This paper presents a unified framework for license plate detection and character recognition on low-power edge hardware. The proposed network uses a shared CSP-Darknet and PANet feature encoder with an anchor-free localization head and a RoIAlign-based sequence-recognition branch. Character decoding is performed directly on aligned feature maps, avoiding image-space plate cropping and a second recognition backbone. The model is trained on an AI-assisted dataset of approximately 37,000 manually verified images covering license plate layouts from all 50 United States. The dataset includes variations in illumination, viewpoint, weather, blur, occlusion, and plate design. The framework is converted to an INT8 TensorRT engine and evaluated on an NVIDIA Jetson Orin Nano with 8 GB memory. The INT8 model shows an mAP50 of 97.8% and a full-sequence accuracy of 90.7%. It records 147 FPS at 6.8 ms end-to-end latency, with 315 MB inference memory and 9.4 W average power consumption. Relative to the YOLOv8s+CRNN baseline, the proposed deployment provides a 7.1x throughput increase and a 77.8% reduction in allocated inference memory
We evaluate AI agents in an online setting where tasks arrive sequentially and, before running each task, we construct an upper predictive bound on its \emph{time-to-success}---the number of interactions until the task is completed successfully. Under a limited budget, some runs are stopped early, so their time-to-success is unobserved (right-censored), and the correctness of the bound may be unknown. We introduce \emph{\textbf{PR}edictive Bounds for \textbf{O}nline \textbf{B}udget \textbf{E}fficiency} (\texttt{PROBE}), a method for constructing calibrated upper predictive bounds while adaptively allocating a limited budget across agent tasks to reduce the anticipated coverage deficit. Our method calibrates the bounds online using inverse-probability-weighted feedback together with a correction designed to minimize coverage deficit while remaining unbiased. We prove that \texttt{PROBE} provides a coverage guarantee for any finite time horizon while satisfying the budget constraint. Experiments on software engineering, retrieval-augmented generation, and instruction following show that \texttt{PROBE} attains a coverage rate closer to the target level with tighter predictive bounds than a calibrated baseline.
Recommendation systems have become more important since the widespread use of the Internet. The large amount of information means that it is difficult for users to discover they really need. However, users' preferences can change over time because of the age and the impact of social networks. Recommendation systems must capture users' preferences and recommend suitable items, which is a difficult challenge. This study proposes a novel recommendation system that adapts to the changing preferences of users. By learning and adapting to users' changing preferences better recommendations are provided. This study uses context factors as additional information to model users' preferences more precisely. For the proposed recommendation system, the context is based on users' interaction items within each defined time interval. This study proposes two novel attention mechanisms for the recommendation system. The contextual item attention module captures contextual information, changing pattern in users' preference and the importance of items. The multi-head attention module extends the diversity of users' preferences and adapts to changing preferences. The recommendation performance is improved using additional item's temporal information to model the contextual item's representation. Experiments compare the proposed algorithm with several state-of-the-art recommendation methods using three real-world datasets. The experimental results demonstrate that the proposed context-aware recommendation model outperforms traditional methods and demonstrate the effectiveness with which contextual information is captured by the attention mechanism.
Stochastic first-order methods use sampled gradients to reduce per-iteration costs, but estimation errors can slow convergence and, in some settings, cause divergence. To improve gradient accuracy while retaining acceleration, we develop scalar and diagonal versions of VR-Adam-SHANG for smooth stochastic convex optimization. The methods combine STORM-type recursive estimation, Adam-style preconditioning, and SHANG's accelerated interpolation and correction, using at most two stochastic gradient evaluations per iteration without full-gradient refreshes. Under unbiased sampling, bounded variance, and mean-square smoothness of gradient noise, polynomial steps with prescribed initialization yield $O(1/K)$ mean-square gradient-estimation error and expected objective-gap bounds combining $O(1/K^2)$ initial-energy decay with an $O(1/\sqrt K)$ stochastic term after $K$ iterations. These guarantees require neither monotonicity of the preconditioner nor a uniform positive lower bound on its eigenvalues. The analysis also establishes polynomial-step convergence of projected nonrecursive Adam-SHANG-s, filling a gap in its original theory. Experiments compare the methods at matched sample-gradient budgets. In every tested setting, at least one of the two versions achieves the lowest mean terminal objective gap among all compared methods.
