TechRxiv Daily Picks

[TechRxiv] A TechRxiv preprint by Ravi Natarajan presents a pattern-based architectural approach for governed agentic systems, called K9-AIF. According to the paper, agentic systems increasingly combine multiple cooperating agents, heterogeneous models and external tools, creating new challenges in governance, traceability and architectural extensibility at enterprise scale, while existing orchestration frameworks often emphasize task coordination and offer limited support for consistent policy enforcement, observability and long-term maintainability. The approach is grounded in a clear separation between Architecture Building Blocks (ABBs) and Solution Building Blocks (SBBs), and uses configuration-driven factories, runtime agent loaders and layered architectural abstractions to structure orchestration, inference integration, persistence and monitoring in a modular, governance-first manner. Diagrams illustrate how these elements compose into a cohesive architecture that supports extensibility while preserving stable architectural contracts. The work is a preprint and has not been peer reviewed.

[TechRxiv] A preprint by Jessica Sciammarelli, dated March 5, 2026, proposes a decentralized AI infrastructure model that integrates blockchain into machine learning operations (MLOps) workflows to enhance data governance, model traceability and lifecycle transparency. According to the article, modern AI systems rely on complex infrastructure stacks spanning data pipelines, MLOps and cloud services, and centralized control of these components introduces risks related to data manipulation, lack of traceability and governance failures. The model uses smart contracts to enforce governance policies, control access to datasets and register model artifacts throughout their lifecycle. The article says the proposed architecture shows how blockchain can complement existing AI infrastructure by providing decentralized coordination, auditability and policy enforcement without compromising performance or scalability. The work is a preprint and has not been peer-reviewed.

[TechRxiv] Rohith Namboothiri proposes Authorization-First Retrieval (AFR), an architectural invariant requiring that authorization constrain the retrieval candidate set before any learned component consumes retrieved content, according to a preprint. In a controlled corpus of 43 chunks across 28 documents with role-scoped rules spanning five enterprise roles, three sensitivity levels and four domains, the author reports that retrieve-then-filter pipelines exposed unauthorized context in 96.8% of 93 structured queries covering six attack categories and violated authorization correctness on 100% of queries that should be denied, while AFR eliminated all structural leaks (0/93) and all answer leaks (0/93). A chunk-size sweep across five granularities confirmed AFR held invariant to chunking strategy, while retrieve-then-filter leakage rose monotonically with chunk size, the author said. The author identifies parametric knowledge leakage as a fundamental boundary on what retrieval architecture can guarantee. The work is a preprint and has not been peer reviewed.

[TechRxiv] A TechRxiv preprint by Ramanayake and co-authors presents a machine learning-based framework for predicting vehicle arrival times when origin and destination are not fixed. The study developed and evaluated Linear Regression, Artificial Neural Networks, Long Short-Term Memory networks, Decision Trees and Random Forests, using data collected at fixed intervals from public transport routes in Sri Lanka, with Mean Absolute Error (MAE) used to assess performance. According to the authors, Linear Regression showed the highest error and Random Forest the lowest, and model effectiveness was directly influenced by both the data collection methodology and the choice of features. The authors said that with further improvements in dataset quality and model design, the framework has strong potential for practical deployment in applications where accurate arrival time prediction is critical, such as public transportation management and taxi services. The paper is a preprint and has not been peer-reviewed.

[TechRxiv] A TechRxiv preprint by Morgan Mekale Smith and co-authors presents VFS-Net, a dual vertex- and face-based deep learning pipeline for segmenting and analyzing dynamic three-dimensional models. According to the authors, the framework showed favorable segmentation and classification metrics on both biological and synthetic 3D model types, indicating broad applicability, and benchmarking across diverse datasets indicated that VFS-Net remained robust to variations in model complexity and mesh density. The authors said traditional machine learning methods require structured grids or volumetric data and are not optimized for data points that vary widely by scale and surface area. The pipeline was designed to operate on 3D models obtained through structured-light scanning, enabling accurate capture of complex surface geometries and topological features. The authors said future extensions will integrate two-dimensional data to broaden the framework's biomedical research relevance.

