Abstract
With the development of artificial intelligence and breakthroughs in deep learning, large-scale Foundation Models (FMs), such as GPT, CLIP, etc., have achieved remarkable results in many fields including natural language processing and computer vision. The application of FMs in autonomous driving holds considerable promise. For example, they can contribute to enhance scene understanding and reasoning. By pre-training on rich linguistic and visual data, FMs can understand and interpret various elements in a driving scene, and provide cognitive reasoning to give linguistic and action commands for driving decisions and planning. Furthermore, FMs can augment data based on its understanding of driving scenarios to provide feasible scenes of those rare occurrences in the long tail distribution that are unlikely to be encountered during routine driving and data collection. The enhancement can subsequently lead to the improvement in the accuracy and reliability of autonomous driving systems. Another testament to the potential of FMs applications lies in the development of World Models, exemplified by the DREAMER series, which showcase the ability to comprehend physical laws and dynamics. Learning from massive data under the paradigm of self-supervised learning, World Model can generate unseen yet plausible driving environment, facilitating the enhancement in the prediction of road users behavior and the off-line training of driving strategies. In this paper, we synthesize the applications and future trends of FMs in autonomous driving. By utilizing the powerful capabilities of FMs, we strive to tackle the potential issues stemming from the long-tail distribution in autonomous driving, consequently advancing overall safety in this domain.
Title:
Early Transformers: A study on Efficient Training of Transformer Models through Early-Bird Lottery Tickets
Authors: Shravan Cheekati
Subjects: Subjects:
Computation and Language (cs.CL); Machine Learning (cs.LG)
Abstract
The training of Transformer models has revolutionized natural language processing and computer vision, but it remains a resource-intensive and time-consuming process. This paper investigates the applicability of the early-bird ticket hypothesis to optimize the training efficiency of Transformer models. We propose a methodology that combines iterative pruning, masked distance calculation, and selective retraining to identify early-bird tickets in various Transformer architectures, including ViT, Swin-T, GPT-2, and RoBERTa. Our experimental results demonstrate that early-bird tickets can be consistently found within the first few epochs of training or fine-tuning, enabling significant resource optimization without compromising performance. The pruned models obtained from early-bird tickets achieve comparable or even superior accuracy to their unpruned counterparts while substantially reducing memory usage. Furthermore, our comparative analysis highlights the generalizability of the early-bird ticket phenomenon across different Transformer models and tasks. This research contributes to the development of efficient training strategies for Transformer models, making them more accessible and resource-friendly. By leveraging early-bird tickets, practitioners can accelerate the progress of natural language processing and computer vision applications while reducing the computational burden associated with training Transformer models.
Title:
PropertyGPT: LLM-driven Formal Verification of Smart Contracts through Retrieval-Augmented Property Generation
Authors: Ye Liu, Yue Xue, Daoyuan Wu, Yuqiang Sun, Yi Li, Miaolei Shi, Yang Liu
Abstract
With recent advances in large language models (LLMs), this paper explores the potential of leveraging state-of-the-art LLMs, such as GPT-4, to transfer existing human-written properties (e.g., those from Certora auditing reports) and automatically generate customized properties for unknown code. To this end, we embed existing properties into a vector database and retrieve a reference property for LLM-based in-context learning to generate a new prop- erty for a given code. While this basic process is relatively straight- forward, ensuring that the generated properties are (i) compilable, (ii) appropriate, and (iii) runtime-verifiable presents challenges. To address (i), we use the compilation and static analysis feedback as an external oracle to guide LLMs in iteratively revising the generated properties. For (ii), we consider multiple dimensions of similarity to rank the properties and employ a weighted algorithm to identify the top-K properties as the final result. For (iii), we design a dedicated prover to formally verify the correctness of the generated prop- erties. We have implemented these strategies into a novel system called PropertyGPT, with 623 human-written properties collected from 23 Certora projects. Our experiments show that PropertyGPT can generate comprehensive and high-quality properties, achieving an 80% recall compared to the ground truth. It successfully detected 26 CVEs/attack incidents out of 37 tested and also uncovered 12 zero-day vulnerabilities, resulting in $8,256 bug bounty rewards.