Clinical speech is increasingly used for disease screening and clinical assessment, but directly distributing speech recordings poses privacy risks because they contain both speaker-identifying voiceprints and sensitive spoken content. Existing speech anonymisation approaches are primarily developed for general speech applications or task-specific clinical speech and overlook that different clinical tasks rely on distinct speech characteristics, necessitating task-adaptive anonymisation strategies to protect patient privacy while preserving clinical utility. Here we introduce CASA, a clinically adaptive speech anonymisation agent that transforms anonymisation from a fixed processing procedure into a task-guided optimisation process, dynamically tailoring voiceprint and content anonymisation to task-specific privacy and clinical utility requirements. Through recursive self-improvement (RSI), CASA iteratively evaluates and refines anonymisation strategies against task-specific privacy and clinical utility requirements, while learning from these outcomes to guide more efficient planning in subsequent clinical tasks. CASA is validated across 14 clinical tasks encompassing five languages, diverse elicitation paradigms, and mental health, cognitive and voice disorders. CASA preserved disease-related classification performance while meeting speaker-anonymisation requirements across all clinical tasks, versus only 8/14 for the strongest fixed method. Assessments of resistance to speaker re-identification by external attackers and identification and redaction of sensitive personal information in clinical speech further demonstrated CASA's dual protection of speaker identity and sensitive spoken content. Clinical expert assessments further supported the preservation of clinically relevant information and the acceptability of anonymised speech for controlled research sharing.
Dichotomous image segmentation requires pixel-accurate foreground-background separation while preserving fine structures such as thin filaments, sharp edges, and intricate textures. BiRefNet addressed this challenge through bilateral references, combining image patches and gradient information, but its inward reference loses local spatial structure and lacks semantic guidance, while its outward reference relies on single-channel gradient attention. We recast this paradigm as controlled information recovery - locality determines what spatial information survives, semantic selectivity what information enters, and focused boundary supervision where refinement occurs - and propose BiRefNet-V2 to implement it: (1) locality-preserving and semantically guided detail injection through space-to-depth transformation, overlapping patches, and a Semantically Guided Gate; (2) a learned boundary stream with focused supervision that generates full-channel residual attention without requiring boundary annotations; and (3) a dual-branch decoder integrating depth-wise boundary convolutions with large-kernel attention. A modernized framework - LayerNorm, GELU, the Lion optimizer, and a DINOv3 backbone - cuts training from 600 to 50 epochs. Extensive experiments demonstrate state-of-the-art performance on DIS5K, camouflaged object detection, and high-resolution salient object detection across all 13 evaluation sets. Code is available at https://github.com/ZhengPeng7/BiRefNet-V2.
As enterprises delegate consequential work to multi-agent LLM systems, benign-looking requests can escalate into unsafe actions as they propagate across sessions through poisoned shared state, long-context documents, and trusted peer messages: no single message is malicious, yet the assembled tool action is. Prior benchmarks isolate a single agent under prompt injection or score multi-agent attacks by their final output, leaving cross-session erosion of shared state out of scope. We introduce ErosionBench, an environment-grounded benchmark that executes such attacks and verifies alignment decay at the process level, scoring the realized tool call against authorized intent and provenance with deterministic contracts, not output text. Each scenario is authored as a deterministic state graph before any prose is rendered, so every state change traces to a known cause. It spans four enterprise domains and four decay mechanisms across 1,728 scenarios with matched benign controls, labeled against the OWASP Top-10 for LLM Applications, plus 198 image-based attacks, each paired with a text equivalent, so a defense must survive the same attack in either channel. Perimeter defenses suppress attacks only by over-refusing 54-92% of legitimate work, so a low attack-success rate can conceal severe over-refusal. As a reference defense, we propose a calibrated Meta-Auditor pairing synchronous state checks with asynchronous trajectory monitoring; on a family-disjoint 906-scenario held-out split it reaches 98.5% recall at a 2.9% false-positive rate. Since it monitors trajectory, not surface content, its detection is channel-invariant: the strongest auditor flags image-delivered attacks about as reliably as their text equivalents (recall gap $\leq$2 points per model). Across GPT, Gemini, and Claude auditors, calibration keeps false positives low while recall tracks model capability, strongest for Claude.
