https://jst.vn/index.php/dsai Journal of Science and Technology - Technical Universities Hanoi University of Science and Technology en-US Data Science and Artificial Intelligence 3093-3331 https://jst.vn/index.php/dsai/article/view/1545 Weight-only post-training quantization (PTQ) pipelines such as AWQ ultimately commit to a deterministic rounding decision for every scalar weight, and recent discrete coordinate-descent quantizers refine those decisions one weight at a time. We show that both are blocked by the same obstacle: correlated activation channels create multi-spin energy barriers in the discrete rounding landscape — configurations in which no single rounding flip helps, yet a small group of coordinated flips strictly reduces the layer reconstruction error. We characterize these barriers exactly under the empirical activation metric and prove that they arise only from frustration between local quantization fields and cross-channel couplings. Building on this analysis we propose Transverse-Field Ising Correction (TFIC), a barrier-aware discrete optimizer that post-processes any AWQ-style pipeline. TFIC casts fixed-grid rounding as a classical Ising ground-state problem, selects the spins near a rounding boundary as a mobile pool, drives that pool to a single-flip local minimum by exact coordinate descent, and then crosses barriers with cluster moves that flip small groups of coupled spins together. Every accepted move is scored against the exact reconstruction energy, and a robust per-layer noise-floor acceptance rule prunes flips whose apparent gain is within calibration-estimation noise; together these guarantee that each layer’s calibration reconstruction error is monotonically non-increasing. TFIC requires no gradients and no architectural changes. On Mistral-7B and Llama-3-8B, applying TFIC on top of a 4-bit AWQ baseline consistently reduces WikiText-2 and C4 perplexity, at a per-layer cost of tens of seconds and with no inference-time overhead. Dr Anh Nguyen Duc Linh Ngo Van Sang Dinh Viet Hoang Ta Duy Tung Nguyen Son Canh Dinh The Tu Vu Van Copyright (c) 2026 Data Science and Artificial Intelligence https://creativecommons.org/licenses/by/4.0/ 36 1 001 008 10.51316/jst.194.dsai.2026.36.1.1 https://jst.vn/index.php/dsai/article/view/1574 This study develops a forward-cross-fitted two-channel EWMA (FCF-DS-EWMA) chart for post-deployment monitoring of predictionerror deterioration in continuous-response machine-learning models. Temporally held-out Phase-I residuals are transformed into robust signed and magnitude channels, and an equal-share rule uses channel-specific finite-horizon limits whose common marginal exceedance level is selected jointly. In a unified Monte Carlo study with 𝑅 = 1000, the native false-alarm probability (PFA) is 0.176–0.196 for the in-sample-residual chart and 0.094–0.097 for FCF-DS across the three in-control environments. Six deterioration mechanisms are compared on a common ΔMSE scale. The signed and magnitude channels specialize to different changes, while equal-share calibration restores directional sensitivity relative to a common-limit maximum (abrupt-shift detection probability 𝑃𝐷 = 0.724 versus 0.582). One-factor analyses of native PFA identify serial dependence, inadequate block length, and monitoring horizon as the clearest sources of false-alarm inflation. An oracle diagnostic shows that the temporary-model/final-model reference mismatch mainly shifts the operating point by a small amount. The ethanol study illustrates channel attribution and reference-location sensitivity, and paired perturbations of actual post-Phase-I Metro traffic residual windows provide quantitative evidence in a second regression domain. MSc Quynh Pham Ba Ph. D Du Nguyen Huu Ph. D. Son Ta Anh MSc Hoa Le Thanh Copyright (c) 2026 Data Science and Artificial Intelligence https://creativecommons.org/licenses/by/4.0/ 36 1 009 017 10.51316/jst.194.dsai.2026.36.1.2 https://jst.vn/index.php/dsai/article/view/1572 Driver action recognition (DAR) is a core capability of advanced driver assistance systems, however, it remains difficult because distracted driver actions involve small interacting objects, subtle hand movements, and severe self-occlusion inside a confined cabin. Most existing methods are trained on pre-segmented video clips with predefined action labels requiring substantial annotation effort and additional labeled data when extending to new action classes, which limits their scalability to real-world applications. Moreover, they cannot localize actions in an untrimmed video stream. This paper presents HybridLVLM, a training-free framework for continuous driver action recognition in untrimmed videos that leverages large pretrained vision-language models. Firstly, the framework transforms every sampled frame into a natural-language caption, filters the captions with a text-only language model to extract distracted action evidence and core frames for video representation. Secondly, an open-vocabulary detector then spatially localizes the bounding boxes corresponding to the driver and several task-relevant objects on each core frame. Then, these bounding boxes together with the set of core frames are fed forward to a two-stage Chain-of-Thought (CoT) reasoning