. The latter encourages false positive features to move away from the positive feature center, optimizing the predictions from the feature level. Some ablation studies demonstrate the efficiency of the SM and the effectiveness of the FCM. Moreover, the deficiency of existing traffic datasets (such as the low-quality annotation or closed source data unavailability) motivated us to collect and annotate a traffic text dataset, which introduces motion blur. In addition, to validate the scene robustness of the SM-Net, we conduct experiments on traffic, industrial, and natural scene datasets. Extensive experiments verify it achieves (SOTA) performance on several benchmarks. The code and dataset are available at: \url{https://github.com/fengmulin/SMNet}.