Existing LLM-based approaches remain limited by simplistic task execution, restricted tool integration, and a lack of contextual reasoning when interacting with professional GIS software. This study investigates the efficacy of a multi-agent architecture designed to enhance geospatial task execution accuracy through collaboration, reasoning and tool-use. The architecture integrates Chain of Thought (CoT) reasoning and Retrieval-Augmented Generation (RAG) and employs specialized agents that collaboratively translate high-level natural language queries into structured, executable workflows using QGIS processing algorithms as tools. Through a structured fine-tuning approach, we evaluated how the multi-agent setup influences spatial task comprehension, geoprocessing tool selection, and code generation. The results demonstrate that the developed framework significantly outperforms baseline single-agent and non-fine-tuned systems. For tasks involving one or two GIS tools, the system achieved up to 100% execution success and 87.5% semantic correctness. However, its effectiveness decreases with more complex, multi-step workflows. Notably, iterative self-refinement and self-debugging led to moderate gains in execution success and semantic correctness. The results indicate that multi-agent frameworks are a promising approach, even though improvements in reasoning depth and tool alignment are still needed. This study represents a foundational step toward more reliable, modular, and adaptable LLM-based systems for geospatial analysis automation.

One of the ways to enhance analyses conducted with large claims databases is by validating the measurement characteristics of the code-based algorithms used to identify health outcomes or other key study parameters of interest. These metrics can be used in quantitative bias analyses to assess the robustness of results for an inferential study, given potential bias from outcome misclassification. However, performing this validation through manual chart review of free-text notes from linked electronic health records requires extensive time and resource allocation. We describe an expedited process for validating code-based algorithms that introduces efficiency using two distinct mechanisms: (1) use of natural language processing to reduce the time spent by human reviewers to review each chart, and (2) a multiwave adaptive sampling approach with predefined criteria to stop the validation study once performance characteristics are identified with sufficient precision. We illustrate this process in a case study that validates the performance of a claims-based outcome algorithm for intentional self-harm in patients with obesity. We empirically demonstrate that the natural language processing-assisted annotation process reduced the time spent on review per chart by 40%, and the use of the predefined stopping rule with multiwave samples would have prevented review of 77% of patient charts with limited compromise to the precision of performance characteristics. This approach could facilitate more routine validation of code-based algorithms used to define key study parameters, ultimately enhancing understanding of the reliability of findings derived from database studies.

Eye-hand coordinated interaction is becoming a mainstream interaction modality in Virtual Reality (VR) user interfaces. Current paradigms for this multimodal interaction require users to learn predefined gestures and memorize multiple gesture-task associations, which can be summarized as an "Operation-to-Intent" paradigm. This paradigm increases users' learning costs and has low interaction error tolerance. In this paper, we propose SIAgent, a novel "Intent-to-Operation" framework allowing users to express interaction intents through natural eye-hand motions based on common sense and habits. Our system features two main components: (1) intent recognition that translates spatial interaction data into natural language and infers user intent, and (2) agent-based execution that generates an agent to execute corresponding tasks. This eliminates the need for gesture memorization and accommodates individual motion preferences with high error tolerance. We conduct two user studies across over 60 interaction tasks, comparing our method with two "Operation-to-Intent" techniques. Results show our method achieves higher intent recognition accuracy than gaze + pinch interaction (97.2% versus 93.1%) while reducing arm fatigue and improving usability, and user preference. Another study verifies the function of eye gaze and hand motion channels in intent recognition. Our work offers valuable insights into enhancing VR interaction intelligence through intent-driven design.

