Research Papers

Explanations of Large Language Models Explain Language Representations in the Brain

Maryam Rahimi, Mohammad Reza Daliri, Yadollah Yaghoobzadeh

Large Language Model (LLM) representations are known to align with brain activity during language processing, but it remains unclear what drives this alignment. We test whether explainable AI (XAI) can help answer this: using attribution methods, we quantify the contribution of each input word to an LLM's next-word predictions and use these explanations to predict fMRI data from participants listening to narratives. We find that gradient-based attribution methods robustly align with brain activity, contribute unique variance beyond acoustic and word-rate confounds, and outperform internal representations in early auditory regions. Using conductance, we extend attribution from words to individual layers, asking what each layer's attribution reveals about the model's computation and how this relates to its brain alignment. Early layers show greater word-type sensitivity and align preferentially with auditory regions, whereas the final layer's attribution is dominated by positional information and exhibits broad cortical alignment. Together, these findings demonstrate that attribution-based explanations can be used not only to measure LLM--brain alignment but to characterize what it reflects.

Perception module of Autonomous vehicles (AVs) are increasingly susceptible to be attacked, which exploit vulnerabilities in neural networks through adversarial inputs, thereby compromising the AI safety. Some researches focus on creating covert adversarial samples, but existing global noise techniques are detectable and difficult to deceive the human visual system. This paper introduces a novel adversarial attack method, AdvSwap, which creatively utilizes wavelet-based high-frequency information swapping to generate covert adversarial samples and fool the camera. AdvSwap employs invertible neural network for selective high-frequency information swapping, preserving both forward propagation and data integrity. The scheme effectively removes the original label data and incorporates the guidance image data, producing concealed and robust adversarial samples. Experimental evaluations and comparisons on the GTSRB and nuScenes datasets demonstrate that AdvSwap can make concealed attacks on common traffic targets. The generates adversarial samples are also difficult to perceive by humans and algorithms. Meanwhile, the method has strong attacking robustness and attacking transferability.

xai_evals : A Framework for Evaluating Post-Hoc Local Explanation Methods

Pratinav Seth, Yashwardhan Rathore, Neeraj Kumar Singh +2

The growing complexity of machine learning and deep learning models has led to an increased reliance on opaque "black box" systems, making it difficult to understand the rationale behind predictions. This lack of transparency is particularly challenging in high-stakes applications where interpretability is as important as accuracy. Post-hoc explanation methods are commonly used to interpret these models, but they are seldom rigorously evaluated, raising concerns about their reliability. The Python package xai_evals addresses this by providing a comprehensive framework for generating, benchmarking, and evaluating explanation methods across both tabular and image data modalities. It integrates popular techniques like SHAP, LIME, Grad-CAM, Integrated Gradients (IG), and Backtrace, while supporting evaluation metrics such as faithfulness, sensitivity, and robustness. xai_evals enhances the interpretability of machine learning models, fostering transparency and trust in AI systems. The library is open-sourced at https://pypi.org/project/xai-evals/ .

Towards Unified Attribution in Explainable AI, Data-Centric AI, and Mechanistic Interpretability

Shichang Zhang, Tessa Han, Usha Bhalla +1

The increasing complexity of AI systems has made understanding their behavior critical. Numerous interpretability methods have been developed to attribute model behavior to three key aspects: input features, training data, and internal model components, which emerged from explainable AI, data-centric AI, and mechanistic interpretability, respectively. However, these attribution methods are studied and applied rather independently, resulting in a fragmented landscape of methods and terminology. This position paper argues that feature, data, and component attribution methods share fundamental similarities, and a unified view of them benefits both interpretability and broader AI research. To this end, we first analyze popular methods for these three types of attributions and present a unified view demonstrating that these seemingly distinct methods employ similar techniques (such as perturbations, gradients, and linear approximations) over different aspects and thus differ primarily in their perspectives rather than techniques. Then, we demonstrate how this unified view enhances understanding of existing attribution methods, highlights shared concepts and evaluation criteria among these methods, and leads to new research directions both in interpretability research, by addressing common challenges and facilitating cross-attribution innovation, and in AI more broadly, with applications in model editing, steering, and regulation.

