Computer Science and Mathematics

Abstract: Background: Diffusion-weighted imaging (DWI) is an important MRI technique for studying tissue microstructure and structural connectivity, but is limited by relatively low spatial resolution, signal-to-noise ratio (SNR), and sensitivity to inter-acquisition motion. A deep-learning framework was developed that jointly performs image registration and super-resolution on repeated low-resolution DWI inputs to generate a single registered, higher apparent spatial-resolution output. Methods: DeepSUM-DWI adapts a DeepSUM multi-image super-resolution architecture for repeated DWI. It comprises a shared 2D super-resolution feature extractor (SISRNet), a registration network (RegNet), and a 3D fusion network (FusionNet). Evaluation used (1) simulated repetitions with synthetic motion, (2) retrospectively downsampled acquisitions from 12 subjects, and (3) a prospective dual-resolution example with matched 1.0 × 1.0 and 2.0 × 2.0 mm² acquisitions. Results: DeepSUM-DWI reconstructed a single high-resolution-appearing DWI image from multiple low-resolution repetitions while preserving structural characteristics of the high-resolution acquisitions. The model remained robust across synthetic motion magnitudes. In retrospectively downsampled data, it improved NRMSE by 71–79% and PSNR by 10.6–13.5 dB, for super-resolution factors of 2–3, relative to averaging interpolated low-resolution images, with a mean SSIM > 0.99. In prospectively acquired data, it recovered finer structures than visible in the acquired low-resolution images while maintaining contrast consistency. Exploratory diffusion analysis suggested that some gains in apparent detail propagated to FA maps. Conclusions: DeepSUM-DWI provides a framework for joint registration and super-resolution of repeated DWI acquisitions, with potential to improve apparent spatial resolution while retaining the SNR benefits of repeated acquisition.

Abstract: Automatic humor detection, the computational task of identifying humorous content, is increasingly critical as Large Language Models (LLMs) integrate into communication platforms such as chatbots and virtual assistants. However, humor poses challenges for AI due to its reliance on context, cultural nuances, linguistic ambiguity, and multimodal cues. Current research is fragmented across humor types, languages, modalities, and evaluation benchmarks, particularly regarding modern LLM capabilities and limitations. This survey synthesizes the automatic humor detection field, tracing its evolution from foundational psychological and linguistic theories through classical machine learning, deep learning, and transformer-based LLMs. We organize and analyze computational methods, feature engineering techniques, benchmark datasets (text-only, multimodal, multilingual), and evaluation metrics. We examine LLM adaptation strategies, including fine-tuning, parameter-efficient methods (PEFT), prompt engineering, and multi-task learning, alongside multimodal and cross-lingual humor understanding developments. Our analysis reveals that while LLMs show improved performance in capturing surface humor patterns, gaps persist in pragmatic reasoning, cultural grounding, multimodal integration, and explainability. We identify open challenges, including data scarcity, evaluation inconsistencies, the humor–offensiveness boundary, and the need for culturally aware and interpretable models. By consolidating progress and pinpointing limitations, this survey guides future interdisciplinary research toward developing socially intelligent AI systems capable of understanding humor.

Abstract: Recognizing an object from a text label, describing an unfamiliar scene, and following an instruction about a small image region require different forms of visual-language correspondence. The development of vision-language models can be understood through the effort to support these increasingly varied outputs. This survey follows the transition from contrastive image-text representations to captioning and grounding systems, and then to instruction-tuned multimodal language models. Within this account, we examine how pretraining objectives, web-data curation, visual connectors, and adaptation methods affect the tasks a model can perform. Particular attention is given to novel object captioning and dense prediction: both expose information that can be lost when image-level alignment is used as the starting point for a more specific task. Relative contrastive alignment [Fan et al., 2023] and uncertainty-guided cross-head co-training [Fan et al., 2022] offer complementary case studies of supervision quality. Evaluation and hallucination are considered alongside the capabilities they are meant to measure. The review brings together architectural comparisons, task-oriented evidence, and unresolved questions about visual resolution, grounding, and benchmark validity. Its central concern is how a system’s training signal and representation support the visual claims it ultimately makes.