Automatic real image editing offers unprecedented freedom to modify the appearance of the image or to edit a few objects through natural language. Recent scalable model families such as diffusion models have showcased remarkable proficiency in editing highly realistic images due to the introduction of vast amounts of training data and large pretrained language models. However, these large diffusion models require iterative evaluation that would significantly hinder the pace of image editing. Moreover, the pioneering work in this field necessitates the learning of a unique textual token that corresponds to each input image, or a group of images containing the same object, leading to the generation of redundant and fragmented models. Given the aforementioned problems, we suggest a novel Large pretrained models Assistant text-guided image Editing adversarial Network (LAE-Net) in this paper. More concretely, we introduce a deep semantic editing network to globally transfer text information among different isolated editing blocks, which would extract features from the source image to differentiate text-required areas from text-irrelevant ones. Furthermore, based on idea that the multi-modal CLIP model, leveraging vision-language alignment, captures comprehensive global semantic cues, whereas the vision-centric DINO model specializes in delivering intricate, fine-grained pixel-level details, the powerful discriminator of LAE-Net is designed by harnessing the visual embeddings derived from both the CLIP and DINO models separately to boost the visual discriminant capability and facilitate training a strong generator for conditioning image generation. Comprehensive experimental evaluations show that our LAE-Net not only delivers outstanding performance but also surpasses several cutting-edge models.

Large Language Models are useless for linguistics, as they are probabilistic models that require a vast amount of data to analyze externalized strings of words. In contrast, human language is underpinned by a mind-internal computational system that recursively generates hierarchical thought structures. The language system grows with minimal external input and can readily distinguish between real language and impossible languages.

The predictive brain: Neural correlates of word expectancy align with large language model prediction probabilities.

We investigate how family characteristics contribute to language produced in homes. Specifically, we investigate the effects of older siblings on the amounts and types of child-directed and overheard speech that infants hear. We analyzed 76 daylong audio home recordings from 26 American families with infants between the ages of 6-12 months. To annotate the recordings, we randomly sampled 2 h from each recording comprised of 30s intervals and coded all sampled utterances for the speaker and addressee. We found that infants with older siblings had about half as much speech directed to them as did infants without older siblings. This effect was entirely driven by contexts in which the older sibling was physically present; we observed no differences in amounts of directed speech when older siblings were not present. However, when combining all child-directed input (to either the infant or the siblings) there were no differences between families with and without siblings. Further, the nature of overheard speech varied with the presence of a sibling, not only as an additional speaker in the home, but as in the physical presence of the sibling during day-to-day activities. The set of individuals in a home have substantial consequences for the speech that is produced and amounts of overheard- and child-directed speech may vary more widely in American homes than previously assumed.

Evolution of artificial intelligence at Hospital Italiano de Buenos Aires: A retrospective review of experience and lessons learned.

Studies based on electronic health records (EHR) often rely on structured data, which may incompletely capture important clinical phenotypes in EHR notes. The purpose of this study was to assess two natural language processing (NLP) tools to extract phenotypes from unstructured EHR notes, and to evaluate the added value of integrating NLP-derived phenotypes with structured EHR data at a health system scale. This retrospective study is based on inpatient and outpatient EHR data from the Mass General Brigham healthcare system between January 1, 2019 and December 31, 2020. Two established rule-based NLP tools were applied to extract smoking and obesity information from 19 215 303 clinical notes of 503 025 patients. NLP performance was evaluated through a manual review of stratified samples. Phenotype prevalence was estimated using structured EHR data alone and compared with prevalence estimates obtained by supplementing structured data with NLP-derived features. Both NLP tools exhibited high performance, with both accuracy and the F1 score of 0.99 for smoking, and 0.92 and 0.91 for obesity, respectively. The combination of NLP and structured data identified 220 714 patients (43.88%) with smoking, compared with 170 396 patients (33.87%) identified using structured data alone, representing a 29.5% relative increase. For obesity, NLP identified 121 360 patients (24.13%) from EHR notes, and 169 905 patients (33.78%) were documented in structured data; the inclusion of NLP-derived features contributed an additional 32 823 patients, corresponding to a 19.3% relative increase. NLP-derived phenotypes from unstructured EHR notes substantially improve patient identification for both smoking and obesity compared with structured EHR data alone at scale.

