Abstract: Contemporary intelligent systems increasingly participate in shaping, rather than merely supporting, human decision-making processes. While these systems enhance efficiency and predictive performance, they also introduce a critical but underexamined challenge: decisions may remain technically valid while becoming difficult for human actors to interpret and internalize. This misalignment is conceptualized in this study as the meaning gap, defined as the divergence between system output and human interpretive understanding.This paper proposes a human-centered governance framework that positions cognitive sovereignty—the capacity of individuals and institutions to interpret, contextualize, and assume responsibility for decisions—as a necessary condition for sustainable sociotechnical systems. The framework is structured around ten interdependent principles that collectively redefine governance as a cognitive architecture embedded within system design.Drawing on interdisciplinary literature in human–AI interaction, interpretability, and decision science, the study introduces latency as a conceptual construct describing temporal misalignment between system output and human interpretive readiness. This construct provides an integrative lens for understanding phenomena such as delayed comprehension, reduced accountability, and unstable trust in algorithmically mediated environments.Rather than treating interpretability as an auxiliary feature, the proposed framework positions it as a core system function. The paper further outlines potential pathways for operationalizing the meaning gap through measurable indicators, including time-to-comprehension, decision override frequency, and confidence misalignment.While existing frameworks have called for interpretability, few have proposed measurable indicators of cognitive alignment. This paper contributes preliminary metrics—time-to-comprehension, decision override frequency, and confidence misalignment—that operationalize the meaning gap for empirical testing.

Abstract: The construction industry faces a dual imperative: continued growth to meet the demands of a rapidly expanding global population, and deep decarbonisation to align with planetary boundaries and climate commitments embedded in frameworks such as the European Green Deal and the EU Bioeconomy Strategy. This paper examines the potential of bio-based construction materials to bridge these competing demands, reviewing evidence across a broad spectrum of material categories — including fast-growing plant-based materials, bio-based admixtures and polymer composites for concrete, bio-based polyurethanes, nanocellulose and cellulose aerogels, plant-based biocomposites, and mycelium-based composites. The review demonstrates that bio-based materials offer compelling environmental advantages over conventional petrochemical-derived alternatives, including superior carbon sequestration potential, reduced embodied carbon, improved indoor environmental quality, and compatibility with circular economy principles. The strategic urgency of this transition has been rendered concrete by the 2026 Strait of Hormuz crisis, which triggered severe disruptions to global petrochemical supply chains and exposed the structural vulnerability of European construction to fossil-derived material inputs — reframing bio-based alternatives as a supply security imperative alongside an environmental one. However, the transition from demonstrator projects to mainstream specification practice remains constrained by persistent technical, economic, and regulatory barriers, including inconsistencies in life cycle assessment methodologies, the absence of harmonised performance standards, certification gaps, high initial costs, and fragmented supply chains. Crucially, the review identifies that resolving these barriers depends not only on continued material innovation but equally on governance configurations, policy stability, and actor coalitions, with the conditions under which green finance, circular procurement, and regulatory instruments successfully accelerate material adoption varying substantially depending on who orchestrates systemic coordination.

Abstract: Herein, starting from (R)-(+)-α-methylbenzylamine, we report an efficient synthesis and full characterization of a new (R)-3-(1-hydroxyethylidene)-1-(1-phenylethyl)piperidine-2,4-dione, a new tetramic acid analogue. The key steps involved a non-classical Corey-Chaykovsky intramolecular cy-clization reaction to access the corresponding zwitterion, followed by a sequential desul-furization/reduction and condensation procedure. The key intermediate was obtained in 5 steps, and the desired product 7 with an overall 58% yield.

