Large Language Models (LLMs) possess a theoretical capability to model information density far beyond the limits of classical statistical methods (e.g., Lempel-Ziv). However, utilizing this capability for lossless compression involves navigating severe system constraints, including non-deterministic hardware and prohibitive computational costs. In this work, we present an exploratory study into the feasibility of LLM-based archival systems. We introduce Hybrid-LLM, a proof-of-concept architecture designed to investigate the "entropic capacity" of foundation models in a storage context. We identify a critical barrier to deployment: the "GPU Butterfly Effect," where microscopic hardware non-determinism precludes data recovery. We resolve this via a novel logit quantization protocol, enabling the rigorous measurement of neural compression rates on real-world data. Our experiments reveal a distinct divergence between "retrieval-based" density (0.39 BPC on memorized literature) and "predictive" density (0.75 BPC on unseen news). While current inference latency (≈ 2600× slower than Zstd) limits immediate deployment to ultra-cold storage, our findings demonstrate that LLMs successfully capture semantic redundancy inaccessible to classical algorithms, establishing a baseline for future research into semantic file systems.

We present a feedforward graph architecture in which heterogeneous frozen large language models serve as computational nodes, communicating through a shared continuous latent space via learned linear projections. Building on recent work demonstrating geometric compatibility between independently trained LLM latent spaces , we extend this finding from static two-model steering to end-to-end trainable multi-node graphs, where projection matrices are optimized jointly via backpropagation through residual stream injection hooks. Three small frozen models (Llama-3.2-1B, Qwen2.5-1.5B, Gemma-2-2B) encode the input into a shared latent space whose aggregate signal is injected into two larger frozen models (Phi-3-mini, Mistral-7B), whose representations feed a lightweight cross-attention output node. With only 17.6M trainable parameters against approximately 12B frozen, the architecture achieves 87.3% on ARC-Challenge, 82.8% on OpenBookQA, and 67.2% on MMLU, outperforming the best single constituent model by 11.4, 6.2, and 1.2 percentage points respectively, and outperforming parameter-matched learned classifiers on frozen single models by 9.1, 5.2, and 6.7 points. Gradient flow through multiple frozen model boundaries is empirically verified to be tractable, and the output node develops selective routing behavior across layer-2 nodes without explicit supervision.

Do large language models contain domain-specific parametric shells: concentrated, causally necessary neuron populations whose removal selectively degrades a target domain while sparing others? We apply a uniform causal methodology across two domain granularities, three model families (1.5B to 7B parameters), and eight domains. At the academic subject level, zero neurons exceed 60% domain selectivity across 939,008 combined FFN neurons and causal damage matrices are flat, despite domain identity being linearly decodable above 85% accuracy. At the language and modality level, 0.65–1.14% of neurons exceed 60% selectivity, damage matrices are near-perfectly diagonal (ratios up to 595:1), and shell neuron sets are essentially disjoint (IoU < 0.003). Masking code-selective neurons reduces mathematical reasoning accuracy by 16–24 percentage points across all models; masking Spanish or Chinese neurons leaves it at or below random. Shell strength increases monotonically with scale and shells are spatially interleaved in a pattern that precludes group-level selective quantization. Parametric shells form where and only where training data was modular at the token level.

Large language models (LLMs) reason over discrete token ID sequences, yet modern subword tokenizers routinely produce non-unique encodings: multiple token ID sequences can detokenize to identical surface strings. This representational mismatch creates an unmeasured fragility wherein reasoning processes can fail. LLMs may treat two internal representations as distinct "words" even when they are semantically identical at the text level. In this work, we show that tokenization can betray LLM reasoning through one-to-many token ID mappings. We introduce a tokenization-consistency probe that requires models to replace designated target words in context while leaving all other content unchanged. The task is intentionally simple at the surface level, enabling us to attribute failures to tokenizer-detokenizer artifacts rather than to knowledge gaps or parameter limitations. Through analysis of over 11000 replacement trials across state-of-the-art open-source LLMs, we find a non-trivial rate of outputs exhibit phantom edits: cases where models operate under the illusion of correct reasoning, a phenomenon arising from tokenizer-induced representational defects. We further analyze these cases and provide a taxonomy of eight systematic tokenizer artifacts, including whitespace-boundary shifts and intra-word resegmentation. These findings indicate that part of apparent reasoning deficiency originates in the tokenizer layer, motivating tokenizer-level remedies before incurring the cost of training ever-larger models on ever-larger corpora.