Abstract: Probabilistic-bit (p-bit) classifiers are trained by distilling a full-precision teacher into the stochastic binary network, usually by matching log-odds. A p-bit's firing probability is a sigmoid of its bias, so when the deployed network sees fewer sensors than its teacher, the output firing probability is whatever log-odds matching produces, which is not in general the Bayes posterior given the student's inputs. We characterise the gap for a frozen, Bayes-correct binary teacher. Given a teacher-confidence bound and the coarse posterior, a tangent construction gives the sharp attainable interval of the output firing probability; a classical third-convex-order result sharpens it given the conditional variance; a monotone optimisation map turns both into sharp log-loss and Brier envelopes, an exact reversal threshold with a safe window, and a read-budget envelope for a majority vote. A p-bit multilayer perceptron on a 16-cell exact-teacher panel, read through a simulated stochastic magnetic tunnel junction model, tracks the predicted firing probabilities up to sampling error and a small systematic shortfall at the two most extreme targets, reverses where predicted (more reads then converge to the wrong class), and shows the predicted read inflation. Four pre-specified tabular comparisons are null. The device is simulated; no energy claim is made.

Abstract: Structural topology optimization of large-scale engineering systems requires repeated compliance evaluations, making conventional finite element solvers computationally prohibitive in iterative design workflows. Physics-informed neural networks (PINNs) offer a mesh-free, differentiable surrogate alternative, but standard PINNs yield unstable sensitivity fields due to error amplification in higher-order automatic differentiation, undermining optimization reliability. This paper introduces a gradient-enhanced PINN surrogate (gSurrogate) that incorporates a residual-gradient penalty into the total potential energy loss, directly regularizing displacement field derivatives and stabilizing compliance sensitivities for practical structural design. The surrogate is validated on cantilever, simply supported, and fixed–fixed beam configurations and embedded within a SIMP-based topology optimization pipeline driven by the Method of Moving Asymptotes. Over 200 optimization iterations, the proposed surrogate reduces unfiltered sensitivity variance by 94.35\% and filtered variance by 81.74\% relative to a standard PINN, with variance standard deviation reduced by 97.16\%. Three-dimensional topology optimization across mesh resolutions of 120×20×20 and 60×60×20 demonstrates that the gSurrogate achieves displacement errors of 1.81\% and 1.63\% with respect to FEM (Abaqus), compared to 3.75\% and 3.45\% for a standard PINN, a consistent twofold accuracy improvement. These results establish the proposed surrogate as a computationally efficient and numerically reliable alternative to finite element solvers for compliance-driven topology optimization of structural systems.

Abstract: Auto-regressive decoding is limited by weight movement, not arithmetic. If an accurate 8-bit copy of the model does not fit in GPU memory, the usual choices are a coarser resident copy, which is fast but changes the model, or keeping part of the accurate copy in CPU memory, which preserves the model but is slow. Progressive Weight Refinement (PWR) splits each 8-bit round-to-nearest weight into a 4-bit high plane and a 4-bit low plane. The high planes stay on the GPU and are themselves a 4-bit model which drafts tokens. The low planes stay in pinned CPU memory and are streamed once per verification pass, overlapping the transfer with drafting, so several accepted tokens share one transfer. Because the draft is half of the target’s own bits, it costs no extra memory, and each verify moves only half of the 8-bit weights. The output matches the 8-bit target: greedy identity on 102,399 tokens from Qwen3-8B and 5,111 tokens from Qwen3-32B, and, for sampling, no deviation detected on 8B at a test resolution of about 0.02 total variation. On a 24 GB RTX 3090 Ti, where Qwen3-32B at 8-bit does not fit, PWR decodes at 8.99 tok/s, 1.80× llama.cpp’s automatic layer offload and 1.65x an upper bound on any whole-layer split. A three-term cost model built from per-layer transfer time and per-step draft time predicts measured round times to within 5% when transfer dominates. PWR is slower than resident 8-bit decode when the model fits and 4.6x slower than resident 4-bit on 32B, which costs 6–9% perplexity. It is a memory-placement method, not a general decoder speedup.