Abstract: Background/Objectives: Children who experience high levels of dental anxiety often show poor cooperation during dental visits, which compromises treatment outcomes and which leads to a vicious cycle with poor outcomes for oral health and caries prevention. The purpose of this study is to analyse the prevalence of dental anxiety in a sample of clinical subjects in children, and to examine possible correlations with other factors as well as the longitudinal nature of anxiety. Methods: This cross-sectional and longitudinal part study was conducted on a sample of 150 children aged 12 to 18 years in the city of Zagreb and Zagreb County. The MDAS (Modified Dental Anxiety Scale) scale was used to collect data related to dental anxiety at three-time intervals; before the procedure (T1), after the procedure (T2) and after a period of 3 months (T3). Results. Significant differences between the MDAS before and after the procedure were found, in the sense that dental anxiety decreased in the majority of children (p< 0.001). The MDAS test result was 9.79 for the T1 period, and 8.03 for the T2 period, which belongs to the mild anxiety group. There are statistically significant differences between individual time points T1, T2, T3 (p< 0.001), and these differences are significant between T1 and T2 (p=0.045), and T1 and T3 (p=0.012), while between T2 and T3 the differences are not statistically significant (p=0.616). At T1, most children had mild dental anxiety (55.3%), moderate (41.3%) and severe dental anxiety (3.3%). At T2, most children had mild dental anxiety (77.2%) and moderate (22.8%), with no severe dental anxiety. Financial impact on dental service use was statistically associat-ed with dental anxiety at T1 (p=0.015), and at T1-T2 period (p=0.032). The long period since the last visit to the dentist also showed significance for T1 (p=0.003) and T2 (p=0.014). The "urgent pain" showed a statistical correlation with the period T2-T1 (p=0.023). The greatest decrease in the dental anxiety scale T2-T1 was in subjects who had a doctorate/master's degree in their family, high income in family, regular dental check-ups within 3 months, brushing their teeth several times a day for over 2 minutes with horizontal brushing tech-nique, use of dental floss and fluoridated toothpaste, no active caries lesions, no bad habits, and use of drinking water and rarely eating sweets. Conclusions: Most of the children in this study have mild to moderate anxiety, and it de-creases after the therapeutic procedure. Increased dental anxiety is associated with urgent dental procedures/urgent pain and irregular check-ups that are longer than 3 months. Bet-ter oral hygiene and oral status, higher socioeconomic status, and a low cariogenic diet in-fluence the level of dental anxiety among children.

Abstract: In the contemporary landscape, natural language processing (NLP) stands as a vital force, empowering computers to comprehend and engage with human languages, thereby enhancing the realm of human-computer interaction (HCI) through the utilisation of large language models (LLMs) and multilingual pre-trained language models (mPLMs). The widespread adoption of these LLMs on a global scale is obvious. However, a critical observation reveals a significant gap in their capacity to effectively recognize some low-resource African languages, a concern observed by numerous researchers. This paper endeavours to contribute to the discourse by conducting a comprehensive metadata analysis of existing African language models. Through this investigation, the aim is to outline the importance, strengths, and weaknesses inherent in these models. By shedding light on these aspects, the paper seeks to not only underscore the current limitations but also to provide valuable insights and recommendations for future research endeavours in the domain of language recognition, particularly focusing on African languages. In doing so, the paper aspires to catalyse advancements that promote inclusivity and a more nuanced understanding of linguistic diversity within the realm of natural language processing. Multilingual Testing shall be used on Cheetah to evaluate the model's proficiency strength in multiple languages, including those that are less widely spoken such as Margi and Ibibio as well as identify any language-specific weaknesses or limitations of the LLMs, especially in recognizing and understanding languages like Margi spoken in the North-East geo-political zone of Nigeria and Ibibio spoken in the South-South geo-political zone of Nigeria.