Title:
Confidential and Protected Disease Classifier using Fully Homomorphic Encryption
Abstract
With the rapid surge in the prevalence of Large Language Models (LLMs), individuals are increasingly turning to conversational AI for initial insights across various domains, including health-related inquiries such as disease diagnosis. Many users seek potential causes on platforms like ChatGPT or Bard before consulting a medical professional for their ailment. These platforms offer valuable benefits by streamlining the diagnosis process, alleviating the significant workload of healthcare practitioners, and saving users both time and money by avoiding unnecessary doctor visits. However, Despite the convenience of such platforms, sharing personal medical data online poses risks, including the presence of malicious platforms or potential eavesdropping by attackers. To address privacy concerns, we propose a novel framework combining FHE and Deep Learning for a secure and private diagnosis system. Operating on a question-and-answer-based model akin to an interaction with a medical practitioner, this end-to-end secure system employs Fully Homomorphic Encryption (FHE) to handle encrypted input data. Given FHE's computational constraints, we adapt deep neural networks and activation functions to the encryted domain. Further, we also propose a faster algorithm to compute summation of ciphertext elements. Through rigorous experiments, we demonstrate the efficacy of our approach. The proposed framework achieves strict security and privacy with minimal loss in performance.
Abstract
Despite the longstanding adage "an image is worth a thousand words," creating accurate and hyper-detailed image descriptions for training Vision-Language models remains challenging. Current datasets typically have web-scraped descriptions that are short, low-granularity, and often contain details unrelated to the visual content. As a result, models trained on such data generate descriptions replete with missing information, visual inconsistencies, and hallucinations. To address these issues, we introduce ImageInWords (IIW), a carefully designed human-in-the-loop annotation framework for curating hyper-detailed image descriptions and a new dataset resulting from this process. We validate the framework through evaluations focused on the quality of the dataset and its utility for fine-tuning with considerations for readability, comprehensiveness, specificity, hallucinations, and human-likeness. Our dataset significantly improves across these dimensions compared to recently released datasets (+66%) and GPT-4V outputs (+48%). Furthermore, models fine-tuned with IIW data excel by +31% against prior work along the same human evaluation dimensions. Given our fine-tuned models, we also evaluate text-to-image generation and vision-language reasoning. Our model's descriptions can generate images closest to the original, as judged by both automated and human metrics. We also find our model produces more compositionally rich descriptions, outperforming the best baseline by up to 6% on ARO, SVO-Probes, and Winoground datasets.
Title:
NegativePrompt: Leveraging Psychology for Large Language Models Enhancement via Negative Emotional Stimuli
Abstract
Large Language Models (LLMs) have become integral to a wide spectrum of applications, ranging from traditional computing tasks to advanced artificial intelligence (AI) applications. This widespread adoption has spurred extensive research into LLMs across various disciplines, including the social sciences. Notably, studies have revealed that LLMs possess emotional intelligence, which can be further developed through positive emotional stimuli. This discovery raises an intriguing question: can negative emotions similarly influence LLMs, potentially enhancing their performance? In response to this question, we introduce NegativePrompt, a novel approach underpinned by psychological principles, involving ten specifically designed negative emotional stimuli. We embark on rigorous experimental evaluations of five LLMs including Flan-T5-Large, Vicuna, Llama 2, ChatGPT, and GPT-4, across a set of 45 tasks. The results are revealing: NegativePrompt markedly enhances the performance of LLMs, evidenced by relative improvements of 12.89% in Instruction Induction tasks and 46.25% in BIG-Bench tasks. Moreover, we conduct attention visualization experiments to decipher the underlying mechanisms of NegativePrompt's influence. Our research contributes significantly to the understanding of LLMs and emotion interaction, demonstrating the practical efficacy of NegativePrompt as an emotion-driven method and offering novel insights for the enhancement of LLMs in real-world applications. The code is available at this https URL.
