Our AU-ROC scores show that GPT-4o, OpenAI's more recent and capable model, demonstrates strong recognition of paywalled O'Reilly book content (AUROC = 82%), compared to OpenAI's earlier model GPT-3.5 Turbo. In contrast, GPT-3.5 Turbo shows greater relative recognition of publicly accessible O'Reilly book samples. GPT-4o Mini, as a much smaller model, shows no knowledge of public or non-public O'Reilly Media content when tested (AUROC 50%). Testing multiple models, with the same cutoff date, helps us account for potential language shifts over time that might bias our findings. These results highlight the urgent need for increased corporate transparency regarding pre-training data sources as a means to develop formal licensing frameworks for AI content training.

Prompt engineering enables Large Language Models (LLMs) to perform a variety of tasks. However, lengthy prompts significantly increase computational complexity and economic costs. To address this issue, we study six prompt compression methods for LLMs, aiming to reduce prompt length while maintaining LLM response quality. In this paper, we present a comprehensive analysis covering aspects such as generation performance, model hallucinations, efficacy in multimodal tasks, word omission analysis, and more. We evaluate these methods across 13 datasets, including news, scientific articles, commonsense QA, math QA, long-context QA, and VQA datasets. Our experiments reveal that prompt compression has a greater impact on LLM performance in long contexts compared to short ones. In the Longbench evaluation, moderate compression even enhances LLM performance. Our code and data is available at https://github.com/3DAgentWorld/Toolkit-for-Prompt-Compression.

These authors contributed equally to this work. Abstract To enable precise and fully automated cell type annotation with large language models (LLMs), we developed a graph-structured feature-marker database to retrieve entities linked to differential genes for cell reconstruction. We further designed a multi-task workflow to optimize the annotation process. Compared to general-purpose LLMs, our method improves human evaluation scores by up to 0.21 and semantic similarity by 6.1% across 11 tissue types, while more closely aligning with the cognitive logic of manual annotation. Keywords: Cell type annotation, Graph RAG, Large language models, Graph data curation, Multi-task workflow, scRNA-seq In single-cell RNA sequencing analysis, achieving precise cell type annotation through manual labeling typically requires two key steps: annotators retrieve relevant marker genes and integrate this information with their domain expertise to make informed decisions.

Background Practical applications such as social media monitoring and customer-feedback analysis require accurate emotion detection for Japanese text, yet resource scarcity and class imbalance hinder model performance. Objective This study aims to build a high-accuracy model for predicting the presence or absence of eight Plutchik emotions in Japanese sentences. Methods Using the WRIME corpus, we transform reader-averaged intensity scores into binary labels and fine-tune four pre-trained language models (BERT, RoBERTa, DeBERTa-v3-base, DeBERTa-v3-large). For context, we also assess two large language models (TinySwallow-1.5B-Instruct and ChatGPT-4o). Accuracy and F1-score serve as evaluation metrics. Results DeBERTa-v3-large attains the best mean accuracy (0.860) and F1-score (0.662), outperforming all other models. It maintains robust F1 across both high-frequency emotions (e.g., Joy, Anticipation) and low-frequency emotions (e.g., Anger, Trust). The LLMs lag, with ChatGPT-4o and TinySwallow-1.5B-Instruct scoring 0.527 and 0.292 in mean F1, respectively. Conclusion The fine-tuned DeBERTa-v3-large model currently offers the most reliable solution for binary emotion classification in Japanese. We release this model as a pip-installable package (pip install deberta-emotion-predictor). Future work should augment data for rare emotions, reduce model size, and explore prompt engineering to improve LLM performance. This manuscript is under review for possible publication in New Generation Computing.

Qualitative research often involves labor-intensive processes that are difficult to scale while preserving analytical depth. This paper introduces The AI Co-Ethnographer (AICoE), a novel end-to-end pipeline developed for qualitative research and designed to move beyond the limitations of simply automating code assignments, offering a more integrated approach. AICoE organizes the entire process, encompassing open coding, code consolidation, code application, and even pattern discovery, leading to a comprehensive analysis of qualitative data.