Positron emission tomography (PET) is pivotal in medicine and healthcare, but requires high radiation doses for quality imaging. Although deep learning has shown promise for low-dose PET reconstruction/enhancement, existing methods primarily rely on fully supervised learning, which requires paired low- and full-dose scans that are clinically impractical to obtain. To overcome this fundamental limitation, we propose DuSS-PET, a novel self-supervised PET reconstruction framework that operates without any paired low- and full-dose data. Our method is built upon a computational Noisier2Noise (Nr2N) paradigm, featuring a Bernoulli corruption strategy that serves as an image-space approximation to emulate low-count noise patterns in image-derived pseudo-sinogram representations. The DuSS-PET framework integrates three core components: an adaptive Swin transformer for pseudo-sinogram restoration, a plug-and-play module for image enhancement, and a consistency-constrained SSL strategy that enforces agreement between the pseudo-sinogram and image representations. Extensive experiments on multi-center datasets demonstrate that DuSS-PET achieves superior reconstruction performance, outperforming state-of-the-art SSL methods and achieving comparable or even better results than fully supervised approaches across various dose levels. In particular, it exhibits exceptional generalization across different scanner vendors. This work is among the first self-supervised low-dose PET reconstruction framework that enforces consistency between reconstructed images and pseudo-sinograms, where the pseudo-sinogram generated in the image space serves as a solid computational consistency constraint for enhanced imaging without full-dose training scans. The source code will be publicly available at https://github.com/ladderlab-xjtu/DuSS-PET.
Methods for detecting phishing accounts on Ethereum are usually evaluated as complete pipelines that differ in features, architecture, and training protocol, which makes the contribution of interaction information hard to isolate. This paper evaluates the Unlabeled Feature Extension (UFE), a block of 12 local interaction statistics computed from each account's incoming and outgoing call and transfer edges without reading class labels. UFE is appended to the 14 source features of HEIG-Phish and tested with six model families: Logistic Regression, Random Forest, LightGBM, Graph Attention Network, GraphSAGE, and Meta-IFD. Within each pair of Baseline and Plus variants, the data partition, model configuration, random seed, decision threshold, and sampled neighborhoods are identical; only the input feature block changes. Across 20 predetermined stratified partitions of the same 2,763 labeled externally owned accounts, UFE increased test Macro-F1 by 1.0 to 9.1 percentage points across the six models. Every pair improved on all 20 partitions, all Nadeau-Bengio corrected 95% confidence intervals lie above zero, and the false-positive rate decreased for all six models. An exact-edge Random Forest ablation over 10 predetermined partitions associates most of the gain with the incoming transfer-edge sum attribute; once this feature is present, an incremental contribution of the remaining 11 UFE columns was not established. The results support reporting feature-augmented baselines alongside graph models, but they show stability under re-partitioning of one static, transductive benchmark rather than temporal or cross-dataset generalization.