Title:
Overconfidence is Key: Verbalized Uncertainty Evaluation in Large Language and Vision-Language Models
Authors: Tobias Groot, Matias Valdenegro-Toro
Subjects: Subjects:
Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Machine Learning (cs.LG)
Abstract
Language and Vision-Language Models (LLMs/VLMs) have revolutionized the field of AI by their ability to generate human-like text and understand images, but ensuring their reliability is crucial. This paper aims to evaluate the ability of LLMs (GPT4, GPT-3.5, LLaMA2, and PaLM 2) and VLMs (GPT4V and Gemini Pro Vision) to estimate their verbalized uncertainty via prompting. We propose the new Japanese Uncertain Scenes (JUS) dataset, aimed at testing VLM capabilities via difficult queries and object counting, and the Net Calibration Error (NCE) to measure direction of miscalibration. Results show that both LLMs and VLMs have a high calibration error and are overconfident most of the time, indicating a poor capability for uncertainty estimation. Additionally we develop prompts for regression tasks, and we show that VLMs have poor calibration when producing mean/standard deviation and 95% confidence intervals.
Title:
Unraveling the Dominance of Large Language Models Over Transformer Models for Bangla Natural Language Inference: A Comprehensive Study
Authors: Fatema Tuj Johora Faria, Mukaffi Bin Moin, Asif Iftekher Fahim, Pronay Debnath, Faisal Muhammad Shah
Subjects: Subjects:
Computation and Language (cs.CL)
Abstract
Natural Language Inference (NLI) is a cornerstone of Natural Language Processing (NLP), providing insights into the entailment relationships between text pairings. It is a critical component of Natural Language Understanding (NLU), demonstrating the ability to extract information from spoken or written interactions. NLI is mainly concerned with determining the entailment relationship between two statements, known as the premise and hypothesis. When the premise logically implies the hypothesis, the pair is labeled entailment''. If the hypothesis contradicts the premise, the pair receives thecontradiction'' label. When there is insufficient evidence to establish a connection, the pair is described as ``neutral''. Despite the success of Large Language Models (LLMs) in various tasks, their effectiveness in NLI remains constrained by issues like low-resource domain accuracy, model overconfidence, and difficulty in capturing human judgment disagreements. This study addresses the underexplored area of evaluating LLMs in low-resourced languages such as Bengali. Through a comprehensive evaluation, we assess the performance of prominent LLMs and state-of-the-art (SOTA) models in Bengali NLP tasks, focusing on natural language inference. Utilizing the XNLI dataset, we conduct zero-shot and few-shot evaluations, comparing LLMs like GPT-3.5 Turbo and Gemini 1.5 Pro with models such as BanglaBERT, Bangla BERT Base, DistilBERT, mBERT, and sahajBERT. Our findings reveal that while LLMs can achieve comparable or superior performance to fine-tuned SOTA models in few-shot scenarios, further research is necessary to enhance our understanding of LLMs in languages with modest resources like Bengali. This study underscores the importance of continued efforts in exploring LLM capabilities across diverse linguistic contexts.
Title:
Can Large Language Models Make the Grade? An Empirical Study Evaluating LLMs Ability to Mark Short Answer Questions in K-12 Education
Authors: Owen Henkel, Adam Boxer, Libby Hills, Bill Roberts
Subjects: Subjects:
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Abstract
This paper presents reports on a series of experiments with a novel dataset evaluating how well Large Language Models (LLMs) can mark (i.e. grade) open text responses to short answer questions, Specifically, we explore how well different combinations of GPT version and prompt engineering strategies performed at marking real student answers to short answer across different domain areas (Science and History) and grade-levels (spanning ages 5-16) using a new, never-used-before dataset from Carousel, a quizzing platform. We found that GPT-4, with basic few-shot prompting performed well (Kappa, 0.70) and, importantly, very close to human-level performance (0.75). This research builds on prior findings that GPT-4 could reliably score short answer reading comprehension questions at a performance-level very close to that of expert human raters. The proximity to human-level performance, across a variety of subjects and grade levels suggests that LLMs could be a valuable tool for supporting low-stakes formative assessment tasks in K-12 education and has important implications for real-world education delivery.
