Computer Science and Mathematics

Abstract: Large language models (LLMs) are being adopted quickly across biomedicine. Their value in clinical practice depends on more than accuracy: it also depends on whether they can run within the latency, hardware, privacy, cost, and staffing limits of real settings. In this scoping review of efficiency-oriented biomedical LLM research, we organize the literature along two axes: a taxonomy of efficiency techniques (prompting and retrieval, parameter-efficient and data-efficient adaptation, model compression, efficient architectures and inference, and agentic workflows) and a map of biomedical application domains. Across the corpus, prompting and parameter-efficient fine-tuning delivered efficiency most often, and studies reported it as savings in memory, trainable parameters, compute time, and human-workflow time, while energy and carbon were almost never measured. The reported gains are large and concrete: low-rank adaptation combined with low-precision quantization often shrinks the memory needed for adaptation enough to train and deploy a model on a single consumer or edge device, usually at a small and measured cost in task quality. Yet the evidence still leans toward retrospective benchmarking, with external validation, prospective evaluation, and clinical deployment all rare. We map which techniques serve which clinical domains, show how to read an efficiency claim against its comparator and clinical context, and identify what the field still needs to measure to turn demonstrated resource savings into validated clinical value.

Abstract: Self-attention couples every pair of positions in a sequence, incurring a cost that grows quadratically with se-quence length and a key–value state that grows without bound during generation. Here we introduce the inductive algebraicresonance memory (IARM), an attention-free sequence-mixing mechanism that replaces pairwise attention with two ingre-dients: a softmax-gated mixture of learned low-rank operators that modulates a positive feature map of the query, and acoordinatewise causal memory that reads each feature channel as a normalized cumulative average of the value stream.The resulting layer runs in time linear in sequence length and maintains a constant-size recurrent state per channel, so au-toregressive decoding requires no growing cache. We instantiate IARM as a 118 003 200-parameter decoder-only languagemodel, pretrain it on a 10-billion-token educational web corpus, and fine-tune it for two epochs on 207 865 UltraChat con-versations, reducing the assistant-token perplexity from 20.4 to 6.90. Diagnostics over all 528 learned operators show thatthey converge to near rank-1 maps, that the four operators in each head are engaged in balance, and that the effectivememory window sharpens monotonically with network depth. A custom groupwise 4-bit export reproduces the weightsto within a mean relative error of 9.3 % while compressing the model 3.67-fold, and is loaded by an accompanying Apple-silicon (MLX) runtime that ports the mechanism one-to-one. IARM is offered as a transparent, fully reproducible testbedfor attention-free causal modelling rather than as a benchmarked production model.

Abstract: Domain shift remains a major obstacle to robust histopathology image segmentation, especially when models trained on several source organs are deployed to unseen anatomical sites. This study addresses cross-organ adenocarcinoma segmentation by introducing an evidence-guided vision-language segmentation framework that incorporates pathology-relevant morphological evidence into dense mask prediction. The proposed method uses a pathology vision-language encoder to extract image and text representations, a semantic query booster to form image-aware segmentation queries, and an evidence-guided decoding module that integrates positive tumour-supporting evidence and negative misleading evidence. A feature-level style regularization branch is further used during training to improve robustness to source-domain appearance variation. Experiments were conducted on cross-organ adenocarcinoma datasets under a source-only domain generalization setting, using colorectum, stomach, and pancreas as source domains and ampullary, gallbladder, and intestine as unseen target domains. The proposed framework achieved 79.48%/88.57% IoU/Dice on seen source organs and 79.89%/88.82% on unseen target organs, with a small seen-unseen gap of 0.41 IoU points and 0.25 Dice points. These results suggest that structured pathology evidence can provide useful semantic guidance for cross-organ tumour segmentation and may reduce reliance on organ-specific visual shortcuts.