Abstract: Natural-language instructions can describe scenes through mandatory requirements, preferences, and contextual information. Conventional Wave Function Collapse (WFC) enforces local pattern compatibility but cannot, by itself, interpret these instructions or guarantee global requirements such as object counts and connectivity. We introduce Semantic-Constrained Probabilistic WFC (SC-PWFC), a compiler-generator-verifier framework that translates language into typed, executable scene specifications. SC-PWFC combines hard constraints, soft preferences, and contextual priors in an explicit probabilistic model. During scene construction, WFC-style propagation enforces local compatibility, global feasibility checks reject impossible partial scenes, and an independent verifier checks completed scenes. The framework also supports constrained completion, localized editing, and explicit handling of ambiguity and infeasibility. Controlled experiments demonstrate more accurate semantic compilation, more efficient constraint-aware generation, improved preference satisfaction and contextual consistency, and reliable preservation of protected content during editing. SC-PWFC thus extends WFC beyond local pattern generation towards language-guided, probabilistic, and verifiable scene synthesis.

Abstract: Generative artificial intelligence (GenAI), synthetic media, social bots, and agentic AI are turning information disorder from a content-verification problem into a governance problem for multimedia information ecosystems. Research on fake news, deepfakes, GenAI misuse, and cyber-enabled influence has developed in largely separate streams, and none of these streams offers an integrated governance model for influence activity coordinated through AI-supported workflows. This study develops such a model through a structured evidence synthesis of 35 peer-reviewed sources published between 2019 and 2026, identified through four structured discovery queries and citation chaining and appraised with evidence-type-specific criteria. The synthesis identifies five mechanisms through which AI reshapes information disorder: influence orchestration, ambiguity production, selective persuasion, hybrid human–machine dissemination, and a trust paradox in countermeasures. On this basis, the study defines agentic influence operations as workflow-level risks and organizes them in a twelve-dimension framework that links AI capabilities, content modality, platform infrastructure, data access, evidentiary trust, organizational harm, and governance control points. It also sets out construct boundary conditions, five propositions with explicit evidence status, a managerial control set, and a phased implementation roadmap. Only the proposition on synthetic media and uncertainty about authentic evidence has direct empirical support; the others are offered for testing through simulation, experimental, design-science, and expert-validation studies.

Abstract: The rapid proliferation of generative artificial intelligence (GenAI) and large language models (LLMs) has fundamentally transformed the landscape of modern higher education, prompting a critical paradigm shift in how academic integrity is conceptualized. This scoping review synthesizes 20 cutting-edge academic publications from 2026 (retrieved via Google Scholar) to examine the evolution of academic integrity from mechanical plagiarism detection and punitive surveillance toward comprehensive institutional, pedagogical, and ethical redesign. The synthesis reveals that the primary systemic catalyst for academic misconduct remains the pervasive “publish or perish” culture and quantitative assessment paradigms that prioritize performance over moral growth. Contemporary scholarship moves beyond identifying individual infractions to address systemic vulnerabilities, student uncertainty, data privacy, and the degradation of higher-order cognitive skills (analyzing, evaluating, creating) resulting from unreflective AI reliance. To mitigate these threats, the review highlights four key integrative strategies: (1) adoption of ecosystemic models (e.g., A. Maleki’s Higher Education Integrity Ecosystem) linking governance, assessment, cultural norms, and individual agency; (2) implementation of process-oriented, authentic assessment frameworks (e.g., S. Lehane, A. Wright and P. Fenton’s Model for Practice and P.D.N. Ncube et al.’s PRISMA-based multi-stage evaluations) that render automated cheating obsolete; (3) operationalization of the AI3 model that integrated AI, academic integrity, and assessment innovation (C. DeLuca, L.Volante and M. Holden), and (4) proactive educational initiatives, AI literacy, and epistemic pluralism to safeguard neurodivergent learning trajectories. Ultimately, the academic community is transitioning from technocratic panic to building ethically resilient, adaptive educational ecosystems aligned with UN Sustainable Development Goals (SDG 4 and SDG 16).