[TechRxiv] Louis Nisiotis and Nikita Markov said in a TechRxiv preprint that they built a RAG-based back-end AI system to support virtual agent conversational capabilities in a virtual-world prototype and tested two small language model families at three quantization levels (8-bit, 4-bit and 3-bit), evaluating system performance through end-to-end latency, time-to-first-token, throughput and memory headroom. They also assessed response quality through expert evaluations of accuracy, relevance, safety and human alignment against non-quantized baselines. According to the authors, the results show that moderate quantization enables performant practical edge deployment with good response quality, and that edge-deployed quantized SLMs can support real-time narrative interactions and remain practically useful for virtual-world agents. The paper is a preprint and has not been peer-reviewed.

[TechRxiv] Silvirianti, Georges Kaddoum, Bassant Selim and Mahdi Chehimi propose in a preprint a quantum adversarial deep reinforcement learning (QADRL) approach to mitigate AI-based jamming attacks and improve quality-of-service (QoS) in low Earth orbit (LEO) satellite networks. According to the paper, QADRL is based on a zero-sum Markov game with two opposing learning networks: one optimizes satellite routing links to avoid jamming and improve QoS, while the other targets the jammer, optimizing unmanned aerial vehicle (UAV) trajectory, jamming nodes and power to maximize jamming success. The paper reports that QADRL outperforms classical adversarial DRL (CADRL) by reducing the jamming success rate by 33.33% and increasing average satellite network QoS by 18.4975%. The findings are from a preprint and have not been peer-reviewed.

[TechRxiv] A TechRxiv preprint by Rohith Namboothiri proposes a measurement construct called the dual-run comprehension delta, which reframes teaching efficacy as the measurable change in a standardized evaluator's comprehension before and after receiving a teacher's explanation. According to the paper, existing approaches rely on classroom observation protocols requiring trained human raters or on student evaluations that correlate poorly with actual learning outcomes, and both suffer from subjectivity, scalability constraints and documented biases. The author argues that large language models can serve as controlled learner proxies and develops a formal framework in which a domain-specific assessment is administered to an LLM twice, once without access to the teacher's explanation and once with it, with the score difference constituting the delta. The paper grounds the construct in validity theory from the Standards for Educational and Psychological Testing, analyzes it through pedagogical content knowledge and explanation quality research, and presents a reference architecture instantiating the methodology. The author says the dual-run delta offers a reproducible and scalable complement to existing teaching quality measures, particularly for asynchronous instruction, teacher certification and professional development, and outlines a validity argument around Kane's interpretation/use framework, discusses construct boundaries and identifies empirical studies needed to evaluate its claims. The paper is a preprint and has not been peer reviewed.

[TechRxiv] A TechRxiv preprint survey by Kai Jiang and co-authors introduces a philosophical taxonomy that reframes bias in large language models, organizing existing work around three core areas: bias evaluation, bias mitigation, and evaluation of mitigation techniques. The survey distinguishes substantial harms — tangible, adverse impacts on individuals or groups — from representational harms, described as demeaning or stereotyping portrayals that violate dignity, and it highlights how these harms often interact and reinforce one another. According to the authors, structuring metrics, datasets and methods within this taxonomy provides a foundation for developing fairer and more inclusive language models. The paper is a preprint and has not been peer reviewed.

[TechRxiv] A TechRxiv preprint by Zongyuan Zhang and co-authors proposes a Transformer-based Congestion Control Optimization (TCCO) framework for multipath transport. TCCO uses a decoupled architecture that offloads control decisions to an external decision engine via a lightweight in-kernel client and a user-space proxy, allowing edge devices to draw on external compute while maintaining TCP/IP compatibility. The Transformer-based deep reinforcement learning agent in the external engine uses self-attention to capture temporal dependencies, filter measurement noise and coordinate subflow control under a unified policy. According to the authors, evaluations on simulated and real dual-band Wi-Fi testbeds show TCCO achieves better adaptability and performance than state-of-the-art baselines. The preprint has not been peer-reviewed.

