Across several benchmarks, it is shown that a plain, fully fine-tuned RoBERTa matches or exceeds the specialized detectors those benchmarks are built around, suggesting that progress in AI-generated text detection should be measured not only by in-distribution performance, but also by robustness under distribution shift.
Abstract
Recent AI-generated text detection work often introduces a new benchmark together with a specialized detector tailored to it. We revisit this practice from a baseline-first perspective. Across several benchmarks, we show that a plain, fully fine-tuned RoBERTa matches or exceeds the specialized detectors those benchmarks are built around. This suggests that much of the recent architectural complexity is not what drives strong in-distribution detection. The remaining challenge is the distribution shift. The same strong baseline degrades sharply when the topic domain or generating model changes at test time, and simply adding more source data does not close the gap. We identify a key failure mode: under distribution shift, the detector can assign high-confidence machine labels to human-written text from unseen domains. We then study two lightweight domain adaptation methods to address this problem: $K$-shot adaptation with first-order MAML over LoRA adapters, and a per-sample confidence-weighted ensemble built on top of the adapted detector. Overall, our results suggest that progress in AI-generated text detection should be measured not only by in-distribution performance, but also by robustness under distribution shift.
This work analyses the linearity and quality of MGT representations and shows that simple linear probes outperform a wide range of detectors while being substantially more sample-efficient, and demonstrates the potential of linear probes as as robust and sample-efficient MGT detectors.
Gerrit Quaremba, Hanqi Yan, E. Black et al.· 0 citations
It is shown that the strength of this signal depends on text length and sampling range: spectral evidence is clearest for long, continuous, constrained generation, while short, fragmented, mixed, and edited settings require complementary confidence and fluctuation views.
Haitong Luo, Xuying Meng, Weiyao Zhang et al.· 1 citation
Distinguishing Large Language Model (LLM) generated text from human writing is a critical and difficult challenge. While LLMs are trained to write like humans, we hypothesize that this training leaves an indelible mark. LLMs develop a particularly strong aversion to token repetition very early in training. This bias persists as a''Vestigial Heuristic''(a developmental artifact) that is activated in LLM-generated text, separating LLM from human writing. To probe this phenomenon, we introduce Telescope Perplexity, a metric that evaluates the token repetition of the model, $P(s_i | s_{1:i})$ . Our empirical investigation reveals that the Telescope Perplexity signature emerges early in pre-training, and Telescope Perplexity empirically enables highly effective zero-shot LLM detection. We show state-of-the-art or competitive performance across diverse datasets (including modern evaluation sets we introduce), reference models, and perturbation schemes with greater efficiency than other methods.
Large Language Models (LLMs) have been rapidly evolving lately, resulting in the need for strong, explainable models to detect the difference between human-generated and machine-generated articles. Existing approaches which are mostly based on fine-tuned transformers suffer from several drawbacks such as rapid obsolescence, paraphrasing attacks, and lack of interpretability. To improve their ability to detect, this paper proposes a novel paradigm called Human vs. LLM Identification (HLI) which introduces a Retrieval-Augmented Generation (RAG)-inspired evidence-based detection strategy alongside a fine-tuned transformer classifier. Our core model, DeBERTa-Sentinel, is built on top of a fine-tuned Microsoft DeBERTa-v3-small model, which uses a disentangled attention mechanism to better capture subtle syntactic and stylistic deviations characteristic of AI-generated text. We evaluate our framework on a balanced dataset of 43,456 text samples, curated from the OpenGPTText corpus and covering AI-generated and human-authored content across diverse domains including news, education, and creative text. The experimental results show improved performance over the selected baselines, with our framework achieving an accuracy of 97.53%, precision of 95.89%, recall of 99.34%, and ROC-AUC of 99.53%. In addition, explainability is integrated into our framework through Local Interpretable Model-agnostic Explanations (LIME) analysis, providing token-level insight into classification decisions. This study establishes a benchmark for scalable, explainable AI text detection, with implications for academic integrity, content moderation, and combating misinformation.
Ibtasam Ur Rehman, Muhammad Islam, Muhammad Yousaf Rehman et al.· Knowledge· 0 citations
The key idea is to smooth adjacent token scores to reduce their variability, while using an adaptive Lepski-type rule to select the bandwidth according to the local authorship structure, and the proposed method achieves favorable mean square error performance in estimating the underlying signal.
Yangjun Lu, Hongyi Zhou, Fabian Spill et al.· 0 citations
The rapid proliferation of digitally distributed news has made large-scale automated verification an urgent research priority, as deceptive content spreads across social platforms faster than manual reviewers can evaluate it. This work introduces a two-stage deep learning pipeline in which a pretrained BERT encoder generates rich contextual token embeddings that are subsequently processed by a Bidirectional Long Short-Term Memory (BiLSTM) network, enabling the classifier to exploit both sentence-level semantics and document-level narrative flow simultaneously. Experiments on the WELFake benchmark [1] following a standardised cleaning pipeline—lowercasing, URL stripping, tokenisation, stop-word removal, and WordNet lemmatisation— yielded 98.7 % accuracy and an F1-score of 0.986. This surpasses a BERT-only baseline by 3.1 percentage points and exceeds several previously published state-of-the-art results. The gains confirm that coupling transformer-based semantic representations with recurrent sequential modelling produces a measurable and reproducible improvement in misinformation detection. Planned extensions include multilingual evaluation and knowledge-distilled encoder replacements targeting real-time throughput constraints.
Ardra P Namboodiri, Archa P S, Honey Mol O· 2026 6th International Confe...· 0 citations