Title:
Traffic Performance GPT (TP-GPT): Real-Time Data Informed Intelligent ChatBot for Transportation Surveillance and Management
Authors: Bingzhang Wang, Zhiyu (Joey)Cai, Muhammad Monjurul Karim, Chenxi Liu, Yinhai Wang
Abstract
The digitization of traffic sensing infrastructure has significantly accumulated an extensive traffic data warehouse, which presents unprecedented challenges for transportation analytics. The complexities associated with querying large-scale multi-table databases require specialized programming expertise and labor-intensive development. Additionally, traditional analysis methods have focused mainly on numerical data, often neglecting the semantic aspects that could enhance interpretability and understanding. Furthermore, real-time traffic data access is typically limited due to privacy concerns. To bridge this gap, the integration of Large Language Models (LLMs) into the domain of traffic management presents a transformative approach to addressing the complexities and challenges inherent in modern transportation systems. This paper proposes an intelligent online chatbot, TP-GPT, for efficient customized transportation surveillance and management empowered by a large real-time traffic database. The innovative framework leverages contextual and generative intelligence of language models to generate accurate SQL queries and natural language interpretations by employing transportation-specialized prompts, Chain-of-Thought prompting, few-shot learning, multi-agent collaboration strategy, and chat memory. Experimental study demonstrates that our approach outperforms state-of-the-art baselines such as GPT-4 and PaLM 2 on a challenging traffic-analysis benchmark TransQuery. TP-GPT would aid researchers and practitioners in real-time transportation surveillance and management in a privacy-preserving, equitable, and customizable manner.
Title:
SketchGPT: Autoregressive Modeling for Sketch Generation and Recognition
Abstract
We present SketchGPT, a flexible framework that employs a sequence-to-sequence autoregressive model for sketch generation, and completion, and an interpretation case study for sketch recognition. By mapping complex sketches into simplified sequences of abstract primitives, our approach significantly streamlines the input for autoregressive modeling. SketchGPT leverages the next token prediction objective strategy to understand sketch patterns, facilitating the creation and completion of drawings and also categorizing them accurately. This proposed sketch representation strategy aids in overcoming existing challenges of autoregressive modeling for continuous stroke data, enabling smoother model training and competitive performance. Our findings exhibit SketchGPT's capability to generate a diverse variety of drawings by adding both qualitative and quantitative comparisons with existing state-of-the-art, along with a comprehensive human evaluation study. The code and pretrained models will be released on our official GitHub.
Title:
Exploring the Potential of the Large Language Models (LLMs) in Identifying Misleading News Headlines
Authors: Md Main Uddin Rony, Md Mahfuzul Haque, Mohammad Ali, Ahmed Shatil Alam, Naeemul Hassan
Subjects: Subjects:
Computation and Language (cs.CL); Computers and Society (cs.CY); Machine Learning (cs.LG)
Abstract
In the digital age, the prevalence of misleading news headlines poses a significant challenge to information integrity, necessitating robust detection mechanisms. This study explores the efficacy of Large Language Models (LLMs) in identifying misleading versus non-misleading news headlines. Utilizing a dataset of 60 articles, sourced from both reputable and questionable outlets across health, science & tech, and business domains, we employ three LLMs- ChatGPT-3.5, ChatGPT-4, and Gemini-for classification. Our analysis reveals significant variance in model performance, with ChatGPT-4 demonstrating superior accuracy, especially in cases with unanimous annotator agreement on misleading headlines. The study emphasizes the importance of human-centered evaluation in developing LLMs that can navigate the complexities of misinformation detection, aligning technical proficiency with nuanced human judgment. Our findings contribute to the discourse on AI ethics, emphasizing the need for models that are not only technically advanced but also ethically aligned and sensitive to the subtleties of human interpretation.
Title:
Anchored Answers: Unravelling Positional Bias in GPT-2's Multiple-Choice Questions
Authors: Ruizhe Li, Yanjun Gao
Subjects: Subjects:
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Abstract
Large Language Models (LLMs), such as the GPT-4 and LLaMA families, have demonstrated considerable success across diverse tasks, including multiple-choice questions (MCQs). However, these models exhibit a positional bias, particularly an even worse anchored bias in the GPT-2 family, where they consistently favour the first choice 'A' in MCQs during inference. This anchored bias challenges the integrity of GPT-2's decision-making process, as it skews performance based on the position rather than the content of the choices in MCQs. In this study, we utilise the mechanistic interpretability approach to identify the internal modules within GPT-2 models responsible for this bias. We focus on the Multi-Layer Perceptron (MLP) layers and attention heads, using the "logit lens" method to trace and modify the specific value vectors that contribute to the bias. By updating these vectors within MLP and recalibrating attention patterns to neutralise the preference for the first choice 'A', we effectively mitigate the anchored bias. Our interventions not only correct the bias but also improve the overall MCQ prediction accuracy for the GPT-2 family across various datasets. This work represents the first comprehensive mechanistic analysis of anchored bias in MCQs within the GPT-2 models, introducing targeted, minimal-intervention strategies that significantly enhance GPT2 model robustness and accuracy in MCQs. Our code is available at this https URL.