VQA explanation task aims to explain the decision-making process of VQA models in a way that is easily understandable to humans. Existing methods mostly use visual location or natural language explanation approaches to generate corresponding rationales. Although significant progress has been made, these frameworks are bottlenecked by the following challenges: 1) Uni-modal paradigm inevitably leads to semantic ambiguity of explanations. 2) The reasoning process cannot be faithfully responded to and suffers from logical inconsistency. 3) Human-annotated explanations are expensive and time-consuming to collect. In this paper, we introduce a new Semi-supervised VQA Multi-modal Explanation (SME) method via self-critical learning, which addresses the above challenges by leveraging both visual and textual explanations to comprehensively reveal the inference process of the model. Meanwhile, in order to improve the logical consistency between answers and rationales, we design a novel self-critical strategy to evaluate candidate explanations based on answer reward scores. More importantly, our method can benefit from a tremendous amount of samples without human-annotated explanations with semi-supervised learning. Extensive automatic measures and human evaluations all show the effectiveness of our method. Finally, the framework achieves a new state-of-the-art performance on the three VQA explanation datasets.

Differences in Optic Pathway Glioma Prevalence Among Children with Neurofibromatosis Type 1.

Established in 1988, Brazil’s Unified Health System (SUS) faces significant challenges regarding funding, regional inequalities, and resource oversight, necessitating innovative management solutions. Current auditing processes remain inefficient, as manually detecting irregularities within vast volumes of documents is both time-consuming and costly. This study applies Named Entity Recognition (NER) and text classification techniques to analyze health news relevant to SUS audits. We compare the performance of BERT, BERT-CRF, and ModBERTBr models to identify the most effective approach for optimizing content selection, thereby aiding investigations and combating corruption. A controlled experimental design was employed, following a pipeline of tokenization, label alignment, supervised training, and statistical analysis. Models were evaluated using accuracy, recall, precision, F1-score, and Mean Training Time (MTT). In the NER task, BERT-CRF demonstrated superior performance, achieving the best results in recall (0.880), precision (0.855), and F1-score (0.860). Conversely, the standard BERT model achieved the best overall performance in text classification, significantly outperforming ModBERTBr across all metrics.

In this response, we outline the central idea of a real pattern as applied in the philosophy of science as well as its relationship to the modal structure of models. We suggest that there may be additional concerns with linguistically real patterns in terms of convergence between natural language processing and formal linguistic theory.

High-quality data selection-driven instruction tuning for biomedical large language models.

Racial disparities in discretionary parole decisions persist, yet their cultural underpinnings remain poorly understood. The US justice system, shaped by White American cultural norms, may implicitly favor parole candidates who explain their crimes through dispositional rather than situational attributions. Importantly, however, findings from studies of cultural psychology suggest that while White Americans tend to frame behavior in dispositional terms, Black, Hispanic, and East Asian Americans are more likely to acknowledge situational contributors. Across four studies, we tested whether these attributional differences can contribute to racial disparities in parole. In studies 1 and 2 (total n = 588), White American mock parole candidates were more likely to use dispositional attributions to explain their hypothetical crimes, while those from racial and cultural minorities either showed no preference or favored situational explanations. In study 3 (n = 512), White American mock parole commissioners perceived candidates providing dispositional (vs. situational) explanations about their crimes as less likely to reoffend. In study 4, analyzing 183 real parole hearings via natural language processing methods, we found that candidates granted parole used dispositional framing of their past criminal offenses more often. These findings highlight a cultural mismatch between minority parole candidates and the expectations of the US justice system, potentially perpetuating racial disparities. Addressing this misalignment may offer a new pathway for reducing systemic bias in discretionary parole decisions.