Abstract: Micro-fragmented adipose tissue (mFAT) is a promising autologous biologic in regenerative medicine because it provides a mechanically processed adipose-derived product that preserves native extracellular matrix architecture and a cellular milieu rich in mesenchymal stem cells, pericytes, growth factors, cytokines, and extracellular vesicles. Mechanistically, mFAT is hypothesized to act largely through paracrine signaling that dampens inflammation, supports vascular stabilization, and promotes cartilage and soft-tissue repair; in vitro data suggest modulation of osteoarthritic synovial macrophage signaling, including reductions in chemokines such as CCL2 and CCL3. Preparation involves liposuction harvest followed by closed, sterile mechanical processing without enzymatic digestion or cell expansion, aligning with “minimal manipulation” concepts relevant to regulatory frameworks. Preclinical animal studies generally demonstrate favorable effects on synovial inflammation and cartilage matrix markers (e.g., glycosaminoglycan content) with limited adverse events. Clinically, the strongest body of evidence is in knee osteoarthritis, where multiple prospective and retrospective studies report improvements in pain and function from months to several years after single injections, though response rates vary and study designs are heterogeneous. Evolving data support potential benefit in hip osteoarthritis and select tendon conditions, but cohorts remain small. Overall, mFAT appears safe and potentially effective, yet larger, standardized, long-term randomized controlled trials and comparative studies versus platelet-rich plasma and bone marrow aspirate concentrates are needed to clarify indications, dosing, durability, and mechanisms in vivo.

Abstract: Nipah Virus (NiV) is a highly virulent zoonotic henipavirus responsible for recurrent outbreaks across South and Southeast Asia and remains a priority global health threat. Fruit bats of the genus Pteropus serve as the natural reservoir, with transmission to humans occurring through contaminated food products, infected intermediate hosts, or person-to-person spread. Clinical illness ranges from nonspecific febrile disease to rapidly progressive encephalitis and severe respiratory failure, with reported case fatality rates often exceeding 40 percent. This narrative synthesis reviews the evolving epidemiology, phylogenetic diversity, transmission pathways, clinical spectrum, diagnostic approaches, and current management strategies of NiV infection. Persistent seasonal spillovers, limited therapeutic options, the absence of licensed human vaccines, and challenges in early detection continue to hinder containment efforts. Strengthening One Health surveillance, expanding laboratory capacity, and accelerating vaccine and antiviral development remain critical priorities to mitigate future outbreaks and reduce associated morbidity and mortality.

Abstract: The self-attention mechanism has revolutionized sequence modeling but suffers from quadratic computational complexity with respect to sequence length, limiting its applicability to long sequences. We propose Sparse Projection Attention (SPA), a novel attention variant that leverages learnable sparse projections to reduce the effective dimensionality of queries and keys while maintaining expressive power. Our method is grounded in the Johnson-Lindenstrauss lemma and provides theoretical guarantees on distance preservation. We introduce a comprehensive mathematical framework including error bounds, convergence analysis, and gradient dynamics. Experimental results on language modeling, machine translation, and long-range sequence classification demonstrate that SPA achieves up to 8 × computational speedup while maintaining competitive performance compared to standard attention and other sparse variants. The proposed approach offers an effective trade-off between computational efficiency and model expressivity for long-sequence tasks, making transformers more accessible for resource-constrained environments and real-time applications.

Abstract: We study the problem of constructing group-invariant embeddings that faithfully represent data modulo group symmetries, a task that arises naturally in signal processing, physics, and machine learning. A central challenge is to design embeddings that are simultaneously orbit-separating, stable, and computationally tractable. To address this, we develop a general lifting framework for constructing such embeddings. The key idea is to start from a group-invariant embedding defined on a low-dimensional reduced space, and lift it to the ambient space by composing it with a finite family of parameterized linear maps, followed by an aggregation step that produces a global embedding. This framework provides a unified perspective that connects classical problems such as phase retrieval and permutation-invariant embeddings. We demonstrate the effectiveness of this framework in the finite group setting. In this setting, we establish general sufficient conditions for orbit separation and prove that any orbit-separating lifting embedding is automatically bi-Lipschitz. We further extend the bi-Lipschitz result to sparse regimes, and show that, when applied to phase retrieval, it yields an equivalence between uniqueness and stability for real sparse phase retrieval.

Abstract: Background Aorto-right ventricular fistula (ARVF) secondary to membranous septum rupture is an exceptionally rare complication after surgical aortic valve replacement (SAVR). While sutureless prostheses such as the Perceval valve have gained wide acceptance due to reduced cross-clamp times and procedural simplification, reported adverse events predominantly include conduction disturbances and paravalvular leaks. Structural septal disruption remains sparsely described. We report a case of early ARVF after Perceval implantation and review the pathophysiological and procedural mechanisms implicated in septal injury following sutureless and transcatheter aortic valve interventions. Case Description A 66-year-old woman with severe bicuspid aortic valve stenosis underwent SAVR via median sternotomy using a Perceval XL prosthesis after meticulous annular decalcification and sizing. Immediate intraoperative transesophageal echocardiography (TEE) confirmed optimal seating without paravalvular regurgitation. Within 24 hours, the patient developed complete atrioventricular block followed by cardiogenic shock. Repeat TEE revealed a large ARVF with significant left-to-right shunt. Emergent re-exploration identified a membranous septum tear. The Perceval prosthesis was explanted, the defect closed with reinforced patch repair, and a 27 mm Inspiris Resilia bioprosthesis was implanted. Peripheral veno-arterial ECMO support was required temporarily. The patient recovered and remained free of prosthetic dysfunction at two-year follow-up. Discussion Membranous septum rupture after AVR has an estimated incidence of 0.4-1.5 % in TAVR cohorts but is virtually unreported with Perceval valves. Mechanisms include chronic radial stress from oversized or malpositioned prostheses. Case reports with TAVR devices emphasize oversizing as a risk factor. Predictive factors for septal injury in sutureless AVR mirror those for conduction disturbances: valve oversizing, shallow infra-annular septal length, heavy calcification, and prior valve surgery. Preventive measures: strict sizing protocols, avoidance of balloon dilation, and optimized implantation depth, have reduced conduction complications and may mitigate septal trauma. Treatment choice percutaneous versus surgical closure, depends on hemodynamic stability, defect size and anatomy, and operative risk. Conclusion Early ARVF after Perceval implantation is exceedingly rare but potentially catastrophic. Strict adherence to sizing principles, awareness of septal anatomy, and prompt management, percutaneous in selected stable cases or surgical in acute large defects, are essential to optimize outcomes in sutureless AVR.

Abstract: Single-machine, single-product inventory models with generalized interarrival times have long lacked a rigorously justified and computationally tractable optimization framework, owing to a sign mis-specification in prior derivations that spawned ad hoc feasibility restrictions and expansive search domains. A corrected, succinct derivation establishes the strict convexity of the minimum-cost objective and proves the existence and uniqueness of an interior optimum without auxiliary conditions, unifying previously fragmented results into a general theorem. Leveraging these structural properties, the maximum-profit formulation is reduced to a single-variable program over natural, finite bounds that tightly bracket the optimizer, supplanting earlier paired bounds defined on an unbounded domain. Numerical evidence on a canonical benchmark shows that the admissible interval is markedly tighter yet attains the same optimum with fewer evaluations, thereby improving numerical efficiency and robustness of implementation. The analysis clarifies the correspondence between cost-minimization and profit-maximization formulations and provides an operationally simple, reproducible solution path for capacity-constrained single-machine systems under generalized interarrival times. Principal contributions are: (i) a corrected optimality theory establishing strict convexity and a unique interior optimum without auxiliary conditions; (ii) a dimensionality reduction of the profit model to a single-variable program with natural finite bounds; and (iii) demonstrably tighter admissible intervals that cut evaluations while preserving optimality.

Abstract: Cryptocurrencies are increasingly gaining traction in the digital realm, promising a decentralized future free from the grip of centralized authorities. This magnetic appeal has led to a surge in the integration of cryptocurrencies within various games and applications based on the robust security provided by blockchain technology. As the world embraces this digital revolution, everyday users are navigating a landscape filled with questions and concerns about the safety, privacy, and reliability of these innovative currencies. Some nations have chosen to ban or heavily regulate cryptocurrencies, further fanning the flames of uncertainty among regular users. While extensive research has been done on the technical dimensions of enhancing cryptocurrency security, there is a lack of appropriate work on critical aspects of parameters influencing users’ perspectives on security, privacy, and trust (SPT) offered by cryptocurrencies. This paper explores the existing gap by investigating the complex relationship between users’ perceptions of SPT in cryptocurrencies and the potential advantages presented by decentralized blockchains in current literature. PRISMA methodology has been followed to systematically review the existing literature targeting SPT parameters of cryptocurrencies with a detailed discussion of the methodologies followed by the researchers on the subject. After the careful selection of a search query, 64 papers have been reviewed in detail from a list of 350 papers obtained from Scopus, WoS, IEEE Explore, and ACM. Dominant use of surveys, the Technology Acceptance Model, and Structural Equation Modelling for analysis is observed in the reviewed literature which may not cover the complete domain of parameters affecting SPT concerns of users about cryptocurrencies. By exploring the existing literature, we highlighted the obstacles that may impede the widespread adoption of cryptocurrencies and the limitations that may be in research methodologies being adopted to measure these parameters.

Abstract: A single paragraph of about 200 words maximum. For research articles, abstracts should give a pertinent overview of the work. We strongly encourage authors to use the following style of structured abstracts, but without headings: (1) Background: Place the question addressed in a broad context and highlight the purpose of the study; (2) Methods: briefly describe the main methods or treatments applied; (3) Results: summarize the article’s main findings; (4) Conclusions: indicate the main conclusions or interpretations. The abstract should be an objective representation of the article and it must not contain results that are not presented and substantiated in the main text and should not exaggerate the main conclusions.

Abstract: We construct quantum Yang-Mills theory on four-dimensional Minkowski spacetime within the Epstein-Glaser causal perturbation theory framework, rigorously establishing the Wightman axioms and proving the existence of a positive mass gap \Delta >0 together with asymptotic freedom. The construction proceeds from two postulates—the massless wave equation \square \phi = 0 and Poincaré invariance—through the angular momentum decomposition of the retarded Green's function on the null cone. The equal-weight condition P_{\ell}(1) = 1, a consequence of the Peter-Weyl theorem for the unit element in every irreducible representation of \mathrm{SO}(3), ensures that all angular momentum modes contribute identically at the light-cone vertex. The spectral sum \Sigma^{(4)}(t) = \sum_{\ell = 0}^{\infty}(2\ell +1)e^{- (2\ell +1)t / 2} admits the closed form \cosh (t / 2) / [2\sinh^{2}(t / 2)], whose small-t expansion encodes the Riemann zeta function at negative odd integers via \zeta (- 1) = - 1 / 12, \zeta (- 3) = 1 / 120, etc. From the constant term 1 / 12 and the group-theoretic factor C_{2}(G), we derive the one-loop \beta-function coefficient b_{1} = 11C_{2}(G) / (12\pi) analytically without Feynman diagrams, establishing asymptotic freedom as a geometric consequence of null-cone causality. The mass gap is proven through two independent arguments: a distributional proof that non-abelian vertices extend propagator support to the timelike region, and a Carleman-Fredholm determinant argument excluding zero-mass poles. We verify all Wightman axioms via the reconstruction theorem. Furthermore, the framework reveals deep structural connections to random matrix theory (Dyson's threefold classification and Migdal's large-N reduction), and to number theory through the \mathrm{SL}(2,\mathbb{C}) holonomy R(2\pi) = - \mathbb{I} and the Selberg trace formula, providing a construction of the Hilbert-Pólya operator whose spectrum corresponds to the non-trivial zeros of the Riemann zeta function.

Abstract: Luminescent gadolinium oxide nanoparticles doped with europium were synthesized through a precipitation reaction using gadolinium and europium nitrates as precursors. The europium- doped gadolinium oxide nanoparticles were incorporated first: into a Gel matrix of silicon dioxide; and second: mixing with Polymethyl Methacrylate. Both processes are synthesized by the simultaneous hydrolysis of tetraethyl orthosilicate and polymerization of 3-(Trimethoxysilyl) propyl methacrylate. The solid samples obtained are round in shape with a size of about 2.5 cm, which makes it easy to handle to test different applications. The inclusion of Gd2O3:Eu3+ nanoparticles increases the level of absorbance in the ultraviolet region, which allows for improved emission of the material at a wavelength of around 609 nm. Furthermore, it enables easy doping of the material and the fabrication of thin films and monoliths with potential optical applications.

Abstract: We study a receiver-side closure built on the operational copy-time core of QICT and analyze its linear receiver-accessible sector to quadratic order in momentum. The logical scope is kept explicit throughout: the operational copy time is the only ingredient imported from QICT itself, whereas validator selection, shell geometry, local link structure, receiver maps, and benchmark dynamics are introduced as closure-level assumptions. Within this framework, we show that a calculable and falsifiable receiver-side inverse problem requires five structural ingredients: a validator sector, a certified transport support, a local covariance law, a closed benchmark normal form, and a declared receiver map. We then establish equivariant validator selection on homogeneous substrates, identify a six-cell contour as the minimal nearest-neighbor audited cycle on a codimension-one receiver shell with opposite pairing, and derive compact ring and spoke gauge sectors within the declared closure. For a real two-channel six-cell benchmark, we obtain the exact propagator, receiver-visible and receiver-dark transfer laws relative to a chirality-selective receiver, and the associated hydrodynamic limit. We further derive an algebraic reconstruction of the benchmark parameters from five spectral observables, analyze small-signal stability against weak nonlinear corrections, and formulate a set of platform-independent consistency tests based on dispersion asymmetry, short-time visible curvature, gap identities, and reduced rest-gap-time extraction. The paper also identifies a receiver-side non-identifiability boundary: the linear observable sector alone does not fix family structure, mixing data, anomaly coefficients, or absolute masses without additional microscopic input. No cosmological dark-matter interpretation is assumed; the receiver-dark sector is defined only relative to the declared receiver map. The result is a mathematically explicit and experimentally testable closure framework, rather than a derivation of the Standard Model from QICT alone.

Abstract: Human T-cell leukemia virus-1 (HTLV-1)-associated myelopathy/tropical spastic paraparesis (HAM/TSP) is a chronic inflammatory disease of the spinal cord induced by immunological activation due to high HTLV-1 proviral load in the peripheral blood. HAM/TSP is representative of HTLV-1-related inflammatory diseases and its main neurological symptoms—namely, motor dysfunction of the lower extremities through spastic paraparesis with urinary disturbance—are progressive and lead to deterioration in the quality of life of patients once these dysfunctions develop. Therefore, novel and safe therapeutic regimens are needed, enabling patients to commence treatment as soon as possible after the diagnosis of HAM/TSP. To date, various treatments have been developed for correction of the associated immunological or virological abnormalities, which have produced some good results. However, there are still many problems, such as insufficient treatment effects and side effects. In addition, most of these treatments are still in the open trial phase, and it remains unclear whether or not they are suitable for the long-term treatment of HAM/TSP induced by a chronic inflammatory status. Thus, we need safe therapeutic regimens for long-term or even lifelong courses of treatment. As treatments for HAM/TSP have not yet been established as therapeutic strategies, in this review, we summarize the clinical trials conducted to date, including representative regimens against HAM/TSP as existing therapeutic approaches, and discuss new regimens showing potential for further investigation as part of new therapeutic strategies in future research.

Abstract: Generative AI has emerged as a transformative force in cybersecurity, offering both opportunities for innovation and challenges in threat detection and mitigation. This systematic literature review and meta-analysis synthesizes existing research to evaluate the efficacy of generative AI in cybersecurity applications, focusing on detection performance, overall impact, and threat detection metrics. We conducted a comprehensive analysis of peer-reviewed studies, employing rigorous statistical methods to quantify effect sizes and their significance. The results reveal a substantial negative effect size for generative AI detection performance (d = −3.41, 95% CI [−3.42, −3.40], p < 1e−5), indicating a strong but counterintuitive trend that warrants further investigation. In contrast, the overall impact of generative AI on cybersecurity was negligible (d = −0.06, 95% CI [−0.31, 0.20], p = 0.68), suggesting a neutral net effect. However, generative AI demonstrated a statistically significant positive effect on threat detection metrics (d = 0.20, 95% CI [0.06, 0.35], p = 0.005), highlighting its potential to enhance specific security tasks. These findings underscore the dual nature of generative AI in cybersecurity, where its capabilities are context-dependent and require careful implementation. The study provides a foundational framework for future research, emphasizing the need for balanced approaches to harness generative AI’s benefits while mitigating its risks.

