With the rapid adoption of Federated Learning (FL) as the training and tuning protocol for applications utilizing Large Language Models (LLMs), recent research highlights the need for significant modifications to FL to accommodate the large-scale of LLMs. While substantial adjustments to the protocol have been introduced as a response, comprehensive privacy analysis for the adapted FL protocol is currently lacking. To address this gap, our work delves into an extensive examination of the privacy analysis of FL when used for training LLMs, both from theoretical and practical perspectives. In particular, we design two active membership inference attacks with guaranteed theoretical success rates to assess the privacy leakages of various adapted FL configurations. Our theoretical findings are translated into practical attacks, revealing substantial privacy vulnerabilities in popular LLMs, including BERT, RoBERTa, DistilBERT, and OpenAI's GPTs, across multiple real-world language datasets. Additionally, we conduct thorough experiments to evaluate the privacy leakage of these models when data is protected by state-of-the-art differential privacy (DP) mechanisms.

Chronic diseases such as diabetes are the leading causes of morbidity and mortality worldwide. Numerous research studies have been attempted with various deep learning models in diagnosis. However, most previous studies had certain limitations, including using publicly available datasets (e.g. MIMIC), and imbalanced data. In this study, we collected five-year electronic health records (EHRs) from the Taiwan hospital database, including 1,420,596 clinical notes, 387,392 laboratory test results, and more than 1,505 laboratory test items, focusing on research pre-training large language models. We proposed a novel Large Language Multimodal Models (LLMMs) framework incorporating multimodal data from clinical notes and laboratory test results for the prediction of chronic disease risk. Our method combined a text embedding encoder and multi-head attention layer to learn laboratory test values, utilizing a deep neural network (DNN) module to merge blood features with chronic disease semantics into a latent space. In our experiments, we observe that clinicalBERT and PubMed-BERT, when combined with attention fusion, can achieve an accuracy of 73% in multiclass chronic diseases and diabetes prediction. By transforming laboratory test values into textual descriptions and employing the Flan T-5 model, we achieved a 76% Area Under the ROC Curve (AUROC), demonstrating the effectiveness of leveraging numerical text data for training and inference in language models. This approach significantly improves the accuracy of early-stage diabetes prediction.

In today's digital world, social media plays a significant role in facilitating communication and content sharing. However, the exponential rise in user-generated content has led to challenges in maintaining a respectful online environment. In some cases, users have taken advantage of anonymity in order to use harmful language, which can negatively affect the user experience and pose serious social problems. Recognizing the limitations of manual moderation, automatic detection systems have been developed to tackle this problem. Nevertheless, several obstacles persist, including the absence of a universal definition for harmful language, inadequate datasets across languages, the need for detailed annotation guideline, and most importantly, a comprehensive framework. This study aims to address these challenges by introducing, for the first time, a detailed framework adaptable to any language. This framework encompasses various aspects of harmful language detection. A key component of the framework is the development of a general and detailed annotation guideline. Additionally, the integration of sentiment analysis represents a novel approach to enhancing harmful language detection. Also, a definition of harmful language based on the review of different related concepts is presented. To demonstrate the effectiveness of the proposed framework, its implementation in a challenging low-resource language is conducted. We collected a Persian dataset and applied the annotation guideline for harmful detection and sentiment analysis. Next, we present baseline experiments utilizing machine and deep learning methods to set benchmarks. Results prove the framework's high performance, achieving an accuracy of 99.4% in offensive language detection and 66.2% in sentiment analysis.

This paper introduces SceneCraft, a Large Language Model (LLM) Agent converting text descriptions into Blender-executable Python scripts which render complex scenes with up to a hundred 3D assets. This process requires complex spatial planning and arrangement. We tackle these challenges through a combination of advanced abstraction, strategic planning, and library learning. SceneCraft first models a scene graph as a blueprint, detailing the spatial relationships among assets in the scene. SceneCraft then writes Python scripts based on this graph, translating relationships into numerical constraints for asset layout. Next, SceneCraft leverages the perceptual strengths of vision-language foundation models like GPT-V to analyze rendered images and iteratively refine the scene. On top of this process, SceneCraft features a library learning mechanism that compiles common script functions into a reusable library, facilitating continuous self-improvement without expensive LLM parameter tuning. Our evaluation demonstrates that SceneCraft surpasses existing LLM-based agents in rendering complex scenes, as shown by its adherence to constraints and favorable human assessments. We also showcase the broader application potential of SceneCraft by reconstructing detailed 3D scenes from the Sintel movie and guiding a video generative model with generated scenes as intermediary control signal.