Abstract: When using the few-shot recognition model, due to the influence of materials, light directions and background textures of the industrial inspection images, there will be domain drift, and the recognition accuracy of the model will be unstable when used in different equipment. We built a new dataset with 54, 000 images of 5 types of equipment and 5 types of defects to improve the generalization ability of defect recognition in new equipment scenarios. We present a cross domain few shot recognition model that combines ConvNeXt texture features extraction, domain adversarial feature alignment and prototype network discrimination. The experimental results demonstrate that the model gets the average accuracy of 88.3% and F1 score of 84.7% in defect classification with 50 labeled target samples per class. For the unseen equipment, the recall rates for superficial cracks and pitting corrosion were 81.6% and 85.2%, respectively. The results show that the method can further improve the cross-domain aggregation function for similar defects and improve the stability of the model transfer and deployment in complex industrial scenes.

Abstract: The textile and clothing value chain faces increasing pressure to improve reuse and recycling rates, particularly in post-consumer scenarios where garments must be rapidly assessed, classified, and routed toward appropriate end-of-life pathways. Manual sorting remains labor-intensive and difficult to scale, especially when garments must be evaluated according to visual condition, type, color, fiber-related characteristics, and reuse potential. This article proposes and evaluates an AI-based framework for automated clothing sorting to support textile reuse and recycling. The proposed framework combines YOLO-based object detection with a Vision-Language Model for semantic interpretation in identifying garments, interpreting visual attributes, and supporting sorting decisions. The framework first distinguishes reusable garments from non-reusable items and then further classifies, using two ConvNeXt-based visual classifiers, non-reusable textiles according to characteristics relevant to recycling. By integrating computer vision and vision-language reasoning, the approach aims to improve the speed, consistency, and scalability of garment classification in circular textile systems. The results demonstrate the potential of AI-assisted sorting as a decision-support tool for increasing the recovery value of post-consumer clothing and reducing the amount of textile waste directed to landfill or incineration.

Abstract: Quantum Machine Learning (QML) has been proposed as a framework that may offer theoretical advantages over classical machine learning, especially in computational complexity and parallel processing of high-dimensional data. However, due to the limitations imposed by the Noisy Intermediate-Scale Quantum (NISQ) era, implementing large-scale quantum algorithms remains infeasible, making hybrid quantum-classical approaches a more viable alternative. In this work, we formulate and implement custom quantum-hybrid versions of classical clustering algorithms and compare their performance on an Autism Spectrum Disorder (ASD) screening dataset, which represents a highly relevant clinical domain characterized by a complex mix of behavioral and demographic features. We evaluate classical k-means, DBSCAN, spectral clustering, and agglomerative clustering against these formulated quantum-hybrid implementations. For the evaluation, we use inner metrics (Silhouette, Davies-Bouldin, Calinski-Harabasz) and outer metrics (AMI, ARI) on the generated partitions. The results show that the formulated hybrid implementations achieve a partition quality comparable to that of classical counterparts, except in the case of Q-DBSCAN, where the classical algorithm outperformed our hybrid implementation due to the high-dimensional mapping. Thus, we present empirical information about strengths, weaknesses and potential viability of these formulated hybrid implementations in real-life applications in the NISQ era.

Abstract: Self-improving autonomous agents are moving from research prototypes to deployed systems. The primary goal is controllable evolution, or adaptation, from experience with minimal or even no human input. This survey frames modern self-improving agents as adaptive systems that convert experience into accumulated capability gains. We offer a system-level framework that represents a modern agent as a configuration coupling a foundation model with an operational scaffold of prompts, memory, tools, and control logic. Within this framework, self-improvement is formalized as a self-induced update operator that obtains and commits updates to model parameters or scaffold components. We organize prior work by update target and by the signals that drive change, then review applications and discuss evaluation, before closing with open problems and future directions. For convenience, we track technical updates on this GitHub page.