We investigate whether independently trained language models converge to geometrically compatible latent representations, and whether this compatibility can be exploited to correct model behavior at inference time without any weight updates. We learn a linear projection matrix that maps activation vectors from a large teacher model into the coordinate system of a smaller student model, then intervene on the student's residual stream during generation by substituting its internal state with the translated teacher representation. Across a fully crossed experimental matrix of 20 heterogeneous teacher-student pairings spanning mixture-of-experts, dense, code-specialized, and synthetically trained architectures, the Ridge projection consistently achieves R^2 = 0.50 on verbal reasoning and R^2 = 0.40 on mathematical reasoning, collapsing to R^2 = -0.22 under permutation control and R^2 = 0.01 under L_1 regularization. Behavioral correction rates range from 14.0

While large language models (LLMs) exhibit significant utility across various domains, they simultaneously are susceptible to exploitation for unethical purposes, including academic misconduct and dissemination of misinformation. Consequently, AI-generated text detection systems have emerged as a countermeasure. However, these detection mechanisms demonstrate vulnerability to evasion techniques and lack robustness against textual manipulations. This paper introduces back-translation as a novel technique for evading detection, underscoring the need to enhance the robustness of current detection systems. The proposed method involves translating AI-generated text through multiple languages before back-translating to English. We present a model that combines these back-translated texts to produce a manipulated version of the original AI-generated text. Our findings demonstrate that the manipulated text retains the original semantics while significantly reducing the true positive rate (TPR) of existing detection methods. We evaluate this technique on nine AI detectors, including six open-source and three proprietary systems, revealing their susceptibility to back-translation manipulation. In response to the identified shortcomings of existing AI text detectors, we present a countermeasure to improve the robustness against this form of manipulation. Our results indicate that the TPR of the proposed method declines by only 1.85 after back-translation manipulation. Furthermore, we build a large dataset of 720k texts using eight different LLMs. Our dataset contains both human-authored and LLM-generated texts in various domains and writing styles to assess the performance of our method and existing detectors. This dataset is publicly shared for the benefit of the research community.

The local news landscape, a vital source of reliable information for 28 million Americans, faces a growing threat from Pink Slime Journalism, a low-quality, auto-generated articles that mimic legitimate local reporting. Detecting these deceptive articles requires a fine-grained analysis of their linguistic, stylistic, and lexical characteristics. In this work, we conduct a comprehensive study to uncover the distinguishing patterns of Pink Slime content and propose detection strategies based on these insights. Beyond traditional generation methods, we highlight a new adversarial vector: modifications through large language models (LLMs). Our findings reveal that even consumer-accessible LLMs can significantly undermine existing detection systems, reducing their performance by up to 40

We present a novel evaluation paradigm for AI text detectors that prioritizes real-world and equitable assessment. Current approaches predominantly report conventional metrics like AUROC, overlooking that even modest false positive rates constitute a critical impediment to practical deployment of detection systems. Furthermore, real-world deployment necessitates predetermined threshold configuration, making detector stability (i.e. the maintenance of consistent performance across diverse domains and adversarial scenarios), a critical factor. These aspects have been largely ignored in previous research and benchmarks. Our benchmark, SHIELD, addresses these limitations by integrating both reliability and stability factors into a unified evaluation metric designed for practical assessment. Furthermore, we develop a post-hoc, model-agnostic humanification framework that modifies AI text to more closely resemble human authorship, incorporating a controllable hardness parameter. This hardness-aware approach effectively challenges current SOTA zero-shot detection methods in maintaining both reliability and stability. (Data and code: https://github.com/navid-aub/SHIELD-Benchmark)

