Understanding where drivers direct their visual attention during driving, known as gaze behavior, is important for developing next-generation advanced driver-assistance systems and improving road safety. This paper addressed this challenge as a semantic identification task from road scenes captured by a vehicle’s front-view camera. Specifically, the collocation of gaze points with object semantics was investigated using three distinct vision-based approaches: direct object detection, segmentation-assisted classification, and query-based vision-language models. The results demonstrate that the direct object detection and vision language model–based approach outperformed other approaches, achieving macro F1-scores exceeding 0.84. The larger vision-language model, in particular, exhibited better robustness and performance for identifying small, safety-critical objects such as traffic lights, especially in nighttime conditions. Conversely, the segmentation-assisted paradigm suffers from a part-versus-whole semantic gap, which led to low recall. The results reveal a trade-off between the real-time efficiency of traditional detectors and the contextual understanding and robustness offered by large vision-language models. These findings provide guidance for the design of future human-aware driver monitoring systems.

In particle and nuclear physics, “detector effects unfolding” can be viewed as a high-dimensional inverse problem whose goal is to recover the true event distributions from observed experimental data corrupted by detector-induced distortions. Recent advances in generative AI have positioned data-driven and machine learning-based approaches as powerful alternatives to traditional unfolding techniques, offering superior scalability to high-dimensional data, capability of learning complex detector responses, and the ability to operate directly at the event level. We survey state-of-the-art generative AI-based models for detector folding and unfolding. We review existing architectures and training strategies, and highlight recent methodological advances and open challenges. Through a detailed discussion of latent space representations, uncertainty quantification, physics insights, and background removals, we demonstrate the potential to significantly advance the field.

Large language models (LLMs) have achieved unprecedented performance in various tasks, such as classification, summarization, reasoning, and complex decision-making. But, as the size and parameter count of these models has increased exponentially, their computational demand, electricity consumption, and carbon footprint has also risen dramatically. This rapid scaling has put the global research community in front of a serious environmental concern. In response to this growing concern, researchers are gradually shifting their focus from red AI (compute-intensive, resource-heavy systems) to green AI. Particularly, small language models and optimization techniques like quantization, pruning, and knowledge distillation are being explored as sustainable alternatives to LLMs for domain-specific tasks. Covering the literature from 2020 to 2026, this review paper presents an extensive, deep, and rigorous bibliometric analysis. Using an advanced rule-based keyword preprocessing pipeline and Biblioshiny network mapping, we have mapped the annual publication growth, global research hubs, author influence networks, and dominant thematic clusters. Further, this review paper highlights a major research gap: Although many papers have already been written on energy optimization, publications on standardized carbon emission reporting frameworks are still limited.

Product attributes can be extracted from free-form text content using a large language model (LLM) with high precision. We propose repeating the extraction several times and selecting the most frequent extraction. We measure consistency as the percentage of the most frequent extraction and accept only the most consistent extractions. This approach reduces error rate from 0.5% with the baseline LLM to 0.2%, and, more importantly, eliminates all hallucinations (specificity increases from 99.1% to 100%). The method could also be used for LLM-assisted extraction where a human verifies the results by focusing on the less consistent extractions, and those attributes that the LLM did not find values for.

One solution to nearest neighbor outliers is to link them with the class cluster by adding new points between the outlier and the cluster. In the age of Generative AI, new data points are legitimate, and if they are added equally and only when needed, they might help re-balance the dataset. In this paper, we describe a nearest neighbor algorithm called Step Nearest Neighbor (Step NN), where incorrectly identified points can be re-linked by adding new points between them and the main cluster. The step points can also be noted, so that they can, for instance, be ignored as real cases. As new points were added, it worked only with distance-based metrics, however. While other adaptive versions may have to fine-tune parameters, this method was easier to use and as it could be pre-trained, this offers options for big data and distributed environments. Test results showed that accuracy comparisons with the classic version of k-NN is almost equal, but Step NN can generalise. In the cases where k-NN did well, k-NN should be equal to or better than Step NN. In the cases where k-NN did not do well, Step NN is likely to be better.