model to infer an action class for each core frame based on task-relevant visual evidences. Finally, a temporal post-processing is applied to refine and extend frame-level recognition to continuous recognition presented by a predicted action tube. Experiments on the UTCDriverAct dataset show that HybridLVLM attains a Video-mAP of 52.0% and a mean F1-score of 61.4% at an STT-IoU threshold of 0.3 without any training on driving data. These findings point to a promising research direction in video understanding for applications without real-time constraints. Additionally, an ablation study on the input sampling rate is also conducted to investigate the effect of the sampling rate on the overall recognition performance. The results indicate that the optimal sampling density depends on action duration, with higher rates for rapid actions and lower rates for gradual actions Mr. Hiep Bui Hoang Dr. Binh Nguyen Thuy Linh Vo Thien Dr. Quan Nguyen Hong Dr. Frédéric Rayar Assoc. Prof. Lan Le Thi Copyright (c) 2026 Data Science and Artificial Intelligence https://creativecommons.org/licenses/by/4.0/ 2026-10-15 2026-10-15 36 1 018 028 10.51316/jst.194.dsai.2026.36.1.3 https://jst.vn/index.php/dsai/article/view/1553 Holding a stable viewpoint is a prerequisite for observation tasks performed by remotely operated underwater vehicles, yet it is difficult to achieve when the only sensor is a monocular camera and all computation must run on board. This work reports the deployment of a complete visual stationkeeping perception pipeline on a Qualcomm QCS8550 platform using its Hexagon HTP/NPU accelerator, in which two networks---an INT8 object detector and a dense monocular depth model---share a single accelerator. A per-stage latency breakdown shows that the throughput difference between an accelerated and a CPU-executed configuration is attributable to one stage: moving depth inference alone to the accelerator makes it 27 times faster and carries the pipeline from 0.46 to 8.47 frames per second, with depth inference still occupying two thirds of the frame budget afterwards and the accelerator occupied at least 69\% of wall-clock time. We document two deployment failure modes that are silent rather than loud: an output polarity that must be established empirically, and a normalisation choice that destroys the range signal. The two available range cues are combined by an analytic fusion rule---the square root of the tracked bounding box area and the depth output both vary inversely with distance, so the two are directly combinable once anchored on a reference captured at target selection, without training data or knowledge of the target dimensions. On the rigid targets evaluated here the zero-cost geometric term carries more of the distance signal than the depth network that consumes two thirds of the frame budget, and the two cues degrade unequally on immersion, which makes their relative weighting an operating parameter rather than a constant. Evaluation is open-loop; closed-loop validation was prevented by a degenerate thruster configuration. Mr Gyan Louisin Assoc. Prof. Vinh La The Dr. Trung Ngo Lam Dr. Nga Nguyen Thi Thanh Dr. Dung Nguyen Anh Assoc. Prof. Duc Duong Minh Dr. Hiep Hoang Van Copyright (c) 2026 Data Science and Artificial Intelligence https://creativecommons.org/licenses/by/4.0/ 36 1 029 039 10.51316/jst.194.dsai.2026.36.1.4 https://jst.vn/index.php/dsai/article/view/1638 Face presentation attack detection (PAD) systems intended for real-world access control must provide not only accurate spoof detection but also reliable on-device operation, multi-sensor integration, traceable decision making, and reproducible evaluation. This paperpresentsacompleteon-deviceRGB–IRfacePADsystemimplementedonacustomAndroidaccess-control device and evaluated independently under ISO/IEC 30107-3-based test protocols. The system employs temporally paired RGB and infrared observations and integrates three complementary sources of evidence: a primary RGB liveness score, an RGB–IR cross-spectral consistency score, and a secondary independent liveness score. Frame-level outputs are combined through a configurable transaction-level decision policy based on multi-frame aggregation, while intermediate scores, decisions, and transaction results are logged to support traceability and post-hoc analysis. The complete hardware–software system was independently evaluated by BixeLab, an ISO/IEC 17025-accredited biometric testing laboratory, at Level A and Level B. For each level, the evaluation involved 10 consenting subjects, 50 bona fide transactions, and 300 presentation attacks covering six presentation attack instrument species. No attack classification errors or attack non-responses were observed in either evaluation, corresponding to an observed APCER of 0/300 and APNRR of 0/300 at both levels. However, 10 of 50 bona fide presentations were incorrectly rejected at each level, corresponding to an observed BPCER of 20%. These results demonstrate the effectiveness of the proposed system-level integration against the evaluated attack species while also highlighting the security–usability trade-off introduced by conservative operating points. The study provides an independently validated example of a deployable