With the increasing reliance on LLMs as research agents, distinguishing between LLM and human-generated ideas has become crucial for understanding the cognitive nuances of LLMs' research capabilities. While detecting LLM-generated text has been extensively studied, distinguishing human vs LLM-generated scientific idea remains an unexplored area. In this work, we systematically evaluate the ability of state-of-the-art (SOTA) machine learning models to differentiate between human and LLM-generated ideas, particularly after successive paraphrasing stages. Our findings highlight the challenges SOTA models face in source attribution, with detection performance declining by an average of 25.4% after five consecutive paraphrasing stages. Additionally, we demonstrate that incorporating the research problem as contextual information improves detection performance by up to 2.97

LLMs offer valuable capabilities, yet they can be utilized by malicious users to disseminate deceptive information and generate fake news. The growing prevalence of LLMs poses difficulties in crafting detection approaches that remain effective across various text domains. Additionally, the absence of precautionary measures for AI-generated news on online social platforms is concerning. Therefore, there is an urgent need to improve people’s ability to differentiate between news articles written by humans and those produced by LLMs. By providing cues in human-written and LLM-generated news, we can help individuals increase their skepticism towards fake LLM-generated news. This paper aims to elucidate simple markers that help individuals distinguish between articles penned by humans and those created by LLMs. To achieve this, we initially collected a dataset comprising 39k news articles authored by humans or generated by four distinct LLMs with varying degrees of fake. We then devise a metric named Entropy-Shift Authorship Signature (ESAS) based on the information theory and entropy principles. The proposed ESAS ranks terms or entities, like POS tagging, within news articles based on their relevance in discerning article authorship. We demonstrate the effectiveness of our metric by showing the high accuracy attained by a basic method, i.e., TF-IDF combined with logistic regression classifier, using a small set of terms with the highest ESAS score. Consequently, we introduce and scrutinize these top ESAS-ranked terms to aid individuals in strengthening their skepticism towards LLM-generated fake news.

User-generated content like comments are valuable sources for various downstream applications. However, access to user comments data is often limited to specific platforms or outlets, which imposes a great limitation on the available data, and may not provide a representative sample of opinions from a diverse population on a particular event. This paper presents a comment crawling system that leverages the Web API of popular third-party commenting systems to collect comments from a large number of websites integrated with the commenting systems. Given a target page, the crawling system utilizes a deep learning model to extract API parameters and send HTTP requests to the API to retrieve comments. The system, Comquest, that we propose to demo is news-oriented and crawls comments regarding specific news topics/stories. Comquest can work with any website that allows commenting. Comquest provides a useful tool for collecting comments that represent a wider range of opinions, stances, and sentiments from websites on a global scale.

Opinion phrase extraction is one of the key tasks in fine-grained sentiment analysis. While opinion expressions could be generic subjective expressions, aspect specific opinion expressions contain both the aspect as well as the opinion expression within the original sentence context. In this work, we formulate the task as an instance of token-level sequence labeling. When multiple aspects are present in a sentence, detection of opinion phrase boundary becomes difficult and label of each word depend not only upon the surrounding words but also with the concerned aspect. We propose a neural network architecture with bidirectional LSTM (Bi-LSTM) and a novel attention mechanism. Bi-LSTM layer learns the various sequential pattern among the words without requiring any hand-crafted features. The attention mechanism captures the importance of context words on a particular aspect opinion expression when multiple aspects are present in a sentence via location and content based memory. A Conditional Random Field (CRF) model is incorporated in the final layer to explicitly model the dependencies among the output labels. Experimental results on Hotel dataset from Tripadvisor.com showed that our approach outperformed several state-of-the-art baselines.

Despite repeated measures, YouTube's comment section has been a fertile ground for scammers.With the growth of the cryptocurrency market and obscurity around it, a new form of scam, namely "Collusion Scam" has emerged as a dominant force within YouTube's comment space.Unlike typical scams and spams, collusion scams employ a cunning persuasion strategy, using the facade of genuine social interactions within comment threads to create an aura of trust and success to entrap innocent users.In this research, we collect 1,174 such collusion scam threads and perform a detailed analysis, which is tailored towards the successful detection of these scams.We find that utilization of the collusion dynamics can provide an accuracy of 96.67% and an F1-score of 93.04%.Furthermore, we demonstrate the robust predictive power of metadata associated with these threads and user channels, which act as compelling indicators of collusion scams.Finally, we show that modern LLM, like chatGPT, can effectively detect collusion scams without the need for any training.