Abstract: Scientific AI systems can generate hypotheses and explanations, but many opti-mize plausibility more than refutability. This paper presents a falsification-drivenmulti-agent framework in which specialized agents propose hypotheses, build causalmodels, design adversarial tests, and verify formal claims. The architecture com-bines hypothesis generation, causal reasoning, falsification, formal verification, andpersistent orchestration through a shared memory state that records assumptions,counterexamples, interventions, and proof obligations. By treating failed predictionsand invalid proof attempts as useful learning signals, the framework shifts discoveryfrom fluent claim production toward disciplined claim survival. On simulated dis-covery tasks, the full system improves verified discovery rate from 0.42 to 0.76 andreduces false-positive hypothesis retention from 0.31 to 0.08. Scaling experimentsshow a peak discovery quality factor of 0.83 with eight agents, supporting the prin-ciple that scientific AI should prioritize systematic refutation, causal identifiability,and machine-checkable proof.

Abstract: Tabular anomaly detection (TAD) assigns anomaly scores to unusual samples in tables and supports applications in finance, healthcare, cybersecurity, industrial monitoring, and data-quality assurance. Its central challenge stems from the heterogeneous and weakly structured nature of tabular data: unlike images, sequences, and graphs, tables lack native spatial, temporal, or relational structure while mixing heterogeneous feature types without a common metric, so the notions of distance, density, and dependency that detectors rely on must be induced through representation and encoding choices. Part of what defines an anomaly may also reside outside processed values, in column semantics, domain rules, or schema information that standard pipelines discard. Recent TAD research spans non-deep detectors, deep task-specific models, large language model-based methods, and foundation model-based methods. These directions differ in supervision, information access, evaluation protocols, and downstream use, yet existing surveys examine neighboring areas largely in isolation. This survey provides a unified account of TAD by organizing existing detectors according to how they form anomaly scores. Beyond method taxonomy, it reviews enhancement and adaptation for deployment, benchmarks and evaluation protocols, and downstream tasks built on TAD outputs. It aims to clarify method assumptions and provide a more comparable basis for future TAD research.

Abstract: Continual Learning (CL) rose to prominence with the rise of deep architectures, and then entered a winter as large language models (LLMs) gave the impression that learning over time was no longer needed. We argue that this winter was not a failure but, in fact, a necessary transition. Intelligence cannot be separated from the time in which an agent lives and acts, and the arrival of agentic systems built on large models returns CL to its proper place: not the narrow question of how representations are formed, but the principle by which agents organise their knowledge as its environment changes. We call these systems Continual Learning Agents, designed from the start for adaptation and consolidating what they learn. We follow this path from the early ideas about machines that learn over time, through the era of deep CL, towards an holistic view in which learning continuously and the design of the agent can no longer be held apart, and in which the hard problems of Artificial Intelligence (AI) and those of CL are seen to converge.

Abstract: This study analyzes patterns of stress and exhaustion among student pilots throughout flight training using a combination of physiological and self-reported measurements. The Perceived Stress Scale (PSS-10) was used to measure perceived stress and exhaustion before and after each flight, while physiological data, including heart rate (HR), electrodermal activity (EDA), skin temperature, and acceleration, were continuously recorded during flight sessions. To identify recurring patterns in arousal and workload, physiological signals were preprocessed and analyzed across the flight stages. The findings indicate a buildup of workload-related weariness over time, as evidenced by steady increases in EDA and skin temperature across flights, as well as post-flight increases in self-reported exhaustion. Heart rate responses were more event-specific, with brief spikes during high-demand phases of flight. Overall, the findings demonstrate the value of combining physiological signals with subjective reports to identify patterns of stress and fatigue during real-world flight training and highlight the potential of data-driven approaches for monitoring pilot well-being.