[TechRxiv] A TechRxiv preprint by Seung Heon Oh, Geon Woong Byeon, Young In Cho, Seungmin Kwon and Jong Hun Woo proposes solving the Weapon-Target Assignment (WTA) problem using reinforcement learning combined with graph neural networks. According to the paper, earlier approaches based on dynamic programming, linear programming, metaheuristics and heuristics were limited by oversimplified models, computational burden, lack of adaptability to disruptive events and the need for recalculation when problem size changes. The authors said the proposed method reflects the real-world OODA-loop (Observe-Orient-Decide) decision framework and was tested across various environments, with effectiveness shown by comparison against existing heuristic and metaheuristic methods. The authors described it as a new methodology for intelligent decision-making in tactical command and control, traditionally considered exclusive to human experts. The work is a preprint and has not been peer-reviewed.

[TechRxiv] Zhe Li, Detong Wang, Sheng Fang and co-authors present in a preprint a Modal-invariant Feature Extraction Network (MFE-Net) for unsupervised multimodal change detection (UMCD). According to the authors, aligning cross-modal images in a common subspace via deep learning without decoupling residual modal noise from intermediate features leads to poor alignment and discrimination errors when features are compared directly. MFE-Net learns land use and land cover semantics through a Hybrid CNN-KAN (HCK) encoder, which combines CNNs' local perception with Kolmogorov-Arnold network global nonlinear modeling, and a Modal Difference Learning Module (MDLM) that uses adversarial multi-task learning to separate modal-invariant components from mode-specific noise. A Differential Feature Enhancement Module (DFEM) computes change intensity by fusing feature-level differences via Spearman's rank correlation. The authors say experiments on six real datasets show MFE-Net outperforms state-of-the-art methods. The work is a preprint and has not been peer-reviewed.

[TechRxiv] A survey posted to TechRxiv on March 6, 2026 positions DouDiZhu as a representative testbed for imperfect-information card games (IICGs) and provides a review and categorization of existing AI methods. The authors, Qian Luo, Daochen Zha, Tianqiao Zhang and Xuewen Tan, note that imperfect-information games pose decision-making challenges due to hidden information and multi-agent interaction, and that IICGs form a strategically rich subclass that most existing reviews of imperfect-information games have given little attention. According to the survey, DouDiZhu has become an emerging AI research topic because of its asymmetric three-player roles, large state-action space and complex strategic interactions; recent work spans search-based methods such as MCTS, deep reinforcement learning methods such as DQN, DMC and PPO, and hybrid approaches integrating learning, search and opponent modeling. The authors said the literature remains fragmented and lacks a systematic synthesis, and the survey analyzes strengths and limitations, highlights trends and compares performance against available baselines. It also identifies open challenges and future directions, including generalization beyond standard rules, unified bidding and playing, continual learning and standardized evaluation. The paper is a preprint and has not been peer reviewed.

[TechRxiv] Sanjay Kumar Mishra has posted a preprint presenting an independent empirical evaluation of Oracle Autonomous Database 26AI's SELECT AI feature, which uses OpenAI GPT-4 to convert natural language into executable SQL. According to the paper, the author used a custom evaluation framework and 22 TPC-H benchmark queries spanning simple, medium and complex difficulty tiers to measure SQL generation accuracy and end-to-end latency. The results showed overall accuracy of 95.5%, including 90.9% exact SQL match, while mean end-to-end latency was 3,906ms, of which 94.5% was attributable to LLM generation overhead. The paper said the model demonstrated intelligent query optimization, selecting targeted columns over broad COUNT(*) patterns, a behavior not reflected in exact-match metrics. The author said the findings establish a reproducible baseline for enterprise NL2SQL deployment and identify LLM generation as the primary optimization target.

[TechRxiv] Qi Lyu, Jiahua Dong and co-authors have released a survey on world models in artificial intelligence on TechRxiv. According to the survey, world models offer a unified approach to modeling how environments evolve over time by learning predictive representations of states and observations; advances in large-scale generative modeling and multimodal foundation models have broadened their applicability across interactive and multimodal domains, yet existing research remains fragmented across modeling paradigms, application domains and evaluation protocols. The authors categorize world models into four branches with formal mathematical formulations: observation-level generative, latent space, reinforcement learning-based and object-centric world models. The survey reviews applications spanning robotics, autonomous driving, scientific discovery, game simulation, GUI-based agents, and interpretability and trustworthiness, and analyzes benchmarks, new evaluation metrics, simulation platforms and comparative results across world models. The authors said long-horizon consistency and generalization remain key challenges and outline directions for future research; the survey maintains an actively updated GitHub repository to track developments and aims to offer a unified reference for understanding, comparing and advancing world models.

[TechRxiv] LXB, a cross-device mobile automation framework described in a preprint by Runguo Chen, uses a hybrid neuro-symbolic architecture to address high latency, excessive API costs and poor reliability of vision-language model (VLM)-driven mobile automation on complex multi-page tasks. According to the paper, LXB introduces VLM-XML Fusion, which aligns semantic detections from VLMs with structural XML accessibility nodes through point-containment matching; a Route-Then-Act paradigm that separates deterministic graph-based navigation from AI-guided execution; and Node-Driven Exploration, which builds navigation graphs through breadth-first traversal of discovered navigation nodes. The paper reports an overall success rate of 100% on 8 real-world tasks across Taobao and Bilibili. On deep tasks requiring at least 2 navigation hops, LXB is said to achieve 100% success with only 1.0 model call, a single LLM query during the planning phase, versus 80.0% with 3.87 model calls for VLM-ReAct, a 74% reduction in model invocations and 58.6% reduction in token consumption. The findings are from a preprint and have not been peer-reviewed.

[TechRxiv] Yahia Hamdi, Mohamed Fnadi, Wenqian Du and others proposed a Swin Transformer-based deep learning framework for detecting and classifying floating debris in water surface debris accumulation areas such as rivers, coastlines and ports. The study performed Bayesian optimization and fine-tuning on Swin-T and compared it with ResNet-50, InceptionV3, VGG-19 and ViT; the dataset used was expanded from public data and newly collected RGB images under multiple hydrological and lighting conditions. The paper claims that the optimized Swin-T achieves a classification accuracy of 98.5% and a detection mAP of 52.3%, outperforming baseline Transformers and conventional CNNs. The authors state that systematic hyperparameter optimization can significantly improve the performance of Transformer models for marine debris monitoring.

[TechRxiv] Shingo Kitayama published a study proposing a tool called sqlew that repurposes architecture decision records (ADRs) — introduced in 2011 but shelved due to the manual burden of maintaining them — as external design memory for LLMs, enabling automatic ADR recording and structured-query selective retrieval via the Model Context Protocol (MCP), and establishing a pattern in which static instructions (CLAUDE.md) and dynamic rationale (ADRs) coexist. The author conducted a three-condition controlled experiment on the development of a 12.5-round TypeScript workflow approval system: no ADR, sqlew adopted from the start, and sqlew adopted retroactively (v2s). The paper reports that the sqlew condition reduced development time by 10.1% (496 minutes versus 552 minutes) and reduced total tool calls by 10.8%; after accounting for the additional round required in the no-ADR condition to add E2E tests (round 12.6, 29 minutes, 21,751K tokens), the time difference widens to 14.6%. Chain-of-thought keyword density analysis shows the reasoning pattern shifting from trial-and-error loops to context-based reasoning, with the density of references to prior decisions diverging at an accelerating rate (+21%→+64%→+137%). The author states that the most prominent finding…

[TechRxiv] Lihui Liu released a review on March 5 focusing on an advanced form of retrieval-augmented generation (RAG), "deep RAG," which integrates reasoning, memory, and multi-step planning into an end-to-end research workflow. The paper states that the review systematically organizes the key components of deep RAG systems, including query planning, information retrieval, memory management, and answer generation, and covers optimization techniques such as supervised fine-tuning, iterative self-evolution, and agentic reinforcement learning. The author also analyzes different planning strategies, adaptive retrieval mechanisms, conflict resolution methods, and multimodal synthesis, points out their advantages and limitations, classifies methods by workflow type, reasoning strategy, and optimization paradigm, outlines the current landscape, and raises open challenges in evaluation, scalability, and multimodal reasoning. The review states that its goal is to provide a roadmap for designing flexible, reliable, and intelligent deep RAG systems that can complete complex real-world research tasks with minimal human supervision.

[TechRxiv] Tong Guo proposes a unified framework called ReLabel for identifying noisy data in deep learning applications and relabeling it, using model predictions as a reference during relabeling. The paper states that initial datasets come from manual annotation, large language model generation, or user behavior logs, all of which may contain a certain amount of noisy data; the relabeling process can be done manually or with the help of large language models for annotation. The framework covers multiple types of tasks, including classification, sequence labeling, object detection, sequence generation, and click-through rate prediction. The author states that evaluation results on the development set and manual evaluation results validate this approach, and mentions that the goal is to achieve a score above 90 on the development set.

[TechRxiv] Laha Ale and others propose a framework that combines agentic workflows with retrieval-augmented generation (RAG), deployed in 6G network edge environments, to reduce hallucinations in generative AI. The paper states that although RAG can improve factual grounding with external knowledge beyond the model's training data, RAG alone cannot fully eliminate hallucinations; agentic workflows that introduce external tools such as APIs, search engines, and self-reflection mechanisms can enable models to iteratively evaluate and correct outputs. The framework consists of a multi-agent system composed of generator and critic agents, reducing errors through iterative self-criticism, and uses 6G edge intelligence for data fusion, dynamic updating of knowledge bases, and delivery of customized AI services. The authors state that this helps provide more reliable generative AI services across a variety of applications.

[TechRxiv] Lihui Liu released a review on March 5, 2026, surveying the concepts, methods, and recent progress in integrating large language models with graph learning, with a focus on text-attributed graphs. The paper states that before the rise of LLMs, graph neural networks were the mainstream approach in graph learning, with node embedding, graph convolution, graph attention, and random walk-based techniques widely adopted; when handling graphs with text node attributes, shallow text embeddings are typically used as initial node representations, but the general knowledge captured by such embeddings is limited. The review examines two main routes around tasks such as node classification, link prediction, and graph-based question answering: one is "LLM as enhancer," in which the LLM enriches node text attributes with its knowledge and graph learning methods then make predictions; the other is "LLM as predictor," in which graph data is fed as input to the LLM, which directly solves graph-related tasks. The author states that through a systematic examination of these two directions, observations and new insights on the integration of LLMs and graph learning are provided.

[TechRxiv] Manish Tripathy released a preprint on March 5 proposing a cross-domain threat actor attribution framework that bypasses the "LLM wall." The paper states that existing attribution methods based on semantic features such as BERT and RoBERTa approach random accuracy under adversarial style transfer. The framework instead uses behavioral signal processing: Logical Time Difference of Arrival (L-TDoA) draws on the mathematics of electronic warfare signal propagation to measure "cognitive delay," while SynGNN learns identity based on deep syntactic embeddings such as dependency tree structure and part-of-speech transitions rather than lexical content. The author reports that under fully structurally stripped conditions, SynGNN recovers 78.6% (95% CI: 75.4-81.8%) of identities, while BERT-type models drop to 12.4%; in a controlled adversarial scenario with 200 candidate suspects, the fused simulation achieves an overall precision of 91.7% (95% CI: 88.3-94.1%). The study is positioned as a TRL 3 foundational framework.

[TechRxiv] Wei Chen, Jiaming Zhang, Haowen Liu and others have published a review that traces the evolution of industrial intelligent fault diagnosis from traditional data-driven methods and deep learning to emerging large language models and foundation models. The paper states that over the past decade, convolutional neural networks, recurrent architectures and generative adversarial networks have significantly improved diagnostic accuracy for rotating machinery, power systems and manufacturing processes; large language models, meanwhile, offer a complement to or surpass these methods in natural language reasoning, multimodal fusion and few-shot generalization. The authors establish a taxonomy organized by signal-level feature extraction, supervised and self-supervised deep architectures, transfer and meta-learning strategies, and LLM pipelines combining prompting, retrieval-augmented generation and fine-tuning, and compare representative methods on standard benchmarks. The paper identifies open challenges including domain-specific pretraining, real-time deployment, multimodal alignment and interpretability, and sets out research directions.