Deceptive text poses a significant threat to users, resulting in widespread misinformation and disorder.While researchers have created numerous cutting-edge techniques for detecting deception in domain-specific settings, whether there is a generic deception pattern so that deception-related knowledge in one domain can be transferred to the other remains mostly unexplored.Moreover, the disparities in textual expression across these many mediums pose an additional obstacle for generalization.To this end, we present a Multi-Task Learning (MTL) based deception generalization strategy to reduce the domain-specific noise and facilitate a better understanding of deception via a generalized training.As deceptive domains, we use News (fake news), Tweets (rumors), and Reviews (fake reviews) and employ LSTM and BERT models to incorporate domain transfer techniques.Our proposed architecture for the combined approach of domain-independent and domain-specific training improves the deception detection performance by up to 5.28% in F1-score.

In this paper, we present a novel method for detecting fake and Large Language Model (LLM)-generated profiles in the LinkedIn Online Social Network immediately upon registration and before establishing connections. Early fake profile identification is crucial to maintaining the platform's integrity since it prevents imposters from acquiring the private and sensitive information of legitimate users and from gaining an opportunity to increase their credibility for future phishing and scamming activities. This work uses textual information provided in LinkedIn profiles and introduces the Section and Subsection Tag Embedding (SSTE) method to enhance the discriminative characteristics of these data for distinguishing between legitimate profiles and those created by imposters manually or by using an LLM. Additionally, the dearth of a large publicly available LinkedIn dataset motivated us to collect 3600 LinkedIn profiles for our research. We release our dataset publicly for research purposes. This is, to the best of our knowledge, the first large publicly available LinkedIn dataset for fake LinkedIn account detection. Within our paradigm, we assess static and contextualized word embeddings, including GloVe, Flair, BERT, and RoBERTa. We show that the suggested method can distinguish between legitimate and fake profiles with an accuracy of about 95% across all word embeddings. In addition, we show that SSTE has a promising accuracy for identifying LLM-generated profiles, despite the fact that no LLM-generated profiles were employed during the training phase, and can achieve an accuracy of approximately 90% when only 20 LLM-generated profiles are added to the training set. It is a significant finding since the proliferation of several LLMs in the near future makes it extremely challenging to design a single system that can identify profiles created with various LLMs.

In this paper, we present a novel method for detecting fake and Large Language Model (LLM)-generated profiles in the LinkedIn Online Social Network immediately upon registration and before establishing connections. Early fake profile identification is crucial to maintaining the platform's integrity since it prevents imposters from acquiring the private and sensitive information of legitimate users and from gaining an opportunity to increase their credibility for future phishing and scamming activities. This work uses textual information provided in LinkedIn profiles and introduces the Section and Subsection Tag Embedding (SSTE) method to enhance the discriminative characteristics of these data for distinguishing between legitimate profiles and those created by imposters manually or by using an LLM. Additionally, the dearth of a large publicly available LinkedIn dataset motivated us to collect 3600 LinkedIn profiles for our research. We release our dataset publicly for research purposes. This is, to the best of our knowledge, the first large publicly available LinkedIn dataset for fake LinkedIn account detection. Within our paradigm, we assess static and contextualized word embeddings, including GloVe, Flair, BERT, and RoBERTa. We show that the suggested method can distinguish between legitimate and fake profiles with an accuracy of about 95% across all word embeddings. In addition, we show that SSTE has a promising accuracy for identifying LLM-generated profiles, despite the fact that no LLM-generated profiles were employed during the training phase, and can achieve an accuracy of approximately 90% when only 20 LLM-generated profiles are added to the training set. It is a significant finding since the proliferation of several LLMs in the near future makes it extremely challenging to design a single system that can identify profiles created with various LLMs.