Abstract: Organizations in the Arab world run employee and stakeholder surveys on tools that were built for English first. Arabic support in those tools is usually a translated interface on top of an English analytical pipeline, and the analysis itself tends to stop at raw response counts. This paper describes OrgPulse AI, a web platform I built to measure organizational health in Arabic and English through surveys structured around weighted thematic axes. The platform computes a deterministic set of metrics without any AI involvement: axis scores normalized to a 0-100 scale, top-box favorability, a consensus index derived from response dispersion, a worst-case margin of error, and an improvement priority ranking defined as axis weight multiplied by its performance gap. A second, optional layer adds inferential statistics on top of these descriptives: Welch's t-test for segment comparisons, normal-approximation confidence intervals, and Cronbach's alpha for axis-level internal consistency, all implemented from first principles and verified against reference values. Large language models sit outside this statistical core as a design and narration layer: GPT-4o drafts survey structures, questions, and weights during creation, and Claude Sonnet 4 writes executive narratives over the computed results. The paper also describes a portal mechanism that lets a parent organization share reports with subsidiary or client entities through PIN-gated pages, a pattern that matches how Saudi government bodies distribute assessment results to affiliated units. I state the platform's limitations plainly: the instrument is user-defined rather than psychometrically validated, scores aggregate at the question level rather than the respondent level, and the exploratory dialect-signal layer described in Section 5.2 has not been validated against human-coded ground truth. The contribution is an architectural and measurement pattern, documented from the actual implementation, for teams building survey analytics in languages that mainstream tools treat as an afterthought.

Abstract: Evaluation of artificial intelligence concentrates on local performance — accuracy, calibration, robustness — while the upstream normative and volitional architecture that fixes what a system is for remains largely unexamined by those metrics. Existing scholarship has established that bias, proxy variables, problem formulation, measurement, alignment, and governance are value-laden; what remains insufficiently integrated is a single propagation account connecting moral grounding, authorized volition, teleological translation, representation, AI execution, and feedback. This article develops such an account: the Normative–Volitional Architecture of AI-Mediated Action, M → W → T → R → E → I_AI → D → A → C. Moral grounding (M) constrains what may legitimately be pursued; human or institutional volition (W) commits to a direction; teleological specification (T) translates that commitment into objectives and proxies; representation (R) determines which reality is available to the system; and AI performs inference and execution within that structure. The central argument is that AI executes an operational representation of authorized will; it does not independently legitimize ultimate ends. Systemic error therefore arises not only from malformed will but from the normative translation gap between defensible purposes and their operational encodings, and it persists under normative closure — the absence of an institutionalized feedback path through which consequences can reopen objectives, representations, and authority, rather than merely retraining models. Documented cases in healthcare allocation, risk assessment, hiring, engagement optimization, and welfare administration illustrate how locally correct outputs can constitute globally misleading trajectories, and a reflexive governance framework specifies remedies at each architectural layer.

Abstract: This study aims to construct a lightweight action recognition pipeline for work environments that does not rely on detailed manual labeling to achieve both the reduction of manual labeling load in the offline stage and the lightweight recognition of user actions in the online stage. The proposed method generates pseudo-labels from accelerometer data using time-series clustering. Extending the labeling scheme to other subjects, it integrates data from multiple subjects into a common set of action classes. Furthermore, it trains a lightweight action recognition model using the obtained pseudo-labels, which enables us to evaluate the trade-off between model complexity and recognition performance when the method is deployed on edge devices. The process verifies not only the transferability of action structures using pseudo-labels, but also the applicability of lightweight models to sequential action recognition.