Title:
The high dimensional psychological profile and cultural bias of ChatGPT
Authors: Hang Yuan (1), Zhongyue Che (1), Shao Li (1), Yue Zhang, Xiaomeng Hu (2), Siyang Luo (1) ((1) Sun Yat-Sen University, (2) Renmin University of China)
Subjects: Subjects:
Computation and Language (cs.CL)
Abstract
Given the rapid advancement of large-scale language models, artificial intelligence (AI) models, like ChatGPT, are playing an increasingly prominent role in human society. However, to ensure that artificial intelligence models benefit human society, we must first fully understand the similarities and differences between the human-like characteristics exhibited by artificial intelligence models and real humans, as well as the cultural stereotypes and biases that artificial intelligence models may exhibit in the process of interacting with humans. This study first measured ChatGPT in 84 dimensions of psychological characteristics, revealing differences between ChatGPT and human norms in most dimensions as well as in high-dimensional psychological representations. Additionally, through the measurement of ChatGPT in 13 dimensions of cultural values, it was revealed that ChatGPT's cultural value patterns are dissimilar to those of various countries/regions worldwide. Finally, an analysis of ChatGPT's performance in eight decision-making tasks involving interactions with humans from different countries/regions revealed that ChatGPT exhibits clear cultural stereotypes in most decision-making tasks and shows significant cultural bias in third-party punishment and ultimatum games. The findings indicate that, compared to humans, ChatGPT exhibits a distinct psychological profile and cultural value orientation, and it also shows cultural biases and stereotypes in interpersonal decision-making. Future research endeavors should emphasize enhanced technical oversight and augmented transparency in the database and algorithmic training procedures to foster more efficient cross-cultural communication and mitigate social disparities.
Title:
UnsafeBench: Benchmarking Image Safety Classifiers on Real-World and AI-Generated Images
Authors: Yiting Qu, Xinyue Shen, Yixin Wu, Michael Backes, Savvas Zannettou, Yang Zhang
Subjects: Subjects:
Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV); Social and Information Networks (cs.SI)
Abstract
Image safety classifiers play an important role in identifying and mitigating the spread of unsafe images online (e.g., images including violence, hateful rhetoric, etc.). At the same time, with the advent of text-to-image models and increasing concerns about the safety of AI models, developers are increasingly relying on image safety classifiers to safeguard their models. Yet, the performance of current image safety classifiers remains unknown for real-world and AI-generated images. To bridge this research gap, in this work, we propose UnsafeBench, a benchmarking framework that evaluates the effectiveness and robustness of image safety classifiers. First, we curate a large dataset of 10K real-world and AI-generated images that are annotated as safe or unsafe based on a set of 11 unsafe categories of images (sexual, violent, hateful, etc.). Then, we evaluate the effectiveness and robustness of five popular image safety classifiers, as well as three classifiers that are powered by general-purpose visual language models. Our assessment indicates that existing image safety classifiers are not comprehensive and effective enough in mitigating the multifaceted problem of unsafe images. Also, we find that classifiers trained only on real-world images tend to have degraded performance when applied to AI-generated images. Motivated by these findings, we design and implement a comprehensive image moderation tool called PerspectiveVision, which effectively identifies 11 categories of real-world and AI-generated unsafe images. The best PerspectiveVision model achieves an overall F1-Score of 0.810 on six evaluation datasets, which is comparable with closed-source and expensive state-of-the-art models like GPT-4V. UnsafeBench and PerspectiveVision can aid the research community in better understanding the landscape of image safety classification in the era of generative AI.