Phishing emails exhibit some unique psychological traits which are not present in legitimate emails. From empirical analysis and previous research, we find three psychological traits most dominant in Phishing emails - A Sense of Urgency, Inducing Fear by Threatening, and Enticement with Desire. We manually label 10% of all phishing emails in our training dataset for these three traits. We leverage that knowledge by training BERT, Sentence-BERT (SBERT), and Character-level-CNN models and capturing the nuances via the last layers that form the Phishing Psychological Trait (PPT) scores. For the phishing email detection task, we use the pretrained BERT and SBERT model, and concatenate the PPT scores to feed into a fully-connected neural network model. Our results show that the addition of PPT scores improves the model performance significantly, thus indicating the effectiveness of PPT scores in capturing the psychological nuances. Furthermore, to mitigate the effect of the imbalanced training dataset, we use the GPT-2 model to generate phishing emails (Radford et al., 2019). Our best model outperforms the current State-of-the-Art (SOTA) model's F1 score by 4.54%. Additionally, our analysis of individual PPTs suggests that Fear provides the strongest cue in detecting phishing emails.

Social media is the ultimate challenge for many natural language processing tools. The constant emergence of linguistic constructs challenge even the most sophisticated NLP tools. Predicting word embeddings for out of vocabulary words is one of those challenges. Word embedding models only include terms that occur a sufficient number of times in their training corpora. Word embedding vector models are unable to directly provide any useful information about a word not in their vocabularies. We propose a fast method for predicting vectors for out of vocabulary terms that makes use of the surrounding terms of the unknown term and the hidden context layer of the word2vec model. We propose this method as a strong baseline in the sense that 1) while it does not surpass all state-of-the-art methods, it surpasses several techniques for vector prediction on benchmark tasks, 2) even when it underperforms, the margin is very small retaining competitive performance in downstream tasks, and 3) it is inexpensive to compute, requiring no additional training stage. We also show that our technique can be incorporated into existing methods to achieve a new state-of-the-art on the word vector prediction problem.

With the increase in volume of daily online news items, it is more and more difficult for readers to identify news articles relevant to their interests. Thus, effective recommendation systems are critical for an effective user news consumption experience. Existing news recommendation methods usually rely on the news click history to model user interest. However, there are other signals about user behaviors, such as user commenting activity, which have not been used before. We propose a recommendation algorithm that predicts articles a user may be interested in, given her historical sequential commenting behavior on news articles. We show that following this sequential user behavior the news recommendation problem falls into in the class of session-based recommendation. The techniques in this class seek to model users' sequential and temporal behaviors. While we seek to follow the general directions in this space, we face unique challenges specific to news in modeling temporal dynamics, e.g., users' interests shift over time, users comment irregularly on articles, and articles are perishable items with limited lifespans. We propose a recency-regularized neural attentive framework for session-based news recommendation. The proposed method is able to capture the temporal dynamics of both users and news articles, while maintaining interpretability. We design a lag-aware attention and a recency regularization to model the time effect of news articles and comments. We conduct extensive empirical studies on 3 real-world news datasets to demonstrate the effectiveness of our method.

Imbalanced data classification has always been one of the most challenging problems in data science especially in the cybersecurity field, where we observe an out-of-balance proportion between benign and phishing examples in security datasets. Even though there are many phishing detection methods in literature, most of them neglect the imbalanced nature of phishing email datasets. In this paper, we examine the imbalanced property by varying legitimate to phishing class ratios. We generate new synthetic instances using a generative adversarial network model for long sentences (LeakGAN) to balance out the training process and ameliorate its impact on classification. These synthetic instances are labeled by positive-unlabeled learning and added to the initial imbalanced training set. The resulting dataset is given to the Bidirectional Encoder Representations from Transformers (BERT) model for sequence classification. We compare several state-of-the-art methods from the literature against our approach, which achieves a high performance throughout all the imbalanced ratios reaching an F1-score of 99.6% for the most extreme imbalanced ratio and an F1-score of 99.8% for balanced cases.