and auditable RGB–IR PAD architecture for real-world biometric access-control systems. Mr. Huynh Duong Huu Dr. Hung Pham Ngoc Assoc. Prof. Vinh La The Dr. Thuan Nguyen Dinh Dr. Van Hiep Hoang Copyright (c) 2026 Data Science and Artificial Intelligence https://creativecommons.org/licenses/by/4.0/ 36 1 040 048 10.51316/jst.194.dsai.2026.36.1.5 https://jst.vn/index.php/dsai/article/view/1577 With the recent improvements in deep learning techniques, many polyp segmentation methods have become increasingly accurate, facilitating prompt diagnosis and treatment of colorectal cancer. However, the scarcity of labeled datasets poses significant challenges for training and developing supervised learning methods, as annotating data is a time-consuming and labor-intensive task. To address this problem, we propose CTC-MT, a semi-supervised framework for colorectal polyp segmentation based on Mean Teacher and Cross Pseudo Supervision, comprising a teacher and two students to utilize both limited labeled data and unlabeled data. With the proposed framework, we aim to encourage the similarity between the predictions from each student and the pseudo-label from the teacher for unlabeled input, and ensure the consistency between the predictions of the two students for the same input image. Additionally, we improve PMFNet architecture as the backbone model by leveraging a spatial MoE block and adding a regression branch to learn boundary features through a signed distance map and cross-task consistency. The experimental results show that our framework achieves promising performance on Kvasir-SEG, ETIS, CVC-300, CVC-ClinicDB, and CVC-ColonDB datasets compared to other state-of-the-art approaches. Ms. Thao Tran Thi Phuong Mr. Lam Tran Tung Mr. Tu Nguyen Ba Assoc. Prof. Thao Tran Thi Assoc. Prof. Truong Pham Van Copyright (c) 2026 Data Science and Artificial Intelligence https://creativecommons.org/licenses/by/4.0/ 36 1 049 059 10.51316/jst.194.dsai.2026.36.1.6 https://jst.vn/index.php/dsai/article/view/1548 Serving a Markowitz efficient frontier at a varying risk-aversion level requires one quadratic program per query. We study a learned alternative: a hypernetwork that maps a normalized preference vector, read as a risk-aversion level, to a long-only portfolio in one forward pass. A softmax output layer enforces the budget and no-short-selling constraints, and training minimizes a Chebyshev scalarization without any pre-solved frontier. The architecture, the scalarizations and the guarantees for the scalarized problem are those of controllable Pareto front learning [1, 2]. Our contribution is an exact-oracle evaluation: for the long-only mean–variance problem the critical-line algorithm yields the exact Chebyshev solution at every preference, so the error of the learned map can be measured exactly, compared with the strongest alternatives and located. On a ten-industry synthetic instance and a fifteen-asset VN30 instance the map attains a mean Euclidean weight error of 0.0083 and 0.0088 over ten seeds and answers a query in 44 microseconds on one CPU core. From five to two hundred assets the error grows from 0.0041 to 0.0734; its largest values occur where the highest expected returns are nearly tied, it concentrates at changes of the active set (median per-seed ratio 2.45 on VN30), and supervised training on exact targets reduces it four- to six-fold. The exact solver computes the entire frontier in 0.22 to 3.35 ms, so on this problem the learned map offers no computational advantage. Khang Truong Tuan Dr. Thang Tran Ngoc Dr. Anh Dao Tuan Copyright (c) 2026 Data Science and Artificial Intelligence https://creativecommons.org/licenses/by-nc/4.0/ 2026-10-15 2026-10-15 36 1 060 071 10.51316/jst.194.dsai.2026.36.1.7 https://jst.vn/index.php/dsai/article/view/1565 Parallel fuzzing can exchange successful inputs, but the behavior of different Python mutation strategies under a shared corpus is not well characterized. We evaluate four Pythonfuzz-anchored two-worker configurations-PP, PD, PO, and PG-on five Python libraries and compare them with isolated single-process references under one common sys.settrace-based line-transition coverage representation. The observed outcomes are target-dependent and do not favor GRPO as a cooperative partner. On python-dateutil, PD and PO each reach a global union of 3407 edges, whereas PG reaches 2913 despite similar local coverage of the learning-guided worker. On cryptography, PD and PO reach a 47-edge union while PG shows no union gain. A provenance-enabled PG run records 106 global-first discoveries whose direct parent was produced by the peer worker, documenting within-run seed reuse without establishing a causal benefit of sharing. Each configuration is represented by one six-hour campaign; the reported values are therefore individual campaign observations rather than statistical comparisons. Two previously undocumented exceptions are also reported as qualitative bug-finding evidence. Mr. Anh Le Minh Viet Mr. Hieu Nguyen Huy Dr. Thang Hoang Nam Dr. Anh Tran Hai Asst. Prof. Truong Tran Xuan Copyright (c) 2026 Data Science and Artificial Intelligence https://creativecommons.org/licenses/by/4.0/ 2026-10-15 2026-10-15 36 1 072 079 10.51316/jst.194.dsai.2026.36.1.8