Jailbreaking in Large Language Models (LLMs) threatens their safe use in sensitive domains like education by allowing users to bypass ethical safeguards. This study focuses on detecting jailbreaks in 2-Sigma, a clinical education platform that simulates patient interactions using LLMs. We annotated over 2,300 prompts across 158 conversations using four linguistic variables shown to correlate strongly with jailbreak behavior. The extracted features were used to train several predictive models, including Decision Trees, Fuzzy Logic-based classifiers, Boosting methods, and Logistic Regression. Results show that feature-based predictive models consistently outperformed Prompt Engineering, with the Fuzzy Decision Tree achieving the best overall performance. Our findings demonstrate that linguistic-feature-based models are effective and explainable alternatives for jailbreak detection. We suggest future work explore hybrid frameworks that integrate prompt-based flexibility with rule-based robustness for real-time, spectrum-based jailbreak monitoring in educational LLMs.

Pre-trained language models (PLMs) demonstrate remarkable intelligence but struggle with emerging tasks unseen during training in real-world applications. Training separate models for each new task is usually impractical. Multi-task learning (MTL) addresses this challenge by transferring shared knowledge from source tasks to target tasks. As an dominant parameter-efficient fine-tuning method, prompt tuning (PT) enhances MTL by introducing an adaptable vector that captures task-specific knowledge, which acts as a prefix to the original prompt that preserves shared knowledge, while keeping PLM parameters frozen. However, PT struggles to effectively capture the heterogeneity of task-specific knowledge due to its limited representational capacity. To address this challenge, we propose Task-Adaptive Low-Rank Representation (TA-LoRA), an MTL method built on PT, employing the low-rank representation to model task heterogeneity and a fast-slow weights mechanism where the slow weight encodes shared knowledge, while the fast weight captures task-specific nuances, avoiding the mixing of shared and task-specific knowledge, caused by training low-rank representations from scratch. Moreover, a zero-initialized attention mechanism is introduced to minimize the disruption of immature low-rank components on original prompts during warm-up epochs. Experiments on 16 tasks demonstrate that TA-LoRA achieves state-of-the-art performance in full-data and few-shot settings while maintaining superior parameter efficiency.

In this paper we provide evidence that a virtual model of U.S. congresspersons based on a collection of language models satisfies the definition of a digital twin. In particular, we introduce and provide high-level descriptions of a daily-updated dataset that contains every Tweet from every U.S. congressperson during their respective terms. We demonstrate that a modern language model equipped with congressperson-specific subsets of this data are capable of producing Tweets that are largely indistinguishable from actual Tweets posted by their physical counterparts. We illustrate how generated Tweets can be used to predict roll-call vote behaviors and to quantify the likelihood of congresspersons crossing party lines, thereby assisting stakeholders in allocating resources and potentially impacting real-world legislative dynamics. We conclude with a discussion of the limitations and important extensions of our analysis.

We present LangVAE, a novel framework for modular construction of variational autoencoders (VAEs) on top of pre-trained large language models (LLMs). Such language model VAEs can encode the knowledge of their pre-trained components into more compact and semantically disentangled representations. The representations obtained in this way can be analysed with the LangVAE companion framework: LangSpace, which implements a collection of probing methods, such as vector traversal and interpolation, disentanglement measures, and cluster visualisations. LangVAE and LangSpace offer a flexible, efficient and scalable way of building and analysing textual representations, with simple integration for models available on the HuggingFace Hub. Additionally, we conducted a set of experiments with different encoder and decoder combinations, as well as annotated inputs, revealing a wide range of interactions across architectural families and sizes w.r.t. generalisation and disentanglement. Our findings demonstrate a promising framework for systematising the experimentation and understanding of textual representations.

Fine-tuning large language models (LLMs) has become an essential strategy for adapting them to specialized tasks; however, this process introduces significant privacy challenges, as sensitive training data may be inadvertently memorized and exposed. Although differential privacy (DP) offers strong theoretical guarantees against such leakage, its empirical privacy effectiveness on LLMs remains unclear, especially under different fine-tuning methods. In this paper, we systematically investigate the impact of DP across fine-tuning methods and privacy budgets, using both data extraction and membership inference attacks to assess empirical privacy risks. Our main findings are as follows: (1) Differential privacy reduces model utility, but its impact varies significantly across different fine-tuning methods. (2) Without DP, the privacy risks of models fine-tuned with different approaches differ considerably. (3) When DP is applied, even a relatively high privacy budget can substantially lower privacy risk. (4) The privacy-utility trade-off under DP training differs greatly among fine-tuning methods, with some methods being unsuitable for DP due to severe utility degradation. Our results provide practical guidance for privacy-conscious deployment of LLMs and pave the way for future research on optimizing the privacy-utility trade-off in fine-tuning methodologies.