The recent advancements and emergence of rapidly evolving AI models, such as large language models (LLMs), have sparked interest among researchers and professionals. These models are ubiquitously being fine-tuned and applied across various fields such as healthcare, customer service and support, education, automated driving, and smart factories. This often leads to an increased level of complexity and challenges concerning the trustworthiness of these models, such as the generation of toxic content and hallucinations with high confidence leading to serious consequences. The European Union Artificial Intelligence Act (AI Act) is a regulation concerning artificial intelligence. The EU AI Act has proposed a comprehensive set of guidelines to ensure the responsible usage and development of general-purpose AI systems (such as LLMs) that may pose potential risks. The need arises for strengthened efforts to ensure that these high-performing LLMs adhere to the seven trustworthiness aspects (data governance, record-keeping, transparency, human-oversight, accuracy, robustness, and cybersecurity) recommended by the AI Act. Our study systematically maps research, focusing on identifying the key trends in developing LLMs across different application domains to address the aspects of AI Act-based trustworthiness. Our study reveals the recent trends that indicate a growing interest in emerging models such as LLaMa and BARD, reflecting a shift in research priorities. GPT and BERT remain the most studied models, and newer alternatives like Mistral and Claude remain underexplored. Trustworthiness aspects like accuracy and transparency dominate the research landscape, while cybersecurity and record-keeping remain significantly underexamined. Our findings highlight the urgent need for a more balanced, interdisciplinary research approach to ensure LLM trustworthiness across diverse applications. Expanding studies into underexplored, high-risk domains and fostering cross-sector collaboration can bridge existing gaps. Furthermore, this study also reveals domains (like telecommunication) which are underrepresented, presenting considerable research gaps and indicating a potential direction for the way forward.

In explainable AI, many explanation methods generate similar yet diverging explanations for machine learning (ML) models. How fair is it then to explain ML model behaviour by such explanations? Arguably, one needs to judge whether those explanations are good at explaining ML model input/output behaviour. We here attempt to formalise ways to judge goodness of such explanations in terms of their correctness. For assessing correctness, one needs to have desirable properties of explanation correctness in mind, as well as was to measure satisfaction of those properties. We submit two high-level properties of soundness and completeness for assessing explanation correctness: explaining is sound if the model behaves the way the explanations say; explaining is complete if explanations can be given for model’s outputs on any inputs. We formulate soundness and completeness properties for three forms of explanations: feature importance, counterfactuals and rules. We further formalise multiple general metrics, at least one for each property and form of explanation, for quantitatively measuring satisfaction of soundness and completeness. We argue that explanations are correct in as much as various aspects of the different forms of explanations are met as quantified by those metrics. We hope that being able to assess correctness of ML model input/output behaviour explanations against formal properties and metrics is a substantial step towards fairly explaining ML-based inference.

In Frequency Division Duplex (FDD) systems, for efficient communication, the downlink Channel State Information (CSI) should be sent to the base station through feedback links. Since such transmissions come with the cost of signaling overhead, the state-of-the-art has approaches for data-driven compression of CSI using auto-encoders and other Machine Learning (ML) algorithms. However, models built on a particular training dataset need additional domain transfer overhead for different test settings and environments. We propose a causality-aware Channel State Information (CSI) encoding system that adapts to changes in input data distribution as follows: (a) Create a model of the underlying constraints that generate the observational data, e.g., using Structural Causal Models (SCMs), where the model (e.g., SCMs) captures the cause-and-effect relationships between the observational or endogenous variable (i.e., channel state information) and the unobserved or exogenous variables (e.g., User Equipment (UE) speed, frequency, etc.). The causal graph is represented by directed acyclic graphs (DAGs), where nodes (vertices) correspond to the endogenous variables, and the directed edges account for the causal parent-child relationship. In this scenario, the endogenous variable is the input vector of vectors, $H$ , whereas the exogenous variables include the vendor and non-vendor specific parameters, e.g., power thresholds, channel quality indicator, etc., as detailed in the document, and (b) Perform domain adaptation by applying the learned Structural Causal Model to the received data, e.g., by using a causal layer in the neural network. To the best of our knowledge, this work first demonstrates the efficacy of causality-aware CSI compression and its usefulness in domain adaptability, out-of-distribution generalization, and power savings.