Characterizing Multi-Energy X-ray Transmission (XRT) hyperspectral images presents significant challenges for traditional classification frameworks due to high-dimensional redundancy and non-linear spectral overlaps. Although manifold learning techniques such as UMAP enhance spectral representation by preserving non-linear correlations, they do not explicitly utilize the geometric structure of the embedded manifold for segmentation. This study introduces a hybrid architecture that integrates topological features derived from persistent homology with convolutional and graph neural networks. The approach computes Vietoris-Rips persistent homology on UMAP-projected superpixel point clouds to capture H 1 cyclic structures that encode spectral heterogeneity. These topological descriptors are combined with pixel-level CNN features through a broadcasting mechanism that propagates superpixel-level graph representations to full spatial resolution. Experimental results on XRT multi-energy imagery demonstrate that the proposed hybrid model significantly outperforms both CNN-only and superpixel-pooling fusion baselines, suggesting that topological features provide complementary discriminative information to local spectralspatial descriptors.
5G networks increasingly resemble distributed systems in their design. It then becomes crucial to move beyond traditional network-based detection methods and to explore alternative approaches. In this work, we propose a formalization of 5G network exchanges as a dynamic graph, on which we apply techniques from the field of graph-based learning to detect anomalies effectively. This representation allows for fine-grained analysis of packet-level semantic content, while simultaneously capturing sequential and topological contexts, both of which are essential for identifying sophisticated and multi-stage attacks. To assess the effectiveness of this formalism, we evaluate it with the most popular dynamic graph models, demonstrating its potential for 5G anomaly detection.
Online algorithms with bounded recourse face three intertwined obstacles: uncertainty about the future, computational limits, and a limited budget for changes. We ask a fundamental question that separates consistency from the other two: what does solution stability cost when the entire future is known and computation is unlimited? We answer this question for monotone submodular maximization, where an algorithm must maintain a solution of at most $k$ arrived elements, with a constant number of changes per step. We determine the optimal approximation thresholds: $3/4$ for deterministic algorithms and $\xi^\star \approx 0.7942$ for randomized algorithms. For every $\varepsilon>0$, we get within $\varepsilon$ of these thresholds with worst-case recourse polynomial in $1/\varepsilon$, independent of $k$ and the stream length. We prove matching impossibility results even for coverage functions. We attain the sharp bounds by optimally balancing two opposing forces: preparing for future arrivals and protecting the current solution. For randomized algorithms, we use addition robustness against prefixes of the known future, which we analyze tightly via an interval-selection problem. For deterministic algorithms, we introduce certificates, paths in a fractional relaxation that maintain high value, which we round into an update schedule.
Deep learning can classify crop diseases accurately from benchmark leaf images, but centralized training requires farms to share raw data, and predictions without explanations offer limited support for treatment decisions. We present FedXCrop, a federated crop disease classification pipeline that quantitatively evaluates explanations of the aggregated global model. A MobileNetV2 model trained centrally on PlantVillage achieved 99.38% test accuracy (standard deviation 0.05 over three seeds) on a fixed test set of 5,428 images across 38 classes. Federated models initialized with ImageNet weights were trained for 30 rounds across five clients. FedAvg achieved 98.81% accuracy with identically distributed clients and 98.81%, 98.59%, and 97.78% under Dirichlet label skew (alpha = 1, 0.5, and 0.1, respectively). FedProx was not significantly different from FedAvg at any heterogeneity level (McNemar p > 0.3). Grad-CAM and SmoothGrad explanations were evaluated on 380 test images. Both attributed more energy to the leaf than chance (0.70-0.78 versus 0.60), while Grad-CAM maps from federated and centralized models showed Spearman correlations of 0.74-0.76. Agreement decreased as heterogeneity increased, lesion localization was only modestly above chance, and a model-randomization test indicated that SmoothGrad maps depended substantially on image structure rather than learned parameters. In this setting, label skew reduced federated accuracy by 0.6-1.6 percentage points relative to centralized training. FedAvg largely preserved Grad-CAM behavior, while FedProx provided no measurable benefit.