Facial expression recognition (FER) has emerged as a promising approach to the development of emotion-aware intelligent agents and systems. However, key challenges remain in utilizing FER in real-world contexts, including ensuring user understanding and establishing a suitable level of user trust. We developed a novel explanation method utilizing Facial Action Units (FAUs) to explain the output of a FER model through both textual and visual modalities. We conducted an empirical user study evaluating user understanding and trust, comparing our approach to state-of-the-art eXplainable AI (XAI) methods. Our results indicate that visual AND textual as well as textual-only FAU-based explanations resulted in better user understanding of the FER model. We also show that all modalities of FAU-based methods improved appropriate trust of the users towards the FER model.

Large language models (LLMs) have shown promising capabilities in visually interpreting medical time-series data. However, their general-purpose design can limit domain-specific precision, and the proprietary nature of many models poses challenges for fine-tuning on specialized clinical datasets. Conversely, small specialized models (SSMs) offer strong performance on focused tasks but lack the broader reasoning needed for complex medical decision-making. To address these complementary limitations, we introduce \ConMIL{} (Conformalized Multiple Instance Learning), a novel decision-support framework distinctively synergizes three key components: (1) a new Multiple Instance Learning (MIL) mechanism, QTrans-Pooling, designed for per-class interpretability in identifying clinically relevant physiological signal segments; (2) conformal prediction, integrated with MIL to generate calibrated, set-valued outputs with statistical reliability guarantees; and (3) a structured approach for these interpretable and uncertainty-quantified SSM outputs to enhance the visual inspection capabilities of LLMs. Our experiments on arrhythmia detection and sleep stage classification demonstrate that \ConMIL{} can enhance the accuracy of LLMs such as ChatGPT4.0, Qwen2-VL-7B, and MiMo-VL-7B-RL. For example, \ConMIL{}-supported Qwen2-VL-7B and MiMo-VL-7B-RL both achieves 94.92% and 96.82% precision on confident samples and (70.61% and 78.02%)/(78.10% and 71.98%) on uncertain samples for the two tasks, compared to 46.13% and 13.16% using the LLM alone. These results suggest that integrating task-specific models with LLMs may offer a promising pathway toward more interpretable and trustworthy AI-driven clinical decision support.

Improved Detection and Diagnosis of Faults in Deep Neural Networks Using Hierarchical and Explainable Classification

Sigma Jahan, Mehil B Shah, Parvez Mahbub +1

Deep Neural Networks (DNN) have found numerous applications in various domains, including fraud detection, medical diagnosis, facial recognition, and autonomous driving. However, DNN-based systems often suffer from reliability issues due to their inherent complexity and the stochastic nature of their underlying models. Unfortunately, existing techniques to detect faults in DNN programs are either limited by the types of faults (e.g., hyperparameter or layer) they support or the kind of information (e.g., dynamic or static) they use. As a result, they might fall short of comprehensively detecting and diagnosing the faults. In this paper, we present DEFault (Detect and Explain Fault) -- a novel technique to detect and diagnose faults in DNN programs. It first captures dynamic (i.e., runtime) features during model training and leverages a hierarchical classification approach to detect all major fault categories from the literature. Then, it captures static features (e.g., layer types) from DNN programs and leverages explainable AI methods (e.g., SHAP) to narrow down the root cause of the fault. We train and evaluate DEFault on a large, diverse dataset of ~14.5K DNN programs and further validate our technique using a benchmark dataset of 52 real-life faulty DNN programs. Our approach achieves ~94% recall in detecting real-world faulty DNN programs and ~63% recall in diagnosing the root causes of the faults, demonstrating 3.92% - 11.54% higher performance than that of state-of-the-art techniques. Thus, DEFault has the potential to significantly improve the reliability of DNN programs by effectively detecting and diagnosing the faults.

As artificial intelligence (AI) becomes integral to economy and society, communication gaps between developers, users, and stakeholders hinder trust and informed decision-making. High-level AI labels, inspired by frameworks like EU energy labels, have been proposed to make the properties of AI models more transparent. Without requiring deep technical expertise, they can inform on the trade-off between predictive performance and resource efficiency. However, the practical benefits and limitations of AI labeling remain underexplored. This study evaluates AI labeling through qualitative interviews along four key research questions. Based on thematic analysis and inductive coding, we found a broad range of practitioners to be interested in AI labeling (RQ1). They see benefits for alleviating communication gaps and aiding non-expert decision-makers, however limitations, misunderstandings, and suggestions for improvement were also discussed (RQ2). Compared to other reporting formats, interviewees positively evaluated the reduced complexity of labels, increasing overall comprehensibility (RQ3). Trust was influenced most by usability and the credibility of the responsible labeling authority, with mixed preferences for self-certification versus third-party certification (RQ4). Our Insights highlight that AI labels pose a trade-off between simplicity and complexity, which could be resolved by developing customizable and interactive labeling frameworks to address diverse user needs. Transparent labeling of resource efficiency also nudged interviewee priorities towards paying more attention to sustainability aspects during AI development. This study validates AI labels as a valuable tool for enhancing trust and communication in AI, offering actionable guidelines for their refinement and standardization.