Abstract: We provide two separate and important results. First, we formalize a provable separation betweensyntactic and semantic computation for AI learning systems by virtue of the existence of hallucinations.We define run-level “hallucination” under a contract κ= (Lκ,Evalκ,winκ), where truth of an assertedproposition is graded by an external evaluator Evalκ within a fixed decision window. This separationallows us to prove the Transparency Impossibility Theorem: there is no total computable procedurethat, from a single run’s activation log, produces a finite, tape-transparent provenance deciding whetherthe run’s assertion is a hallucination without invoking Evalκ or a Halting/Oracle equivalent. The proofis a halting-encoded diagonal reduction. Rice’s theorem (Rice, 1953) and Tarski’s undefinabilitytheorem (Tarski, 1956) provide independent, complementary impossibility results — Rice rulesout global deciders for nontrivial extensional properties of programs; Tarski rules out an internaltruth predicate for sufficiently expressive Lκ — which we treat as background and supportingmotivation rather than as part of the activation-based reduction itself. Second, in this work we presenta layered account of computation and meaning. The base layer captures effective methods by Turingmachines (Turing, 1937; Church, 1936a). The next layers treat definability, truth, and semantic fixedpoints (Gödel, 1962; Tarski, 1956; Kripke, 1975). Then, we then connect these layers to compression,and description length, which act as practical limits on representation and inference (Li & Vitányi,2008). The aim is clarity about limits. However, we do not enlarge the class of computable sets.Instead, we separate internal effective acceptance from externally grounded acceptance with finitetranscripts. This separation lets us ask when explanation should work, and when it must fail. Rice’stheorem marks undecidable semantic properties that matter for explanation (Rice, 1953). Buildingon this, five corollaries organize the space: completeness, incompleteness, undefinability, groundedacceptance under budgets, and compression ceilings. Each yields a concrete probe or prediction. Theresult is a theoretical framework Church-Turing-Kripke-Meyer (CTKM) to explain how systems canproduce correct but non-derivable behavior without implying hypercomputation. The frameworkalso provides a falsification route insofar as we propose a Diophantine test to refute claims that crossthe classical boundary. Additionally, we offer a conditional description-length lens to mark whenfinite grounding changes acceptance without changing computability. In short, the framework keepscomputability classical while making the role of semantics and resources explicit and testable.

Abstract: Pool-Gated Retrieval (PGR) established that retrieval in LLM-based agents should be governed by an explicit Warrant gate rather than similarity scores, treating Knowledge Gaps as first-class epistemic facts. However, a stable multi-cycle implementation of PGR has remained elusive. Prior internal SCL-PGR runtime implementations relied on pattern matching and domain-specific dictionaries to control retrieve_knowledge behavior, producing a characteristic failure trajectory: each encountered failure mode prompted the addition of a new guard rule, accumulating implicit domain knowledge in the runtime and reducing generality without converging on correctness. This paper diagnoses the root cause of this instability and presents a restructured implementation that resolves it. We show that the class of failures attributed to retrieval quality in prior implementations are not retrieval failures: they are failures of semantic warrant delegation and evidence obligation lifecycle management. Five structurally distinct failure modes are identified and documented through execution traces from a live debugging record: surface-form synonym collapse, raw chunk promotion as committed fact, cross-entity score promotion, compound query premature finalization, and conditional branch over-retrieval. All five share a common origin — pattern-based control attempts to answer semantic questions using lexical features, a task for which it is structurally inadequate. The proposed architecture introduces two mechanisms. First, a Warrant Judge — a single additional LLM call within retrieve_knowledge — resolves all semantic judgment tasks (synonym equivalence, entity alignment, section-heading-to-value warrant) without domain-specific encoding, replacing the entire class of alias tables, location markers, and entity mismatch heuristics. Second, a Turn-Scoped Evidence Ledger manages retrieve_knowledge as an obligation lifecycle rather than a query-response transaction, tracking each knowledge obligation through states of pending, found, gap, and reused_gap. The Ledger enables structural guarantees that pattern-based guards cannot provide: finalization is permitted if and only if all active obligations are in a terminal state, regardless of query surface form, domain, or language. We demonstrate that these two mechanisms together resolve all five identified failure modes across multi-attribute queries, conditional branch queries, comparative queries, and partial-coverage retrieval — without any domain-specific rules. The result is a PGR runtime in which the boundary between code and LLM responsibility is principled: semantic warrant is delegated to the LLM; obligation state management remains in code. We argue that this separation — not the addition of computational resources — constitutes the core engineering contribution and a necessary condition for stable PGR deployment across domains.