Title:
MAmmoTH2: Scaling Instructions from the Web
Authors: Xiang Yue, Tuney Zheng, Ge Zhang, Wenhu Chen
Subjects: Subjects:
Computation and Language (cs.CL)
Abstract
Instruction tuning improves the reasoning abilities of large language models (LLMs), with data quality and scalability being the crucial factors. Most instruction tuning data come from human crowd-sourcing or GPT-4 distillation. We propose a paradigm to efficiently harvest 10 million naturally existing instruction data from the pre-training web corpus to enhance LLM reasoning. Our approach involves (1) recalling relevant documents, (2) extracting instruction-response pairs, and (3) refining the extracted pairs using open-source LLMs. Fine-tuning base LLMs on this dataset, we build MAmmoTH2 models, which significantly boost performance on reasoning benchmarks. Notably, MAmmoTH2-7B's (Mistral) performance increases from 11% to 34% on MATH and from 36% to 67% on GSM8K without training on any in-domain data. Further training MAmmoTH2 on public instruction tuning datasets yields MAmmoTH2-Plus, achieving state-of-the-art performance on several reasoning and chatbot benchmarks. Our work demonstrates how to harvest large-scale, high-quality instruction data without costly human annotation or GPT-4 distillation, providing a new paradigm for building better instruction tuning data.
Title:
Enabling High-Sparsity Foundational Llama Models with Efficient Pretraining and Deployment
Authors: Abhinav Agarwalla, Abhay Gupta, Alexandre Marques, Shubhra Pandit, Michael Goin, Eldar Kurtic, Kevin Leong, Tuan Nguyen, Mahmoud Salem, Dan Alistarh, Sean Lie, Mark Kurtz
Subjects: Subjects:
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Abstract
Large language models (LLMs) have revolutionized Natural Language Processing (NLP), but their size creates computational bottlenecks. We introduce a novel approach to create accurate, sparse foundational versions of performant LLMs that achieve full accuracy recovery for fine-tuning tasks at up to 70% sparsity. We achieve this for the LLaMA-2 7B model by combining the SparseGPT one-shot pruning method and sparse pretraining of those models on a subset of the SlimPajama dataset mixed with a Python subset of The Stack dataset. We exhibit training acceleration due to sparsity on Cerebras CS-3 chips that closely matches theoretical scaling. In addition, we establish inference acceleration of up to 3x on CPUs by utilizing Neural Magic's DeepSparse engine and 1.7x on GPUs through Neural Magic's nm-vllm engine. The above gains are realized via sparsity alone, thus enabling further gains through additional use of quantization. Specifically, we show a total speedup on CPUs for sparse-quantized LLaMA models of up to 8.6x. We demonstrate these results across diverse, challenging tasks, including chat, instruction following, code generation, arithmetic reasoning, and summarization to prove their generality. This work paves the way for rapidly creating smaller and faster LLMs without sacrificing accuracy.
Title:
GREEN: Generative Radiology Report Evaluation and Error Notation
Authors: Sophie Ostmeier, Justin Xu, Zhihong Chen, Maya Varma, Louis Blankemeier, Christian Bluethgen, Arne Edward Michalson, Michael Moseley, Curtis Langlotz, Akshay S Chaudhari, Jean-Benoit Delbrouck
Subjects: Subjects:
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Abstract
Evaluating radiology reports is a challenging problem as factual correctness is extremely important due to the need for accurate medical communication about medical images. Existing automatic evaluation metrics either suffer from failing to consider factual correctness (e.g., BLEU and ROUGE) or are limited in their interpretability (e.g., F1CheXpert and F1RadGraph). In this paper, we introduce GREEN (Generative Radiology Report Evaluation and Error Notation), a radiology report generation metric that leverages the natural language understanding of language models to identify and explain clinically significant errors in candidate reports, both quantitatively and qualitatively. Compared to current metrics, GREEN offers: 1) a score aligned with expert preferences, 2) human interpretable explanations of clinically significant errors, enabling feedback loops with end-users, and 3) a lightweight open-source method that reaches the performance of commercial counterparts. We validate our GREEN metric by comparing it to GPT-4, as well as to error counts of 6 experts and preferences of 2 experts. Our method demonstrates not only higher correlation with expert error counts, but simultaneously higher alignment with expert preferences when compared to previous approaches."