LLMs are powerful but incur substantial inference cost, which can be exploited by malicious users to induce overly long outputs, increasing latency and operational expense. Prior resource consumption attacks either coerce repetition or suppress the end-of-sequence token by lowering its logit value. However, with autoregressive, sampling-based generation, the search space grows exponentially with length and small deviations or unstable trajectories can derail optimization, limiting attack effectiveness and stability. To address this issue, we introduce Chatter Attack, a novel class of resource consumption attacks that optimize adversarial prompts to induce predefined short target phrases, thereby avoiding search-space explosion while still eliciting long responses. Specifically, Chatter Attack triggers self-doubt using short doubt-inducing phrases as the target, and sustains this state throughout generation via inference attention loss and global entropy loss. To extend to black-box settings, we further propose a novel Bayesian Prompt Optimization (BPO) method, which models the objective function globally by constructing a discrete token kernel, efficiently exploring the entire search space using prior information. Extensive experiments across multiple models and datasets show that our method outperforms baselines, with maximum response length increases to 31.5$\times$, and revealing practical resource-depletion risks for real-world LLM inference services.
Does reinforcement learning teach language models new capabilities, or merely amplify ones acquired during pre-training? The literature offers evidence for both: pass@$k$ studies suggest RL reweights latent behaviours while some controlled composition studies demonstrate gains beyond simple amplification. We argue that this question has no answer independent of the capability state of the base model, and propose a two-dimensional state space for post-trainability. The first axis, \emph{effective full-solution support}, measures whether successful trajectories can be sampled under a feasible exploration budget. The second axis, \emph{atomic primitive coverage}, measures whether the constituent skills of a solution can be executed when the burden of composition is removed. Together, these dimensions motivate three regimes -- amplification, composition-limited, and acquisition-limited -- each predicting a distinct empirical signature and calling for a distinct intervention. We propose a prospective measurement protocol and investigate the framework using synthetic composition tasks across three small language models. We find that primitive mastery alone does not guarantee that RL can learn a composition. Checkpoints that execute every primitive reliably can fail to improve when complete solutions are never sampled during training, whereas modest composition supervision that establishes nonzero support enables substantial RL gains. The framework reframes mid-training as preparation for post-training: its role is to move models into states from which RL can operate effectively.
Interpretable glaucoma detection requires a transparent connection between fundus images, clinically meaningful biomarkers, and diagnostic predictions. Knowledge graph (KG) reasoning provides this structure, but performance measured using annotated biomarkers does not establish performance when only images are available. We present Auto-GlaKG, an automated framework that combines image-based biomarker prediction with fixed clinical-prior rules. A ResNet50 network predicts the cup-to-disc ratio (CDR) and six morphological signs using biomarker supervision, and a KG reasoner aggregates their signed rule contributions into a glaucoma score. On 689 fundus images evaluated using stratified, group-aware five-fold cross-validation, Auto-GlaKG achieves a pooled area under the receiver operating characteristic curve (AUC) of 0.951 (95\% confidence interval: 0.931--0.969), with sensitivity of 0.970 and specificity of 0.779. The matched image classifier also achieves an AUC of 0.951, while the rule stage adds 0.2~ms per image. Replacing predicted CDR with its annotation increases fold-mean AUC from 0.953 to 0.995. The framework provides an explicit rule-level account of image-based predictions and identifies CDR estimation as a priority for validation using independently measured biomarkers.
Well-calibrated forecasting for non-linear dynamical systems is essential in science and engineering applications such as fluid dynamics and weather prediction. We present a forecasting method combining probabilistic machine learning with the structured linear dynamics of Koopman mode decomposition. Within this framework, the observation, linear dynamics, and state reconstruction operators are constructed using standard machine learning architectures, producing predictive distributions via the pushforward of injected noise samples. The model is trained using a proper scoring rule, the continuous ranked probability score, to directly target calibrated, accurate forecasts. The approach is validated on two non-linear ordinary differential equation examples and two partial differential equation examples, vortex shedding and cavity flow, using both large and small, noisy datasets. Results demonstrate that known Koopman eigenfunctions and spectra can be identified and used for accurate forecasting. The model can also produce multi-modal predictive distributions, demonstrated for a trajectory initialised on the boundary between two basins of attraction. Uncertainty is captured by the sample ensemble spread, with spread--skill ratios close to one over the training horizon in several experiments, although for extended rollouts and noisy datasets, the ensemble may become under- or over-dispersed. In low-noise experiments, temporal rollout is generally accurate for at least twice the training horizon, and up to twenty times in some cases. Approximation quality is comparable to a deep ensemble approximation over the training horizon, but at a reduced computational cost and with better robustness in identifying Koopman eigenvalues.