Multimodal models have often scored below text-only baselines in the BabyLM Challenge (Klerings et al., 2024; Takmaz et al., 2025). We test whether additional recurrent computation helps close this gap. Recurrence increases computation by reusing layers without adding parameters or training data. Using a single implementation, we compare standard and looped Transformers under a 100M-word budget, with three seeds per main condition. Each image is represented by one feature vector from a frozen DINOv2 encoder, used as a prefix token. Caption words count toward the budget, so multimodal training also changes the language data. In a pre-registered text-only comparison, a 12-layer stack repeated four times exceeds a standard 48-layer Transformer by 5.4 BLiMP percentage points at matched inference compute. However, multimodal training lowers BLiMP accuracy by about 3.8 points in both architectures, with no clear mitigation from recurrence. An exploratory loop-side no-image control, trained with the multimodal language data and batching schedule, scores 5.7 points below the text-only baseline. Adding the visual prefix raises accuracy by 1.85 points on average, but this contrast is uncertain with three seeds. These results suggest that changes in language data and its presentation are a major source of the drop. Diagnostic tests show that the models use visual information; its contribution to grammatical generalization remains uncertain under this single-vector interface.
Underwater acoustic communication faces challenges such as limited available bandwidth and complex channel environments, which impose demanding requirements on efficient image transmission and reliable reconstruction. To address these challenges, this paper proposes a generative image communication framework based on cross-modal semantic successive refinement. At the transmitter, text descriptions, compact bottleneck features, and visual side information compressed by residual vector quantization are transmitted. At the receiver, text and bottleneck features are first fused to predict a base estimate of the intermediate visual representation. The received RVQ indices are then used to recover supplementary visual information and perform residual correction, followed by image reconstruction using a one-step Pixel Mean Flow model. During training, a CLIP-based text-image semantic consistency constraint is introduced to preserve the semantic correspondence between generated results and the original text descriptions. During transmission, protection resources are allocated according to the importance of different information branches and RVQ quantization stages. Simulations using COCO 2017 images and measured underwater acoustic channel records from Watermark show that text and bottleneck features provide complementary semantic and structural conditions, RVQ visual side information further improves reconstruction quality, and layered unequal error protection helps improve recovery at low SNRs. Code-rate allocation experiments reveal a tradeoff between the reliability of fundamental conditions and visual-detail reconstruction under a limited visual-feature coding budget, providing guidance for allocating protection resources in underwater acoustic image transmission.
We estimate the quality of a specific target region (e.g., a train body) in an image, rather than the whole image. This differs fundamentally from whole-image No-Reference Image Quality Assessment (NR-IQA): when the target occupies a small fraction of the frame, the whole-image score is dominated by background, and no public benchmark provides target-level human Mean Opinion Score (MOS). We estimate target quality via nested sampling - a sequence of patches expanding outward from the target box - and infer quality from the relative structure of this sequence. This motivates BiCG-Net, a student network that distills an 11-teacher ensemble through a linear observation model and an in-image differencing mechanism, which provably cancels the target term and teacher-specific bias. We train and test the student independently on each dataset, and evaluate zero-shot transfer across benchmarks. On CSIQ and LIVEC, the student reaches PLCCs of 0.9009 and 0.9133 against human MOS, exceeding the best single teacher and the teacher consensus. On a self-built MOS-free railway dataset, the student reaches a PLCC of 0.9855 against a weak label. The student exceeds its teachers because it learns a new mapping from the relative structure of teacher readings to the ground-truth label, rather than mimicking them. Residual analysis on CSIQ and LIVEC shows the prediction residual is significantly correlated with background contamination, teacher dispersion, and label extremity, demonstrating the student extracts information the consensus does not contain. Under CSIQ-to-LIVEC transfer, the frozen student reaches a PLCC of 0.8416, exceeding the consensus by 0.1947. We position BiCG-Net as an offline target-level quality auditing and label-generation engine; its 39,482-parameter student is compact by design, acting as a unifying layer mapping heterogeneous teacher readings onto a common quality scale.