Nowadays both commercial and open-source academic LLM have become the mainstream models of NLP. However, there is still a lack of research on LLM consistency, meaning that throughout the various stages of LLM research and deployment, its internal parameters and capabilities should remain unchanged. This issue exists in both the industrial and academic sectors. The solution to this problem is often time-consuming and labor-intensive, and there is also an additional cost of secondary deployment, resulting in economic and time losses. To fill this gap, we build an LLM consistency task dataset and design several baselines. Additionally, we choose models of diverse scales for the main experiments. Specifically, in the LightGBM experiment, we used traditional NLG metrics (i.e., ROUGE, BLEU, METEOR) as the features needed for model training. The final result exceeds the manual evaluation and GPT3.5 as well as other models in the main experiment, achieving the best performance. In the end, we use the best performing LightGBM model as the base model to build the evaluation tool, which can effectively assist in the deployment of business models. Our code and tool demo are available at https://github.com/heavenhellchen/Consistency.git

Instruction Fine-tuning~(IFT) is a critical phase in building large language models~(LLMs). Previous works mainly focus on the IFT's role in the transfer of behavioral norms and the learning of additional world knowledge. However, the understanding of the underlying mechanisms of IFT remains significantly limited. In this paper, we design a knowledge intervention framework to decouple the potential underlying factors of IFT, thereby enabling individual analysis of different factors. Surprisingly, our experiments reveal that attempting to learn additional world knowledge through IFT often struggles to yield positive impacts and can even lead to markedly negative effects. Further, we discover that maintaining internal knowledge consistency before and after IFT is a critical factor for achieving successful IFT. Our findings reveal the underlying mechanisms of IFT and provide robust support for some very recent and potential future works.

Understanding and shaping the behaviour of Large Language Models (LLMs) is increasingly important as applications become more powerful and more frequently adopted. This paper introduces a machine unlearning method specifically designed for LLMs. We introduce a selective pruning method for LLMs that removes neurons based on their relative importance on a targeted capability compared to overall network performance. This approach is a compute- and data-efficient method for identifying and removing neurons that enable specific behaviours. Our findings reveal that both feed-forward and attention neurons in LLMs are specialized; that is, for specific tasks, certain neurons are more crucial than others.

Knowledge graph question answering (KGQA) is a well-established field that seeks to provide factual answers to natural language (NL) questions by leveraging knowledge graphs (KGs). However, existing KGQA datasets suffer from two significant limitations: (1) no existing KGQA dataset requires commonsense reasoning to arrive at an answer and (2) existing KGQA datasets focus on popular entities for which large language models (LLMs) can directly answer without hallucinating and without leveraging the KG. In this work, we seek a novel KGQA dataset that supports commonsense reasoning and focuses on long-tail entities (e.g., non-mainstream and recent entities) where LLMs frequently hallucinate, and thus create the need for novel methodologies that leverage the KG for factual and attributable commonsense inference. We create a novel Commonsense Reasoning (CR) and Long-Tail (LT) KGQA dataset with two subtasks -- question answering and claim verification -- that address both limitations (1) and (2). We construct CR-LT-KGQA by building extensions to existing reasoning datasets StrategyQA and CREAK over Wikidata. While existing KGQA methods are not applicable due to their lack of commonsense inference support, baseline evaluation of LLMs on CR-LT KGQA demonstrate a high rate of hallucination. Thus, CR-LT KGQA poses significant challenges for hallucination-prone LLMs, hence paving the way for future commonsense KGQA research to provide accurate and factual answers for long-tail entities in the era of LLMs.

Wide usage of ChatGPT has highlighted the potential of reinforcement learning from human feedback. However, its training pipeline relies on manual ranking, a resource-intensive process. To reduce labor costs, we propose a self-supervised text ranking approach for applying Proximal-Policy-Optimization to fine-tune language models while eliminating the need for human annotators. Our method begins with probabilistic sampling to encourage a language model to generate diverse responses for each input. We then employ TextRank and ISODATA algorithms to rank and cluster these responses based on their semantics. Subsequently, we construct a reward model to learn the rank and optimize our generative policy. Our experimental results, conducted using two language models on three tasks, demonstrate that the models trained by our method considerably outperform baselines regarding BLEU, GLEU, and METEOR scores. Furthermore, our manual evaluation shows that our ranking results exhibit a remarkably high consistency with that of humans. This research significantly reduces training costs of proximal policy-guided models and demonstrates the potential for self-correction of language models.

Pre-trained computational language models have recently made remarkable progress in harnessing the language abilities which were considered unique to humans. Their success has raised interest in whether these models represent and process language like humans. To answer this question, this paper proposes MulCogBench, a multi-modal cognitive benchmark dataset collected from native Chinese and English participants. It encompasses a variety of cognitive data, including subjective semantic ratings, eye-tracking, functional magnetic resonance imaging (fMRI), and magnetoencephalography (MEG). To assess the relationship between language models and cognitive data, we conducted a similarity-encoding analysis which decodes cognitive data based on its pattern similarity with textual embeddings. Results show that language models share significant similarities with human cognitive data and the similarity patterns are modulated by the data modality and stimuli complexity. Specifically, context-aware models outperform context-independent models as language stimulus complexity increases. The shallow layers of context-aware models are better aligned with the high-temporal-resolution MEG signals whereas the deeper layers show more similarity with the high-spatial-resolution fMRI. These results indicate that language models have a delicate relationship with brain language representations. Moreover, the results between Chinese and English are highly consistent, suggesting the generalizability of these findings across languages.