To find agreement, conversation partners might focus solely on aligning their distinct perspectives. Alternatively, agreement might require partners to explore, by considering different, new perspectives. Using functional magnetic resonance imaging (fMRI) hyperscanning and natural language processing, we tested whether dyadic alignment or exploration during decision-making conversations was a more effective route to agreement. Dyads (N = 60 dyads; 120 people) discussed pressing societal problems after being instructed to either persuade or compromise with their partner, strategies associated with attenuated and amplified exploration, respectively. Analysis uncovered four key results: First, individuals instructed to compromise rather than persuade tended to agree more at the end of the conversation. Second, fMRI hyperscanning and linguistic analyses revealed that encouraging compromise resulted in increased exploration during conversations; dyads given compromise instructions traversed more diverse mental states and topics. Third, heightened exploration was linked to greater agreement at the end of the conversation. Fourth, the effect of the compromise instructions on agreement was mediated by the degree of exploration. Together, these results suggest that finding agreement may be spurred by exploration, something that happens spontaneously when people are motivated to compromise. (PsycInfo Database Record (c) 2026 APA, all rights reserved).

We argue that language models (LMs) have strong potential as investigative tools for probing the distinction between possible and impossible natural languages and thus uncovering the inductive biases that support human language learning. We outline a phased research program in which LM architectures are iteratively refined to better discriminate between possible and impossible languages, supporting linking hypotheses to human cognition.

Natural Language Processing (NLP) systems are increasingly deployed in high-stakes systems including healthcare, education, recruitment, and law enforcement, yet they have frequently coded and magnified biases that undercut their system’s fairness and trust. This review synthesizes and critically analyzes 121 studies that were published in the year 2014 and up to date that address bias in NLP. We present a novel taxonomy of 18 bias types, such as previously underexplored categories like geographic, disability, and annotation bias, and project them onto the NLP lifecycle, taking data as the starting point to deployment. Four key detection paradigms are examined (statistical, model-probing, benchmark-based, and human-centric), alongside mitigation strategies at the data, model, and post-processing levels. Unlike prior surveys, this study offers a lifecycle-aware framework that connects bias origins, detection methods, and mitigation practices, while focusing on persistent challenges such as intersectionality, generalization, and fairness–performance trade-offs in large language models (LLMs). We argue that achieving fairness in NLP requires not only technical interventions but also socio-technical approaches that integrate community participation, transparency, and governance. Offering a structured, critical, and forward-looking synthesis, this work contributes a roadmap for building transparent, equitable, and socially responsible NLP systems.

This study identifies specific language used for common problems seen in gynaecology. To better identify language used by patients for gynaecological conditions and understand how these terms may differ between communities. Observational qualitative study. An urban academic hospital. Adult female patients (aged 18 and older) seen in the gynaecology clinic were included, and those younger than 18 and non-English speakers were excluded. Interviews using a structured interview guide were conducted. New patients were enrolled in the study until the point of theoretical saturation. Key terms were identified and trends involving symptoms were investigated using natural language processing (NLP) techniques including lemmatization, text similarity and sentiment. Terminology gathered from study interviews was described. Thirty patients completed the study. Study participants included 12 (40%) Black patients, 9 (30%) White, 4 (13.3%) Hispanic and 3 (10.3%) Asian. Participants were equally distributed in age and most participants had attended some college 26 (86.6%). Almost half of the women had not been pregnant, 14 (46.7%). While many patients used some medical specific terminology, many also used symptoms specific common terms that were more descriptive of their experience. Patients identified 'cysts' more commonly than 'fibroids' and were concerned about cancer. Overwhelmingly patients identified communication problems between providers and patients resulting in misunderstandings and a lack of trust. By starting with identifying specific language used and expanding to identifying overall themes, we can improve understanding between patients and providers.

Language models excel at finding patterns in linguistic data, and can therefore prove insightful for statistical approaches to linguistics in that they provide further evidence for the strong reliance of natural languages on recurrent, fixed patterns. Nevertheless, regarding actual usage-based language processing, their implications are severely limited as they lack a crucial aspect of language use: interaction.