Abstract

The White House Executive Order on Artificial Intelligence highlights the risks of large language models (LLMs) empowering malicious actors in developing biological, cyber, and chemical weapons. To measure these risks of malicious use, government institutions and major AI labs are developing evaluations for hazardous capabilities in LLMs. However, current evaluations are private, preventing further research into mitigating risk. Furthermore, they focus on only a few, highly specific pathways for malicious use. To fill these gaps, we publicly release the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of 3,668 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. WMDP was developed by a consortium of academics and technical consultants, and was stringently filtered to eliminate sensitive information prior to public release. WMDP serves two roles: first, as an evaluation for hazardous knowledge in LLMs, and second, as a benchmark for unlearning methods to remove such hazardous knowledge. To guide progress on unlearning, we develop RMU, a state-of-the-art unlearning method based on controlling model representations. RMU reduces model performance on WMDP while maintaining general capabilities in areas such as biology and computer science, suggesting that unlearning may be a concrete path towards reducing malicious use from LLMs. We release our benchmark and code publicly at https://wmdp.ai

Abstract

Fake news being spread on social media platforms has a disruptive and damaging impact on our lives. Multimedia content improves the visibility of posts more than text data but is also being used for creating fake news. Previous multimodal works have tried to address the problem of modeling heterogeneous modalities in identifying fake news. However, these works have the following limitations: (1) inefficient encoding of inter-modal relations by utilizing a simple concatenation operator on the modalities at a later stage in a model, which might result in information loss; (2) training very deep neural networks with a disproportionate number of parameters on small multimodal datasets result in higher chances of overfitting. To address these limitations, we propose GAME-ON, a Graph Neural Network based end-to-end trainable framework that allows granular interactions within and across different modalities to learn more robust data representations for multimodal fake news detection. We use two publicly available fake news datasets, Twitter and Weibo, for evaluations. GAME-ON outperforms on Twitter by an average of 11% and achieves state-of-the-art performance on Weibo while using 91% fewer parameters than the best comparable state-of-the-art baseline. For deployment in real-world applications, GAME-ON can be used as a lightweight model (less memory and latency requirements), which makes it more feasible than previous state-of-the-art models.

Abstract

The widespread adoption of Large Language Models (LLMs) and awareness around multilingual LLMs have raised concerns regarding the potential risks and repercussions linked to the misapplication of AI-generated text, necessitating increased vigilance. While these models are primarily trained for English, their extensive training on vast datasets covering almost the entire web, equips them with capabilities to perform well in numerous other languages. AI-Generated Text Detection (AGTD) has emerged as a topic that has already received immediate attention in research, with some initial methods having been proposed, soon followed by the emergence of techniques to bypass detection. In this paper, we report our investigation on AGTD for an indic language Hindi. Our major contributions are in four folds: i) examined 26 LLMs to evaluate their proficiency in generating Hindi text, ii) introducing the AI-generated news article in Hindi (AG_hi) dataset, iii) evaluated the effectiveness of five recently proposed AGTD techniques: ConDA, J-Guard, RADAR, RAIDAR and Intrinsic Dimension Estimation for detecting AI-generated Hindi text, iv) proposed Hindi AI Detectability Index (ADI_hi) which shows a spectrum to understand the evolving landscape of eloquence of AI-generated text in Hindi. The code and dataset is available at https://github.com/ishank31/Counter_Turing_Test

Abstract

The integration of Vision-Language Models (VLMs) into various applications has highlighted the importance of evaluating these models for inherent biases, especially along gender and racial lines. Traditional bias assessment methods in VLMs typically rely on accuracy metrics, assessing disparities in performance across different demographic groups. These methods, however, often overlook the impact of the model’s disabilities, like lack spatial reasoning, which may skew the bias assessment. In this work, we propose an approach that systematically examines how current bias evaluation metrics account for the model’s limitations. We introduce two methods that circumvent these disabilities by integrating spatial guidance from textual and visual modalities. Our experiments aim to refine bias quantification by effectively mitigating the impact of spatial reasoning limitations, offering a more accurate assessment of biases in VLMs.

Abstract

Indoor air pollution monitoring has been a region of interest in recent times. Multiple Internet of Things (IoT) enabled devices are available for this purpose. With the growing number of sensors in our daily environment, huge amounts of data are being collected and pushed to the servers through the Internet. This study aims to reveal that seemingly trivial indoor air pollution data containing particulate matter, carbon dioxide, and temperature can reveal complex insights about an individual’s lifestyle. Data was collected over a period of four months in a real-world environment. The study demonstrates the inference of cooking activities by using machine learning and deep learning techniques. The study further demonstrates that different food items and culinary practices have different air pollution signatures, which can be identified and distinguished with great accuracy (>90%). In the practice of inferential analysis, it is not necessary to rely on data characterised by high frequency or granularity. Less detailed data like hourly averages, can be used to make meaningful conclusions that might intrude on an individual’s privacy. With the rapid advancement in machine learning and deep learning, a proactive approach to privacy is needed to ensure that the collected data and its usage do not intentionally or unintentionally breach individual privacy.

Abstract

Certain self-supervised approaches to train image encoders, like CLIP, align images with their text captions. However, these approaches do not have an a priori incentive to learn to associate text inside the image with the semantics of the text. Humans process text visually; our work studies the semantics of text rendered in images. We show that the semantic information captured by image representations can decisively classify the sentiment of sentences and is robust against visual attributes like font and not based on simple character frequency associations.

Abstract

We present a method to compress the final linear layer of language models, reducing memory usage by up to 3.4 x without significant performance loss. By grouping tokens based on Byte Pair Encoding (BPE) merges, we prevent materialisation of the memory-intensive logits tensor. Evaluations on the TinyStories dataset show that our method performs on par with GPT-Neo and GPT2 while significantly improving throughput by up to 3x, making it suitable for low-compute environments.

Abstract

Continual learning has primarily focused on the issue of catastrophic forgetting and the associated stability-plasticity tradeoffs. However, little attention has been paid to the efficacy of continually learned representations, as representations are learned alongside classifiers throughout the learning process. Our primary contribution is empirically demonstrating that existing online continually trained deep networks produce inferior representations compared to a simple pre-defined random transforms. Our approach embeds raw pixels using a fixed random transform, approximating an RBF-Kernel initialized before any data is seen. We then train a simple linear classifier on top without storing any exemplars, processing one sample at a time in an online continual learning setting. This method, called RanDumb, significantly outperforms state-of-the-art continually learned representations across all standard online continual learning benchmarks. Our study reveals the significant limitations of representation learning, particularly in low-exemplar and online continual learning scenarios. Extending our investigation to popular exemplar-free scenarios with pretrained models, we find that training only a linear classifier on top of pretrained representations surpasses most continual fine-tuning and prompt-tuning strategies. Overall, our investigation challenges the prevailing assumptions about effective representation learning in online continual learning.

Abstract

In this paper, we tackle the complex task of analyzing televised debates, with a focus on a prime time news debate show from India. Previous methods, which often relied solely on text, fall short in capturing the multimedia essence of these debates [27]. To address this gap, we introduce a comprehensive automated toolkit that employs advanced computer vision and speech-to-text techniques for large-scale multimedia analysis. Utilizing state-of-the-art computer vision algorithms and speech-to-text methods, we transcribe, diarize, and analyze thousands of YouTube videos of prime-time television debates in India. These debates are a central part of Indian media but have been criticized for compromised journalistic integrity and excessive dramatization [18]. Our toolkit provides concrete metrics to assess bias and incivility, capturing a comprehensive multimedia perspective that includes text, audio utterances, and video frames. Our findings reveal significant biases in topic selection and panelist representation, along with alarming levels of incivility. This work offers a scalable, automated approach for future research in multimedia analysis, with profound implications for the quality of public discourse and democratic debate. We will make our data analysis pipeline and collected data publicly available to catalyze further research in this domain.

Abstract

Language models often exhibit undesirable behavior, e.g., generating toxic or gender-biased text. In the case of neural language models, an encoding of the undesirable behavior is often present in the model's representations. Thus, one natural (and common) approach to prevent the model from exhibiting undesirable behavior is to steer the model's representations in a manner that reduces the probability of it generating undesirable text. This paper investigates the formal and empirical properties of steering functions, i.e., transformation of the neural language model's representations that alter its behavior. First, we derive two optimal, in the least-squares sense, affine steering functions under different constraints. Our theory provides justification for existing approaches and offers a novel, improved steering approach. Second, we offer a series of experiments that demonstrate the empirical effectiveness of the methods in mitigating bias and reducing toxic generation.