Object detection research remains constrained by high annotation costs and the closed-set assumptions of supervised models, limiting scalability to novel categories. We present a target-dataset annotation-free object detection framework that eliminates the need for human labeling on the target dataset by leveraging the complementary strengths of pretrained foundation models, including segmentation and vision-language systems, combined with a self-refining training loop. To further reduce reliance on expensive LVLM queries, we introduce a landmark-based clustering mechanism that amortizes labeling across image clusters, improving label consistency while lowering computational cost. Our method autonomously generates segmentation proposals, assigns semantic labels through multimodal reasoning, and iteratively improves both localization and classification using a lightweight open-set detector. This modular pipeline enables adaptation to completely unseen object classes and domains without any human-provided target-dataset annotations, while maintaining compatibility with existing detection architectures. Evaluations on COCO, LVIS, ODinW-35, and Roboflow-100 demonstrate substantial performance gains over state-of-the-art open-vocabulary and long-tail detection methods, achieving 60.8 AP on COCO and 47.4 AP on LVIS under zero-shot settings. Beyond strong benchmark results, our approach provides a scalable foundation for real-world detection systems in unknown environments, where annotation is infeasible.
Dimensionality reduction (DR) techniques are fundamental to the analysis of high-dimensional data, yet their substantial computational cost limits their use in interactive settings. To address this limitation, we present GPU-accelerated Progressive Uniform Manifold Approximation and Projection (GPUMAP), an efficient progressive DR algorithm designed for interactive use. We identify three atomic operations within UMAP and restructure them into a progressive pipeline. As part of this adaptation, we introduce modifications to UMAP, including operation queues, instance-wise local join, and scheduling strategies. We further accelerate computation by leveraging GPU parallelism, enabling interactive analysis of million-scale datasets on commodity GPU hardware. Finally, we employ our implementation to systematically study the behavior of progressive DR, investigating how projections evolve over time relative to the ground-truth projection and how biased data streams influence the quality and stability of the resulting projections.
Underwater wireless sensor networks (UWSNs) face critical challenges in energy imbalance, connectivity, and data collection due to harsh environments. To address these issues, this paper focuses on the problem is how to achieve energy-balanced data collection in IoT-enabled UWSNs with a mobility-aware mobile edge computing (MEC), named MECEBDC. Unlike existing approaches, the proposed scheme jointly optimizes energy-aware clustering, cluster head selection, and AUV-assisted data collection. Furthermore, we design a novel metaheuristic-based AUV routing algorithm that incorporates realistic underwater mobility dynamics, explicitly considering gravitational pressure and buoyancy forces acting on the AUV. By leveraging the computing, storage, and mobility capabilities of MEC-enabled AUVs, the proposed approach enables efficient trajectory planning and adaptive data gathering in a stable 3D underwater environment. Extensive simulations and theoretical analysis demonstrate that the proposed method significantly reduces node energy consumption, improves data collection efficiency, and prolongs network lifetime compared with existing schemes, making it well suited for long-term underwater IoT applications.
Knowledge editing has become a key technique for updating knowledge in large language models. However, existing methods neglect the geometric structure of the knowledge representation space, leading to two key limitations: they cannot distinguish which pre-trained knowledge should be preferentially preserved (core vs. non-core) nor which new edits are inherently difficult (hard vs. easy). To address this, we propose GEdit, a geometry-aware framework built on a Low-Rank Plus Diagonal (LR+D) factor model that captures the intrinsic low-rank structure of Transformer FFN hidden representations. GEdit features two complementary mechanisms. First, the low-rank regularization applies targeted protection on a low-rank core knowledge subspace. Second, the adaptive regularization dynamically adjusts the regularization strength based on each edit request's editability: relaxing constraints for hard edits to ensure success, and tightening them for easy edits to better preserve pretrained knowledge. Extensive experiments on GPT2-XL, GPT-J, and LLaMA-3-8B under large-scale editing scenarios, including both mini-batch and fully sequential editing settings with up to 10,000 continuous updates, show that GEdit consistently and significantly outperforms state-of-the-art baselines in both editing success and knowledge preservation.