Learning neural solvers for spatiotemporal partial differential equations (PDEs) under real-world constraints remains a key challenge in scientific machine learning, especially for inverse tasks with sparse and noisy boundary observations. We present the Aletheia dataset, the first 3D benchmark for learning data-driven solvers in the context of nondestructive testing (NDT). The dataset simulates eddy-current-induced heating in conductive solids and models the resulting transient heat propagation governed by the heat equation. Aletheia contains over 4,700 high-resolution samples across 10 excitation frequencies (1-100,kHz), each providing volumetric heat source and temperature fields over time. It supports both forward prediction of temperature evolution and inverse reconstruction of internal heat sources or defects from surface infrared measurements. Real infrared thermography data from cracked rail specimens are included for calibration and comparison. We define four canonical tasks on both regular and irregular grids and benchmark them using various neural operators. Aletheia establishes a unified platform for evaluating neural PDE solvers under realistic NDT conditions, enabling progress in reliable, data-driven inverse modeling. The dataset and code are publicly available at https://github.com/sherlock-0327/ALETHEIA.
Self-supervised speech models (S3Ms) encode a rich hierarchy of acoustic, prosodic, and phonetic information, yet when and where it emerges during pretraining remains unclear. We train frozen-encoder probes on every Transformer layer across pretraining checkpoints of two matched models: HuBERT-Base, whose stage 2 is a separately initialized model trained on new pseudo-label targets, and single-objective wav2vec 2.0-Base. Our five tasks span an acoustic-to-abstract gradient: absolute pitch (F0), speaker-relative pitch, lexical stress, phoneme boundaries, and 38-way phoneme identity. Acoustic-prosodic cues and phone boundaries are decodable from the earliest checkpoints and barely shift across layers, whereas phoneme identity is constructed during pretraining, re-localizing to upper-middle layers under HuBERT's stage-2 targets at a matched update count. Acquisition order therefore tracks abstraction from the acoustic surface, not the segmental/suprasegmental distinction, extending to prosody an account established for phonetic, lexical and syntactic structure; the no-switch wav2vec 2.0 shows no comparable re-localization.
EmbeddingGemma-300M is a small open embedder whose tokenizer represents every Tajik letter, and as released it is weak at Tajik retrieval: on our Tajik law test the right passage is in its top ten results for 49.3% of questions, on tax documents for 45.9%, and on the human-written Belebele Tajik set for 69.8%, against 98.6% on the Russian version of the same set. We present Saidzoda-Tajik-EmbeddingGemma-300M, the same model fully fine-tuned for one epoch on 235,140 question and passage pairs written by our own Tajik language model for laws, tax documents, Wikipedia and web text, with the same mined hard negatives, English and Russian replay, loss, global batch, learning rate and epoch as our earlier Saidzoda-Tajik-E5-Large (arXiv:2611.00592) and BGE-M3 models. After training, law reaches 92.4% and tax 91.7%; on Belebele, the one test written by people, recall@10 rises from 69.8% to 92.6% and the first result is right for 76.6% of questions instead of 51.7%. A checkpoint at step 100 of 466 already holds 88% of the law gain at recall@10. At recall@10 the result is no more than 2.8 points below the two 560M models trained on the same pairs on law, tax and Wikipedia, within the intervals, with 768-dimensional vectors and about half their parameters; at recall@1, for which we have no intervals, it stays 7.0 points behind the E5 model on Belebele. The cost is a drop of 3.0 and 2.8 points in Russian and English recall@1, larger than for the two earlier models. We also report that, according to our run notes, a bfloat16 test run produced a non-finite loss at step 3 which the training framework logged as zero, so training was done in fp32. Our generated test questions come from the same model as the training questions, so we treat those scores as an upper limit until questions written by Tajik speakers exist.