Abstract: Human oversight is a load-bearing premise of AI governance: in several regimes, human review determines which obligations apply. We examine when this premise fails for verification-based oversight, in which a designated human checks AI outputs before they take effect. It fails when accountability for detected errors falls on reviewers or deployers, reviewer capacity is bounded, and output from open populations persistently outgrows that capacity. Review then becomes nominal, and recent experiments show that fluent AI output undermines error recognition. Because accountability and feedback respond only to detected errors, undetected harms are externalized and surface as threshold shocks. We analyze how these dynamics limit six oversight configurations. As capability grows, the share of AI output that oversight in high-loss domains can responsibly approve contracts. Friction is the one role of human involvement that keeps functioning under load; bounding inflow to reviewer capacity (flow design) is a necessary complement to other safeguards.

Abstract: Agentic artificial intelligence combines goal-directed reasoning, planning, memory, tool use, adaptation and multi-agent coordination with the ability to affect external systems. This capability raises problems of authorization, safety, human oversight, accountability and the evidence required to justify reliance. This non-systematic integrative survey distinguishes AI ethics, responsible AI, governance, trustworthiness, assurance, explainability, security and controllability; synthesizes preventive, detective, corrective, human-based and agent-based controls; and compares responsibility and oversight arrangements across the agent lifecycle. Building on established governance and meaningful-human-control research, it proposes an authority-control-assurance-accountability mapping that connects normative requirements to delegated decision rights, enforceable controls, observable evidence, escalation conditions and identifiable organizational owners. Two fictional cases illustrate the proposed mapping for document access and software changes. The framework is a conceptual synthesis rather than an empirically validated assurance standard. The survey identifies limitations of supervisory-agent independence, long-horizon explanation, dynamic authorization and continuous re-assurance, and specifies research questions for future evaluation.

Abstract: Some sign language avatars work like a phrase book. They play a recorded sign for each word in their dictionary and fingerspell the other meaningful words, such as nouns, verbs and names, spelling them letter by letter with the hand alphabet, which takes longer than signing. This paper measures how much an International Sign avatar with several hundred signs fingerspells in English meeting transcripts and written sentences, and tests whether rewording the text into words that the avatar knows avoids the fingerspelling without changing the meaning. Lists of the most frequent English words stood in for larger dictionaries. Three kinds of rewording were compared, replacing each missing word by a similar word that has a sign, asking a small language model to simplify the text, and restricting that model to words with a sign. With the avatar’s own dictionary, most content words had to be fingerspelled, which took most of the signing time. Word replacement removed more than half of the fingerspelling almost instantly, at some cost in meaning, while simplification barely helped. The restricted model removed almost all fingerspelling and more than halved the signing time but lost much of the meaning, partly because it completed the first letters of a missing word into an unrelated word. With a dictionary of the 5,000 most frequent words, every method tested with it kept most of the meaning. Word replacement is the practical choice for small dictionaries, which should grow, starting with the most frequent words.

Abstract: Multimodal large language models (MLLMs) can generate plausible responses without using the relevant evidence or meeting task constraints. We separate the Modality Gap in cross-modal correspondence and evidence use from the Intent Gap in behavior relative to specified constraints, goals, and human expectations. Six dimensions organize the review: architecture, objective functions, data construction, training, evaluation, and applications. The first four control evidence access, optimization targets, supervision, and updates. Comparing mechanisms and supervision assumptions shows what interface compatibility, target imitation, preference learning, and verifiable rewards can each demonstrate about alignment. Evaluation at the representation, task, and behavior levels tests these choices through controlled input changes and reward-model validation, and we interpret each gain together with its model, task, and evaluation setup instead of comparing scores across conditions. Studies in robotics, safety, healthcare, recommendation, and climate decision support share three requirements: traceable evidence, valid supervision, and constraint-sensitive decisions. Stronger representations or higher benchmark scores can leave the claimed alignment mechanism unresolved, which motivates four research directions: evidence-consistent alignment across modalities, verified self-correction and test-time alignment, targeted repair from failure diagnosis, and reliable transfer to underrepresented settings.