We present the Rohbau3D-Semantics dataset, which adds dense semantic and instance-level labels to laser scans of the previously released Rohbau3D domain data from shell construction sites. The dataset comprises 504 point cloud scenes from 14 survey sites, each captured with a terrestrial laser scanner and stored in a scanner-centered local coordinate system. Unlike point clouds sampled from reconstructed meshes or synthetic building models, the data preserve the characteristic properties of real laser scanning, including non-uniform sampling, measurement artifacts, reflections, and fine geometric detail. Combined with manual labels, expert oversight during pointwise annotation, and full-scene review, this yields a high-quality benchmark for semantic scene understanding in complex built environments. The dataset includes 17 class labels of building components and one background class, along with instance annotations for object-level analysis. In addition to the 3D point clouds, the release also provides scripts for deriving geometrically aligned panoramic and cube-map representations of color, LiDAR reflection intensity, depth, surface normals, semantic labels, and instance labels from the 3D arrays. We describe the acquisition pipeline, annotation protocol, quality-control process, and panorama-generation procedure, and provide benchmark splits and baseline experiments for semantic segmentation. By combining precise terrestrial laser scans, dense annotations, and a wide variety of shell-construction scenes, Rohbau3D-Semantics bridges the gap among synthetic indoor datasets, mesh-derived benchmarks, and smaller construction-specific resources. The dataset is intended to support research on 3D scene understanding, construction-site analytics, and multimodal scene analysis in the built environment. Access and download the full dataset via https://github.com/RauchLukas/rohbau3d
AI clusters keep data in remote object stores and rely on local caches to stop GPUs from waiting for data. A cache performs well only when its prefetching and eviction policies match the access pattern, yet these patterns differ across AI workloads and change between phases of one workflow. Existing methods either need code changes, which compromises user transparency, or use low-end heuristics, which cannot follow such changes. Large language models (LLMs) can infer access behavior without user hints, but direct LLM control of cache actions suffers from long reasoning delay, high token cost, and poor decision quality. We present Chiron, a cache system based on advisory switching control. An LLM asynchronously diagnoses the access pattern and recommends a policy. A deterministic harness gates this advice, executes expert-designed policies locally, and returns cache feedback for review. We implement Chiron atop JuiceFS and evaluate it on nine AI workloads. It reduces job completion time by 14.3-41.2% over the best baseline on popular AI workloads.
Deploying an explainable intrusion detection model from a validated laboratory setting into a forensic workflow requires guarantees that benchmarks do not test: that the certified model is the exact artifact applied to a case, that heterogeneous evidence can be classified reliably, and that every step is verifiable. We present MODEXRE, a forensic pipeline that extends a previously validated synthetic-data-trained XGBoost classifier with SHAP explainability (arXiv:2607.00763) into a dual-mode architecture: a Laboratory mode, where models are trained and certified on public and synthetic data, and a Formal mode, where a frozen, hash-verified model performs inference on case evidence under a SHA-256 hash-chained chain of custody. Suricata alerts, CEF firewall logs and packet captures are normalized through OCSF. Four design decisions, each triggered by an observed failure, shaped the architecture: a timestamp leak that inflated Kitsune performance (F1-macro 0.984, 0.662 after correction); a feature-coverage gap that made the 78-feature reference model misclassify a real SSH alert; causal temporal flow aggregation, which raised Kitsune multiclass F1-macro from 0.394 to 0.932; and a certification defect, exposed only by a real packet capture, whereby the pipeline had always trained a binary classifier. After correction, the same capture went from 100% of events flagged as malicious to 87% classified as Normal, with the scanning host identified as Reconnaissance. The default four-class certified model reaches F1-macro = 0.9538 on CICIDS2017-derived evidence; the recertified 18-category model reaches 0.6214, reflecting the real multiclass task. Chain-of-custody verification, model integrity hashing and per-prediction SHAP explanations are reported as measured outcomes.
While Large Language Models (LLMs) excel in high-resource languages, their performance often degrades in low-resource languages due to tokenization inefficiencies, generative instability, and linguistic interference. This paper presents AssameseLM-E2B and its instruction-tuned variant, AssameseLM-E2B-it, a 2-billion-parameter model adapted specifically for low resource Assamese language. Built upon the Gemma-4 architecture, the model incorporates an expanded byte-level BPE vocabulary to reduce tokenization overhead by nearly 500% compared with multilingual 2B--3B baselines. We employ a parameter-efficient Continual Pre-Training (CPT) pipeline with Hybrid LoRA, followed by Supervised Fine-Tuning (SFT) on a balanced, multi-domain dataset to enhance instruction-following and Retrieval-Augmented Generation (RAG) capabilities. Evaluation demonstrates a Bits-per-Byte (BPB) score of 0.7488 on FLORES-200, alongside an overall RAG score of 96.7% and 100% refusal accuracy on adversarial unanswerable prompts. Comparative analysis reveals significant limitations in general-purpose baselines, including receptive-productive decoupling, generative degradation on Romanized Assamese, and unintended Bengali script assimilation. These results demonstrate the potential of parameter-efficient adaptation for developing capable, locally deployable language models for low-resource languages under constrained computational resources.
Multimodal large reasoning models (MLRMs) integrate visual inputs with textual and parametric knowledge, allowing sensitive information to be accessed through cross-modal associations during reasoning. This raises privacy concerns when a visual input should no longer provide access to such information. Machine unlearning can remove this influence without retraining the model from scratch. However, multimodal unlearning methods mainly target samples, concepts, or knowledge content, without separating the knowledge from the cross-modal association through which it is accessed. In reasoning models, target information may disappear from the final answer while remaining in intermediate reasoning, or be replaced by incorrect substitute knowledge. We therefore study cross-modal association unlearning, where the forget target is a designated visual-to-knowledge association: the model should suppress target-bearing behavior under the designated visual context while preserving access to the same knowledge in permitted contexts. We propose Visual-Keyed Counterfactual Rebinding (VKCR), a parameter-efficient framework for cross-modal association editing. VKCR first localizes the edit to a compact association-conditioned parameter subspace. It then reduces support for the designated target at task outputs and counterfactually rebinds the affected reasoning behavior by discouraging target-bearing or unsupported continuations while favoring valid non-sensitive alternatives. We further introduce CLEAR-R and MLLMU-R to evaluate target leakage, hallucinated reassociation, reasoning preservation, and modality-specific utility. Experiments show that VKCR substantially reduces image-conditioned target leakage and hallucinated reassociation, including across heterogeneous task formulations, while preserving non-sensitive visual reasoning and permitted text-only knowledge.