Title:
Can LLMs Deeply Detect Complex Malicious Queries? A Framework for Jailbreaking via Obfuscating Intent
Authors: Shang Shang, Xinqiang Zhao, Zhongjiang Yao, Yepeng Yao, Liya Su, Zijing Fan, Xiaodan Zhang, Zhengwei Jiang
Subjects: Subjects:
Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
Abstract
To demonstrate and address the underlying maliciousness, we propose a theoretical hypothesis and analytical approach, and introduce a new black-box jailbreak attack methodology named IntentObfuscator, exploiting this identified flaw by obfuscating the true intentions behind user prompts.This approach compels LLMs to inadvertently generate restricted content, bypassing their built-in content security measures. We detail two implementations under this framework: "Obscure Intention" and "Create Ambiguity", which manipulate query complexity and ambiguity to evade malicious intent detection effectively. We empirically validate the effectiveness of the IntentObfuscator method across several models, including ChatGPT-3.5, ChatGPT-4, Qwen and Baichuan, achieving an average jailbreak success rate of 69.21\%. Notably, our tests on ChatGPT-3.5, which claims 100 million weekly active users, achieved a remarkable success rate of 83.65\%. We also extend our validation to diverse types of sensitive content like graphic violence, racism, sexism, political sensitivity, cybersecurity threats, and criminal skills, further proving the substantial impact of our findings on enhancing 'Red Team' strategies against LLM content security frameworks.
Title:
Towards A Human-in-the-Loop LLM Approach to Collaborative Discourse Analysis
Abstract
LLMs have demonstrated proficiency in contextualizing their outputs using human input, often matching or beating human-level performance on a variety of tasks. However, LLMs have not yet been used to characterize synergistic learning in students' collaborative discourse. In this exploratory work, we take a first step towards adopting a human-in-the-loop prompt engineering approach with GPT-4-Turbo to summarize and categorize students' synergistic learning during collaborative discourse. Our preliminary findings suggest GPT-4-Turbo may be able to characterize students' synergistic learning in a manner comparable to humans and that our approach warrants further investigation.
Title:
Large Language Models Reveal Information Operation Goals, Tactics, and Narrative Frames
Authors: Keith Burghardt, Kai Chen, Kristina Lerman
Subjects: Subjects:
Computation and Language (cs.CL); Machine Learning (cs.LG)
Abstract
Adversarial information operations can destabilize societies by undermining fair elections, manipulating public opinions on policies, and promoting scams. Despite their widespread occurrence and potential impacts, our understanding of influence campaigns is limited by manual analysis of messages and subjective interpretation of their observable behavior. In this paper, we explore whether these limitations can be mitigated with large language models (LLMs), using GPT-3.5 as a case-study for coordinated campaign annotation. We first use GPT-3.5 to scrutinize 126 identified information operations spanning over a decade. We utilize a number of metrics to quantify the close (if imperfect) agreement between LLM and ground truth descriptions. We next extract coordinated campaigns from two large multilingual datasets from X (formerly Twitter) that respectively discuss the 2022 French election and 2023 Balikaran Philippine-U.S. military exercise in 2023. For each coordinated campaign, we use GPT-3.5 to analyze posts related to a specific concern and extract goals, tactics, and narrative frames, both before and after critical events (such as the date of an election). While the GPT-3.5 sometimes disagrees with subjective interpretation, its ability to summarize and interpret demonstrates LLMs' potential to extract higher-order indicators from text to provide a more complete picture of the information campaigns compared to previous methods.
Keyword: gpt
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
entailment''. If the hypothesis contradicts the premise, the pair receives thecontradiction'' label. When there is insufficient evidence to establish a connection, the pair is described as ``neutral''. Despite the success of Large Language Models (LLMs) in various tasks, their effectiveness in NLI remains constrained by issues like low-resource domain accuracy, model overconfidence, and difficulty in capturing human judgment disagreements. This study addresses the underexplored area of evaluating LLMs in low-resourced languages such as Bengali. Through a comprehensive evaluation, we assess the performance of prominent LLMs and state-of-the-art (SOTA) models in Bengali NLP tasks, focusing on natural language inference. Utilizing the XNLI dataset, we conduct zero-shot and few-shot evaluations, comparing LLMs like GPT-3.5 Turbo and Gemini 1.5 Pro with models such as BanglaBERT, Bangla BERT Base, DistilBERT, mBERT, and sahajBERT. Our findings reveal that while LLMs can achieve comparable or superior performance to fine-tuned SOTA models in few-shot scenarios, further research is necessary to enhance our understanding of LLMs in languages with modest resources like Bengali. This study underscores the importance of continued efforts in exploring LLM capabilities across diverse linguistic contexts.Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title:
Title: