Research Papers

Conversational Complexity for Assessing Risk in Large Language Models

John Burden, Manuel Cebrian, Jose Hernandez-Orallo

Large Language Models (LLMs) present a dual-use dilemma: they enable beneficial applications while harboring potential for harm, particularly through conversational interactions. Despite various safeguards, advanced LLMs remain vulnerable. A watershed case in early 2023 involved journalist Kevin Roose's extended dialogue with Bing, an LLM-powered search engine, which revealed harmful outputs after probing questions, highlighting vulnerabilities in the model's safeguards. This contrasts with simpler early jailbreaks, like the "Grandma Jailbreak," where users framed requests as innocent help for a grandmother, easily eliciting similar content. This raises the question: How much conversational effort is needed to elicit harmful information from LLMs? We propose two measures to quantify this effort: Conversational Length (CL), which measures the number of conversational turns needed to obtain a specific harmful response, and Conversational Complexity (CC), defined as the Kolmogorov complexity of the user's instruction sequence leading to the harmful response. To address the incomputability of Kolmogorov complexity, we approximate CC using a reference LLM to estimate the compressibility of the user instructions. Applying this approach to a large red-teaming dataset, we perform a quantitative analysis examining the statistical distribution of harmful and harmless conversational lengths and complexities. Our empirical findings suggest that this distributional analysis and the minimization of CC serve as valuable tools for understanding AI safety, offering insights into the accessibility of harmful information. This work establishes a foundation for a new perspective on LLM safety, centered around the algorithmic complexity of pathways to harm.

Breaking Down Financial News Impact: A Novel AI Approach with Geometric Hypergraphs

Anoushka Harit, Zhongtian Sun, Jongmin Yu +1

In the fast-paced and volatile financial markets, accurately predicting stock movements based on financial news is critical for investors and analysts. Traditional models often struggle to capture the intricate and dynamic relationships between news events and market reactions, limiting their ability to provide actionable insights. This paper introduces a novel approach leveraging Explainable Artificial Intelligence (XAI) through the development of a Geometric Hypergraph Attention Network (GHAN) to analyze the impact of financial news on market behaviours. Geometric hypergraphs extend traditional graph structures by allowing edges to connect multiple nodes, effectively modelling high-order relationships and interactions among financial entities and news events. This unique capability enables the capture of complex dependencies, such as the simultaneous impact of a single news event on multiple stocks or sectors, which traditional models frequently overlook. By incorporating attention mechanisms within hypergraphs, GHAN enhances the model's ability to focus on the most relevant information, ensuring more accurate predictions and better interpretability. Additionally, we employ BERT-based embeddings to capture the semantic richness of financial news texts, providing a nuanced understanding of the content. Using a comprehensive financial news dataset, our GHAN model addresses key challenges in financial news impact analysis, including the complexity of high-order interactions, the necessity for model interpretability, and the dynamic nature of financial markets. Integrating attention mechanisms and SHAP values within GHAN ensures transparency, highlighting the most influential factors driving market predictions. Empirical validation demonstrates the superior effectiveness of our approach over traditional sentiment analysis and time-series models.

The Exame Nacional do Ensino Médio (ENEM) is a pivotal test for Brazilian students, required for admission to a significant number of universities in Brazil. The test consists of four objective high-school level tests on Math, Humanities, Natural Sciences and Languages, and one writing essay. Students' answers to the test and to the accompanying socioeconomic status questionnaire are made public every year (albeit anonymized) due to transparency policies from the Brazilian Government. In the context of large language models (LLMs), these data lend themselves nicely to comparing different groups of humans with AI, as we can have access to human and machine answer distributions. We leverage these characteristics of the ENEM dataset and compare GPT-3.5 and 4, and MariTalk, a model trained using Portuguese data, to humans, aiming to ascertain how their answers relate to real societal groups and what that may reveal about the model biases. We divide the human groups by using socioeconomic status (SES), and compare their answer distribution with LLMs for each question and for the essay. We find no significant biases when comparing LLM performance to humans on the multiple-choice Brazilian Portuguese tests, as the distance between model and human answers is mostly determined by the human accuracy. A similar conclusion is found by looking at the generated text as, when analyzing the essays, we observe that human and LLM essays differ in a few key factors, one being the choice of words where model essays were easily separable from human ones. The texts also differ syntactically, with LLM generated essays exhibiting, on average, smaller sentences and less thought units, among other differences. These results suggest that, for Brazilian Portuguese in the ENEM context, LLM outputs represent no group of humans, being significantly different from the answers from Brazilian students across all tests.

Algorithmic bias in healthcare mirrors existing data biases. However, the factors driving unfairness are not always known. Medical devices capture significant amounts of data but are prone to errors; for instance, pulse oximeters overestimate the arterial oxygen saturation of darker-skinned individuals, leading to worse outcomes. The impact of this bias in machine learning (ML) models remains unclear. This study addresses the technical challenges of quantifying the impact of medical device bias in downstream ML. Our experiments compare a "perfect world", without pulse oximetry bias, using SaO2 (blood-gas), to the "actual world", with biased measurements, using SpO2 (pulse oximetry). Under this counterfactual design, two models are trained with identical data, features, and settings, except for the method of measuring oxygen saturation: models using SaO2 are a "control" and models using SpO2 a "treatment". The blood-gas oximetry linked dataset was a suitable test-bed, containing 163,396 nearly-simultaneous SpO2 - SaO2 paired measurements, aligned with a wide array of clinical features and outcomes. We studied three classification tasks: in-hospital mortality, respiratory SOFA score in the next 24 hours, and SOFA score increase by two points. Models using SaO2 instead of SpO2 generally showed better performance. Patients with overestimation of O2 by pulse oximetry of > 3% had significant decreases in mortality prediction recall, from 0.63 to 0.59, P < 0.001. This mirrors clinical processes where biased pulse oximetry readings provide clinicians with false reassurance of patients' oxygen levels. A similar degradation happened in ML models, with pulse oximetry biases leading to more false negatives in predicting adverse outcomes.

On the Evaluation Consistency of Attribution-based Explanations

Jiarui Duan, Haoling Li, Haofei Zhang +5

Attribution-based explanations are garnering increasing attention recently and have emerged as the predominant approach towards \textit{eXplanable Artificial Intelligence}~(XAI). However, the absence of consistent configurations and systematic investigations in prior literature impedes comprehensive evaluations of existing methodologies. In this work, we introduce {Meta-Rank}, an open platform for benchmarking attribution methods in the image domain. Presently, Meta-Rank assesses eight exemplary attribution methods using six renowned model architectures on four diverse datasets, employing both the \textit{Most Relevant First} (MoRF) and \textit{Least Relevant First} (LeRF) evaluation protocols. Through extensive experimentation, our benchmark reveals three insights in attribution evaluation endeavors: 1) evaluating attribution methods under disparate settings can yield divergent performance rankings; 2) although inconsistent across numerous cases, the performance rankings exhibit remarkable consistency across distinct checkpoints along the same training trajectory; 3) prior attempts at consistent evaluation fare no better than baselines when extended to more heterogeneous models and datasets. Our findings underscore the necessity for future research in this domain to conduct rigorous evaluations encompassing a broader range of models and datasets, and to reassess the assumptions underlying the empirical success of different attribution methods. Our code is publicly available at \url{https://github.com/TreeThree-R/Meta-Rank}.

The development of AI-driven generative audio mirrors broader AI trends, often prioritizing immediate accessibility at the expense of explainability. Consequently, integrating such tools into sustained artistic practice remains a significant challenge. In this paper, we explore several paths to improve explainability, drawing primarily from our research-creation practice in training and implementing generative audio models. As practical provisions for improved explainability, we highlight human agency over training materials, the viability of small-scale datasets, the facilitation of the iterative creative process, and the integration of interactive machine learning as a mapping tool. Importantly, these steps aim to enhance human agency over generative AI systems not only during model inference, but also when curating and preprocessing training data as well as during the training phase of models.

Identifying the relevant input features that have a critical influence on the output results is indispensable for the development of explainable artificial intelligence (XAI). Remove-and-Retrain (ROAR) is a widely accepted approach for assessing the importance of individual pixels by measuring changes in accuracy following their removal and subsequent retraining of the modified dataset. However, we uncover notable limitations in pixel-perturbation strategies. When viewed from a geometric perspective, we discover that these metrics fail to discriminate between differences among feature attribution methods, thereby compromising the reliability of the evaluation. To address this challenge, we introduce an alternative feature-perturbation approach named Geometric Remove-and-Retrain (GOAR). Through a series of experiments with both synthetic and real datasets, we substantiate that GOAR transcends the limitations of pixel-centric metrics.

Self-supervised models have revolutionized speech processing, achieving new levels of performance in a wide variety of tasks with limited resources. However, the inner workings of these models are still opaque. In this paper, we aim to analyze the encoded contextual representation of these foundation models based on their inter- and intra-model similarity, independent of any external annotation and task-specific constraint. We examine different SSL models varying their training paradigm -- Contrastive (Wav2Vec2.0) and Predictive models (HuBERT); and model sizes (base and large). We explore these models on different levels of localization/distributivity of information including (i) individual neurons; (ii) layer representation; (iii) attention weights and (iv) compare the representations with their finetuned counterparts.Our results highlight that these models converge to similar representation subspaces but not to similar neuron-localized concepts\footnote{A concept represents a coherent fragment of knowledge, such as ``a class containing certain objects as elements, where the objects have certain properties. We made the code publicly available for facilitating further research, we publicly released our code.

The neonatal period is the most vulnerable time for the development of seizures. Seizures in the immature brain lead to detrimental consequences, therefore require early diagnosis. The gold-standard for neonatal seizure detection currently relies on continuous video-EEG monitoring; which involves recording multi-channel electroencephalogram (EEG) alongside real-time video monitoring within a neonatal intensive care unit (NICU). However, video-EEG monitoring technology requires clinical expertise and is often limited to technologically advanced and resourceful settings. Cost-effective new techniques could help the medical fraternity make an accurate diagnosis and advocate treatment without delay. In this work, a novel explainable deep learning model to automate the neonatal seizure detection process with a reduced EEG montage is proposed, which employs convolutional nets, graph attention layers, and fully connected layers. Beyond its ability to detect seizures in real-time with a reduced montage, this model offers the unique advantage of real-time interpretability. By evaluating the performance on the Zenodo dataset with 10-fold cross-validation, the presented model achieves an absolute improvement of 8.31% and 42.86% in area under curve (AUC) and recall, respectively.

Large Language Models (LLMs) are increasingly adopted in educational contexts to provide personalized support to students and teachers. The unprecedented capacity of LLM-based applications to understand and generate natural language can potentially improve instructional effectiveness and learning outcomes, but the integration of LLMs in education technology has renewed concerns over algorithmic bias which may exacerbate educational inequities. In this review, building on prior work on mapping the traditional machine learning life cycle, we provide a holistic map of the LLM life cycle from the initial development of LLMs to customizing pre-trained models for various applications in educational settings. We explain each step in the LLM life cycle and identify potential sources of bias that may arise in the context of education. We discuss why current measures of bias from traditional machine learning fail to transfer to LLM-generated content in education, such as tutoring conversations because the text is high-dimensional, there can be multiple correct responses, and tailoring responses may be pedagogically desirable rather than unfair. This review aims to clarify the complex nature of bias in LLM applications and provide practical guidance for their evaluation to promote educational equity.

Radiomics and artificial Intelligence for thyroid cancer diagnosis: Concepts, challenges, and solutions

Milad Yousefi, Hadi Vahedi, Shadi Farabi Maleki +8

Thyroid cancer is an increasing global health concern that requires advanced diagnostic methods. The application of AI and radiomics to thyroid cancer diagnosis is examined in this review. A review of multiple databases was conducted in compliance with PRISMA guidelines until October 2024. A combination of keywords led to the discovery of an English academic publication on thyroid cancer and related subjects. 368 papers were returned from the original search after 112 duplicates were removed. Relevant studies were selected according to predetermined criteria after 176 articles were eliminated based on an examination of their abstract and title. After the comprehensive analysis, an additional six studies were excluded. Among the 42 included studies, radiomics analysis, which incorporates ultrasound (US) images, demonstrated its effectiveness in diagnosing thyroid cancer. Various results were noted, some of the studies presenting new strategies that outperformed the status quo. The literature has emphasized various challenges faced by AI models, including interpretability issues, dataset constraints, and operator dependence. The synthesized findings of the 42 included studies mentioned the need for standardization efforts and prospective multicenter studies to address these concerns. Furthermore, approaches to overcome these obstacles were identified, such as advances in explainable AI technology and personalized medicine techniques. The review focuses on how AI and radiomics could transform the diagnosis and treatment of thyroid cancer. Despite challenges, future research on multidisciplinary cooperation, clinical applicability validation, and algorithm improvement holds the potential to improve patient outcomes and diagnostic precision in the treatment of thyroid cancer.

Causal networks are widely used in many fields, including epidemiology, social science, medicine, and engineering, to model the complex relationships between variables. While it can be convenient to algorithmically infer these models directly from observational data, the resulting networks are often plagued with erroneous edges. Auditing and correcting these networks may require domain expertise frequently unavailable to the analyst. We propose the use of large language models such as ChatGPT as an auditor for causal networks. Our method presents ChatGPT with a causal network, one edge at a time, to produce insights about edge directionality, possible confounders, and mediating variables. We ask ChatGPT to reflect on various aspects of each causal link and we then produce visualizations that summarize these viewpoints for the human analyst to direct the edge, gather more data, or test further hypotheses. We envision a system where large language models, automated causal inference, and the human analyst and domain expert work hand in hand as a team to derive holistic and comprehensive causal models for any given case scenario. This paper presents first results obtained with an emerging prototype.