Existing reasoning segmentation frameworks typically employ multimodal large language models (MLLMs) with broad world knowledge to extract semantics in a human-like manner and generate corresponding segmentation masks. Nevertheless, these frameworks lack the necessary scene-specific awareness. This paper introduces driving scene reasoning segmentation (DSRS), a scene-specific task that requires models to interpret implicit instructions, infer spatial relations, detect potential risks, handle challenging conditions, and comprehend underlying traffic rules, ultimately generating accurate segmentation masks. Meanwhile, we establish a benchmark dataset to facilitate future research on the DSRS task. To pioneer this task, we propose DSRS-Net, a novel architecture with advanced reasoning segmentation capability for automated driving. Specifically, DSRS-Net seamlessly infuses geometric features for deeper reasoning segmentation, without compromising the original multimodal fusion mechanism. This strategy effectively mitigates the modality representation mismatch in MLLMs caused by RGB-only pretraining, while enhancing the integration of geometric, contextual, and linguistic features. Extensive experiments demonstrate the superior performance of DSRS-Net in both driving scene reasoning segmentation and other language-driven segmentation tasks across various public datasets. The source code and dataset will be released upon publication.
We study adversarial market making with orders that remain active until execution. The maker operates on two markets: a primary market where orders are made and a secondary market where the resulting positions are hedged. In each round, the maker posts a quote on the primary market before observing both primary and secondary market prices. An order executes when the primary price reaches its quote, and the resulting position is immediately offset at the secondary price. When both price sequences are chosen by an adversary over $T$ rounds, we give a simple deterministic policy with regret of order $\log T$ against the best fixed quote (static comparator) and $\sqrt T$ against the best quote for each submission date (dynamic comparator), plus two terms measuring exposure to future primary and secondary prices. These rates remain optimal even when both exposure terms vanish and lower bounds show that the exposure terms are necessary. Our policy naturally extends to simultaneous bids and asks, where we show that static and dynamic regret are both characterized (up to constant factors) by the exposures terms (where the primary price exposure $V_T^{\mathrm{two}}$ now accounts for both bids and asks) and a term of order $\sqrt{V_T^{\mathrm{two}}\log T}$. Our policy is simple, efficient and fully parameterless. We extend this setting by demonstrating how natural assumptions on the price processes lead to improved guarantees without changing the policy.
Many high-stakes regression decisions ask three things of a model at once: a point estimate to act on, an interval that bounds the risk of acting, and the option to abstain when neither can be trusted. Deep learning provides these in isolation: prediction-interval methods return an interval, and at best a value estimate inside it, for every input, while selective-prediction methods learn when to abstain, but only for point predictions and with no statement of how far the truth may lie from the estimate. We introduce SelPI, which trains interval, value and abstention heads jointly under a single loss with an explicit acceptance-rate budget. A natural objection is that interval width already tells us when to abstain. We show that it does not: across 18 tabular benchmarks, width predicts large point errors well (AUROC 0.72) but is anti-correlated with interval misses (AUROC 0.38), so abstaining on wide intervals discards precisely the inputs the interval got right. SelPI therefore selects with two gates, one trained on the interval loss and one on the value loss, coupled by a Fr'echet penalty that ties their separate budgets to the joint acceptance rate they deliver. With the base predictor held fixed, the learned gate beats width thresholding on a proper interval score on all 18 tabular datasets, closing a median 87% of the gap to an oracle selector, and it also beats selection by deep-ensemble disagreement on every dataset. The advantage grows with the heteroscedasticity of the task, and we report the settings in which it disappears.
Segment Anything Model 3 (SAM 3) can segment objects from text prompts and track them across video, but its memory and computation costs make deployment difficult. Post-training quantization (PTQ) reduces these costs by using fewer bits. However, a quantized model can select the wrong object or lose track of its target across frames. We call this prompt-alignment drift. Controlled tests show that perturbing low-variance feature directions strongly affects target selection. We propose Semantics-Preserving Quantization (SPQ), which gives more calibration weight to concept-sensitive parameters, preserves residual features, and corrects the mean error in video memory. Fixed text concepts guide calibration without task annotations. All three components operate offline and add no inference operators. With a W4A8 image encoder, SPQ improves RefCOCO Pr@0.5 from 42.1 to 72.5 over PTQ4SAM while retaining 54.8 COCO mask AP. Experiments on LVIS, RefCOCO variants, and SA-V show consistent gains. A CLIP study also shows improved image-text retrieval.