Abstract: AI agents are increasingly expected to interact with changing users, tools, and environments over extended periods of time. However, current systems remain largely static: their parameters are fixed during deployment, while new experiences, domain knowledge, corrections, and behavioral adaptations are handled primarily through external memory management and in-context learning. In this position paper, we focus on lifelong AI: AI systems that can continually learn from their long-term experience throughout an extended deployment lifetime. This position paper argues that existing in-context mechanisms are insufficient for lifelong AI, and advocate In-Parameter Learning (IPL) as a necessary complement to In-Context Learning (ICL), referring to mechanisms that allow a deployed model to continually and safely update its parameters during use. More broadly, we aim to renew the community’s attention to in-parameter learning and to recalibrate the overly optimistic expectations placed on in-context learning alone. Our central claim is that future lifelong AI systems should be built around a hybrid paradigm: in-context learning for immediate, provisional, and reversible information, and in-parameter learning for durable, cumulative, and generalizable growth.

Abstract: Dyslexia is a common problem that affects how people read and write. It is not because they are not smart. Because their brain works a little differently. It is really important to find out if someone has dyslexia so we can help them with their reading and writing. The problem is that the ways we have now to figure out if someone has dyslexia are not very good. They take a lot of time. Are not always right. In this paper we are talking about a way to find out if someone has dyslexia. This way uses computers to look at how people write and talk. We use something called Natural Language Processing to look at the words and sentences that people write. We also use something called Sequence-to-Sequence to look at the patterns in their writing. This helps us to see if someone has dyslexia or not. We tried out our way with some tests and it worked pretty well. It was able to tell if someone had dyslexia or not 63 percent of the time. This is really good because it means that we can use this way to help kids who have dyslexia. It is also a way because it is not too expensive. This means that teachers and parents can use it to help kids with dyslexia. Dyslexia is a problem that we can help with if we can just find out who has it. Our new way is a step, in the right direction. Dyslexia detection is important. Our system can help with dyslexia detection.

Abstract: Machine learning is increasingly applied to midstream oil and gas pipelines, yet reported performance rarely establishes readiness for field use. This narrative review organizes peer-reviewed and industry sources from 2015–2025 using a five-level operational maturity framework (sensing and detection, diagnosis, prediction, decision and optimization, and human-supervised closed-loop automation), defined by a rubric of decision authority, operator role, and the validation evidence each level requires. A second axis records where each study’s performance was demonstrated, from simulation to closed-loop use. Applications include leak and anomaly detection, defect and corrosion assessment, hard spots, pump cavitation, predictive maintenance, failure-risk modeling, scheduling, and digital twins. Cross-classifying the coded sources on both axes shows that validation realism does not keep pace with claimed operational responsibility. Peer-reviewed evaluations are concentrated at Levels 1–3 and rely largely on laboratory, simulated, or retrospective data; the only operational Level 4 example is a practitioner case, and no coded source demonstrates Level 5 closed-loop control. The corpus thus shows a validation gap at the lower levels and an evidence-coverage gap at the upper ones; beyond the performance shown in controlled evaluations, validation, transferability, and governance remain major unresolved barriers to deployment. A ranked agenda led by graduated field validation follows.

Abstract: Yoga pose recognition and posture assessment have gained attention with advances in computer vision and machine learning. However, computationally intensive models and cloud-dependent processing can limit deployment on low-cost, resource-constrained devices. Tiny Machine Learning (TinyML) offers a promising approach by enabling on-device inference with limited memory, processing power, and energy consumption. This paper reviews existing research on yoga pose recognition, human pose estimation, lightweight neural networks, and embedded machine learning. A conceptual framework is proposed that integrates data acquisition, preprocessing, pose feature extraction, lightweight model inference, posture assessment, and feedback generation. Both camera-based and wearable sensor-based approaches are considered, highlighting their advantages and limitations for embedded deployment. Model optimization techniques, including quantization and pruning, are discussed as potential methods for reducing memory and computational requirements. The paper also examines challenges related to dataset diversity, model generalization, sensor reliability, privacy, and the accuracy of posture feedback. As this study presents a literature review and conceptual framework, no new implementation or experimental performance results are claimed. Future research should evaluate the proposed framework on physical embedded devices using recognition accuracy, memory usage, inference latency, and energy consumption. This work provides a foundation for developing efficient, privacy-conscious, and accessible yoga monitoring systems using TinyML.

Abstract: Generative artificial intelligence (GenAI) is increasingly embedded in students’ learning and knowledge-work routines, yet perceived productivity gains do not necessarily arise from technology acceptance alone. This study develops and empirically examines an integrated Technology Acceptance Model (TAM) and Self-Regulated Learning (SRL) framework for explaining learning productivity among Google/Gemini users in Bangladesh. The study dataset contains 400 respondent IDs, of which 385 records contain complete questionnaire responses and a valid attention-check response; 15 rows are blank. The analysis therefore uses 385 usable cases and conducts an additional sensitivity analysis on 302 respondents classified as regular or occasional AI-feature users. Perceived ease of use (PEOU), perceived usefulness (PU), SRL and learning productivity (LP) were operationalised using three five-point Likert items each. Transparent composite-score path analysis with 5,000 bootstrap resamples was used because the available analytical environment did not contain a dedicated CFA/PLS-SEM implementation; consequently, the reported measurement diagnostics are explicitly treated as pre-publication validation evidence rather than as a claim of completed covariance-based CFA or PLS-SEM. PEOU positively predicted PU (β = .278, p < .001), while both PEOU (β = .191, p < .001) and PU (β = .218, p < .001) predicted SRL. Learning productivity was predicted by PEOU (β = .140, p = .004), PU (β = .256, p < .001) and SRL (β = .202, p < .001), with R² = .188. Bootstrap estimates also supported indirect pathways from PEOU to LP through PU and SRL, including a serial PEOU → PU → SRL → LP effect. The results suggest that the productivity value of GenAI is partly behavioural: users benefit not only because the technology is perceived as useful and easy to operate, but also because those perceptions are associated with planning, monitoring and addressing learning gaps. The study contributes an emerging-market, mechanism-oriented account of GenAI-enabled learning productivity and identifies measurement and sampling issues that should be resolved in subsequent validation research.

Abstract: Hallucinations in large language models (LLMs) remain a critical challenge for deploying these systems in high-stakes domains. We introduce Calib-Abstain, a framework that combines token-level uncertainty estimation with multi-path consistency checking to generate calibrated confidence scores. When confidence falls below a learned threshold, the system triggers tiered mitigation actions: retrieval augmentation, query reformulation, or abstention with explanation. Our approach employs temperature scaling and isotonic regression to minimize Expected Calibration Error (ECE). Experiments on mathematical reasoning (GSM8K, MATH), code generation (HumanEval), and general knowledge tasks (MMLU) demonstrate that Calib-Abstain reduces hallucination rates among answered queries by 24--31% while maintaining 80--88% coverage, with ECE improvements of 0.08--0.12 across datasets.

Abstract: Large language models (LLMs) reach users only after a chain of transfer steps: full or parameter-efficient fine-tuning, preference alignment, prompting, and changes to the decoding procedure. Each step alters the model, the distribution it faces, or both. Yet fairness and robustness are usually assessed once, on a single fixed test set. This paper presents a unified group-conditional framework to analyze how transfer affects the fairness and robustness of LLMs, and how existing methods try to preserve these properties. To characterize these trade-offs, we introduce a analytical framework that views fair and robust transfer as constrained distributionally robust optimization. We view fair and robust transfer as constrained distributionally robust optimization whose feasible set is determined by the transfer interface, and we split any change in a fairness gap into an update term, caused by changing the model, and shift term, caused by changing the data. Building on existing fairness-transfer bounds, we record two elementary consequences that the LLM literature has not made explicit: the group-conditional bound on the shift term is tight, and when only one group shifts, the marginal shift equals exactly a \( \pi_{\min} \) fraction of the fairness-relevant shift, where \( \pi_{\min} \) is that group's proportion. We use these observations to classify mitigation methods by interface and by the term they target, to offer a reading under which conflicting findings on low-rank adaptation are compatible, and to derive reporting recommendations. We close with open problems, including per-group certificates, fairness of approximate decoding, and cumulative drift under continual transfer.