Abstract: Background: Surface electromyography (sEMG) and movement analysis are increasingly applied to characterize neuromuscular impairments and guide rehabilitation after stroke. Objectives: To synthesize recent literature on the application of sEMG and movement analysis in adult stroke rehabilitation, identify trends and gaps, and discuss implications for clinical practice and future research. Methods: A non-systematic scoping search was performed across PubMed, Scopus, Web of Science, and Google Scholar using combinations of “Movement analysis”, “Gait analysis”, “Electromyography”, and “Stroke.” The first 100 relevant articles (determined by title and abstract relevance) reaching data saturation were included. Data were extracted into a comparative table with fields for study descriptors, outcomes, main results, and clinical implications. Results: Publications increased from the 1990s with a concentration around 2017. Rehabilitation journals accounted for the largest share, followed by neuroscience and engineering. Motion analysis dominated study aims (62%); experimental designs were predominant (82%). Only a minority of studies used sEMG as a primary outcome measure. Most research focused on chronic stroke and lower-limb gait, though a substantial portion addressed upper-limb function. Limitations included methodological heterogeneity, underrepresentation of acute/subacute phases, and limited use of randomized designs. Conclusions: sEMG and movement analysis offer complementary, clinically relevant insights for personalized post-stroke rehabilitation, but broader, standardized adoption—particularly in acute/subacute settings and as routine outcome measures—is needed to translate advances into improved patient care.

Abstract: Downstream adaptation of a contrastively pretrained vision--language model can improve in-domain accuracy while degrading performance on unseen transfer tasks. This study examines how full fine-tuning and low-rank adaptation alter attention heatmaps under a controlled design that matches learning rate across adaptation methods. The completed matched-learning-rate matrix contains 80 runs using the OpenAI Contrastive Language--Image Pretraining model with a base 32-patch vision transformer image encoder, two datasets (EuroSAT and Oxford-IIIT Pets), four shared learning rates (1e-6, 5e-6, 1e-5, and 5e-5), and five random seeds. We measure classification-token-to-patch attention entropy, the fraction of patches required to capture 95\% of attention mass, attention concentration, head diversity, in-domain validation accuracy, and adapter-aware zero-shot accuracy on CIFAR-100. Three findings emerge. First, learning rate is a primary determinant of structural drift: on EuroSAT, full fine-tuning moves from entropy broadening at 1e-6 (+1.83%) to marked contraction at 5e-5 (-3.99%), whereas low-rank adaptation remains entropy-positive across the full matched grid (+0.68% to +1.50%). Second, low-rank adaptation preserves out-of-domain transfer substantially better than full fine-tuning at matched learning rates: averaged across the EuroSAT grid, zero-shot accuracy on CIFAR-100 is 45.13% for low-rank adaptation versus 11.28% for full fine-tuning; on Oxford-IIIT Pets, the corresponding averages are 58.01% and 8.54%. Third, Oxford-IIIT Pets exhibits a clear interaction with optimization scale: low-learning-rate low-rank adaptation underfits the in-domain task, so method-only averages can obscure the regime in which it becomes competitive. Additional rollout, patch-to-patch, centered-kernel-alignment, and backbone analyses are directionally consistent with these controlled results. Across both controlled datasets, runs with broader retained attention support also retain more zero-shot performance. Taken together, these findings support attention heatmap drift as an informative descriptive lens on model adaptation while arguing against a universal interpretation of the observed behavior as a single collapse phenomenon.