Long-form mental health assessments pose unique challenges for large language models (LLMs), which often exhibit hallucinations or inconsistent reasoning when handling extended, domain-specific contexts. We introduce Stacked Multi-Model Reasoning (SMMR), a layered framework that leverages multiple LLMs and specialized smaller models as coequal 'experts'. Early layers isolate short, discrete subtasks, while later layers integrate and refine these partial outputs through more advanced long-context models. We evaluate SMMR on the DAIC-WOZ depression-screening dataset and 48 curated case studies with psychiatric diagnoses, demonstrating consistent improvements over single-model baselines in terms of accuracy, F1-score, and PHQ-8 error reduction. By harnessing diverse 'second opinions', SMMR mitigates hallucinations, captures subtle clinical nuances, and enhances reliability in high-stakes mental health assessments. Our findings underscore the value of multi-expert frameworks for more trustworthy AI-driven screening.

As artificial intelligence (AI) becomes increasingly embedded in healthcare delivery, this chapter explores the critical aspects of developing reliable and ethical Clinical Decision Support Systems (CDSS). Beginning with the fundamental transition from traditional statistical models to sophisticated machine learning approaches, this work examines rigorous validation strategies and performance assessment methods, including the crucial role of model calibration and decision curve analysis. The chapter emphasizes that creating trustworthy AI systems in healthcare requires more than just technical accuracy; it demands careful consideration of fairness, explainability, and privacy. The challenge of ensuring equitable healthcare delivery through AI is stressed, discussing methods to identify and mitigate bias in clinical predictive models. The chapter then delves into explainability as a cornerstone of human-centered CDSS. This focus reflects the understanding that healthcare professionals must not only trust AI recommendations but also comprehend their underlying reasoning. The discussion advances in an analysis of privacy vulnerabilities in medical AI systems, from data leakage in deep learning models to sophisticated attacks against model explanations. The text explores privacy-preservation strategies such as differential privacy and federated learning, while acknowledging the inherent trade-offs between privacy protection and model performance. This progression, from technical validation to ethical considerations, reflects the multifaceted challenges of developing AI systems that can be seamlessly and reliably integrated into daily clinical practice while maintaining the highest standards of patient care and data protection.

Deep learning models in computer vision have made remarkable progress, but their lack of transparency and interpretability remains a challenge. The development of explainable AI can enhance the understanding and performance of these models. However, existing techniques often struggle to provide convincing explanations that non-experts easily understand, and they cannot accurately identify models' intrinsic decision-making processes. To address these challenges, we propose to develop a counterfactual explanation (CE) model that balances plausibility and faithfulness. This model generates easy-to-understand visual explanations by making minimum changes necessary in images without altering the pixel data. Instead, the proposed method identifies internal concepts and filters learned by models and leverages them to produce plausible counterfactual explanations. The provided explanations reflect the internal decision-making process of the model, thus ensuring faithfulness to the model.

Explainable AI (XAI) provides methods to understand non-interpretable machine learning models. However, we have little knowledge about what legal experts expect from these explanations, including their legal compliance with, and value against European Union legislation. To close this gap, we present the Explanation Dialogues, an expert focus study to uncover the expectations, reasoning, and understanding of legal experts and practitioners towards XAI, with a specific focus on the European General Data Protection Regulation. The study consists of an online questionnaire and follow-up interviews, and is centered around a use-case in the credit domain. We extract both a set of hierarchical and interconnected codes using grounded theory, and present the standpoints of the participating experts towards XAI. We find that the presented explanations are hard to understand and lack information, and discuss issues that can arise from the different interests of the data controller and subject. Finally, we present a set of recommendations for developers of XAI methods, and indications of legal areas of discussion. Among others, recommendations address the presentation, choice, and content of an explanation, technical risks as well as the end-user, while we provide legal pointers to the contestability of explanations, transparency thresholds, intellectual property rights as well as the relationship between involved parties.