Abstract: Target-oriented dialogue systems have demonstrated strong capabilities in completing user goals through interactive conversations. However, existing studies are primarily designed for single, explicit goal completion, while phone call assistants face a proxy setting that requires coordinating the device owner’s explicit preset goal with the caller’s implicit and dynamic goal. We introduce CALLBENCH, a Chinese bench-mark for evaluating dual-goal coordination in phone call assistants. CALLBENCH contains 50,000 complete multi-turn phone call dialogues across six scenarios: takeout, delivery, taxi, work, life, and harassment. It covers regular presets, emergent presets, and no-preset cases, and includes diverse relations between owner-side and caller-side goals, such as alignment, complementarity, irrelevance, and conflict. We further design a preset-aware turn-level evaluation protocol covering semantic understanding, context use, active guidance, response quality, preset compliance, dialogue rhythm, and safety. Experiments on representative dialogue methods show that existing approaches still struggle with this task, highlighting the need for phone call assistants that can make reliable turn-level decisions between two independent goals under proxy constraints.

Abstract: Large language models (LLMs) have become a common center for building intelligent agents, yet many real tasks still need abilities that a single LLM cannot provide. A growing line of work therefore connects LLMs with specialized non-LLM models, including object detectors, segmentation models, diffusion generators, and robot policies. We call this emerging paradigm Heterogeneous Multi-Model Agents (HMMAs). This survey provides a systematic review of 572 HMMA papers published at major AI venues between 2023 and 2026. We first define the scope of HMMAs and distinguish them from adjacent paradigms. We then organize existing systems into five interaction patterns according to the roles played by perception, generation, and action models. Building on this taxonomy, we analyze architectural choices across information flow, interface design, feedback structure, uncertainty handling, and model coupling, and review major application domains. Finally, we summarize open challenges and outline future directions for building reliable HMMAs.

Abstract: Organizations and individuals increasingly delegate consequential action—email, iden-tity, payments, hiring, clinical triage—to autonomous AI agents supplied by many inde-pendent vendors. No vendor holds end-to-end visibility into the delegator's intent, ac-tions, and consequences, so each optimizes its own metrics behind an opaque boundary. The result is coherence debt: the accumulating, largely invisible gap between what a principal expects its delegated actions to produce and what they actually produce. Be-cause AI acts at machine speed, this debt compounds faster than fragmented human oversight can detect it, and its consequences arrive with greater gravity and less warning. This article states the problem precisely, surveys the current state of practice and the emerging architectural response, and sets out what remains to be built and demonstrated. It formalizes coherence debt as unreconciled, severity-weighted divergence, and shows—via a requisite-variety argument—that governance bolted on externally cannot close the gap in principle. It then distinguishes two loci of the problem: intra-system co-herence debt, which the emerging Mindful Machines paradigm addresses by making governance an intrinsic architectural property (a Digital Genome, an autopoietic control system, and a continuous Discover–Reflect–Apply–Share loop), and boundary coherence debt, which no current approach closes. It proposes the Sovereignty Boundary Ledger, a six-layer reference architecture that extends intrinsic-governance discipline across the trust boundary to agents the principal cannot rewrite, and it specifies concretely what is already demonstrated and what is required to demonstrate the approach at enterprise scale.

Abstract: Adversarial machine learning is an important area of research in computer science, focusing on understanding and mitigating attacks that make use of the vulnerabilities of machine learning models. In such attacks, adversaries aim to exploit these vulnerabilities in order to harm model's utility or violate its privacy or availability. These attacks include evasion, poisoning, exploratory, and explainability. Evasion, poisoning, and explainability attacks aim to harm the models' utility, while exploratory attacks violate the privacy of the models. To reduce the negative impacts of these attacks, a plethora of defense mechanisms have been proposed. In this survey, we review a substantial body of works and propose a comprehensive and novel taxonomy of defense strategies. We divide defense systems into eight main categories including: training-based, architecture-based, uncertainty-based, detection-based, optimization-based, transformation-based, information-theoretic, and hybrid approaches. For each category, we analyze and summarize key techniques and algorithms. Moreover, we investigate the limitations associated with these defenses, such as model generalization, robustness-accuracy trade-off, and adaptability challenges. By highlighting the strengths and weaknesses of existing defenses, this survey aims to enlighten future research towards more robust and efficient adversarial defense mechanisms.