Large language models (LLMs) are increasingly asked to judge whether a news article is true, and their stated confidence is often taken at face value. We test what happens to that confidence when an article's writing style, but not its facts, is manipulated, as in recent \emph{style attacks} on fake-news detectors. Tracing how the prediction and its confidence form layer by layer inside the network, we study three model families (LLaMA-3-8B, Qwen2.5-7B, Mistral-7B) on three fake-news benchmarks, and find the same three patterns in every model--dataset pairing. First, on the original articles the models are competent and largely honest judges: they settle on an answer, briefly become overconfident, and then correct their confidence in the final layers. Second, style-rewritten articles break this self-correction: the models become far less accurate, under the strongest attack, worse than random guessing, yet remain just as confident, so they are \emph{confidently wrong} precisely when a warning is most needed. Third, nearly all of the damage comes from rewriting into sensational or emotionally charged styles, while sober, neutral rewrites leave the models' honesty largely intact. The practical implication is that an LLM's confidence about veracity cannot be trusted as a safety signal when an article's style may have been manipulated.
Asynchronous pipeline parallelism can accelerate distributed model training, but suffers from instability caused by inconsistent gradient computations. We identify a fundamental source of this inconsistency: error signals are coupled to forward-time features, not just parameter states. This means that parameters may evolve as long as they preserve forward-equivalence. Leveraging this insight, we introduce FRESCO, which frames asynchronous updates as a constrained optimization problem: it finds the minimal update modification that acts trivially on the activation subspace. We demonstrate that FRESCO consistently outperforms representative asynchronous baselines in LLM pre-training, with the gains most pronounced under deep pipelines and high inconsistency levels: at pipeline depth 32, it reaches the same loss as the strongest baseline with 43% fewer iterations and 30% less wall-clock time, while requiring 8~921$\times$ less memory overhead. Moving beyond the current limitations, FRESCO establishes a new foundation for high-utilization pipeline parallel training, providing a scalable path for large-model development on diverse, resource-constrained systems.
Frontline sales staff spend much of their day re-coordinating quotes after customers change quantities, ask for discounts or new payment terms, or costs move. We present LoMASA (Local Multi-Agent Sales Assistant), a training-free multi-agent system that runs entirely on one desktop computer. A 35B mixture-of-experts model (System-1) makes each decision in a single forward pass and reads its confidence from the probabilities of the answer options. A 27B dense model (System-2) writes customer replies and escalation briefs and is consulted when System-1 is unsure. Code computes every price, margin and option, and hard pricing rules cannot be overridden. On a test split of 500 simulated episodes with 1512 decisions built from public supply-chain data, the salesperson has to decide only 19.5% of decisions: 38.0% are settled automatically, with a median latency of 2.67 seconds, and 42.5% are escalated to a manager by the hard rules. On the automatically settled decisions LoMASA is as accurate as a rule baseline (0.894 vs.\ 0.894) and makes fewer commitments whose margin later falls below the floor (2.61% vs 4.53%); the price is additional escalations to the manager. Option probabilities are well calibrated without any training, unlike confidence stated in words by the same model. We also report where the design does not help: the 27B model rarely changes a decision.
Direct Preference Optimization (DPO) is a simple and widely used alternative to RLHF for aligning large language models (LLMs), but its sequence-level objective cannot precisely locate or suppress the local patterns that make an otherwise acceptable response unsafe or undesirable. We propose TLRS-DPO (Token-Level Risk Scoring for Better Direct Preference Optimization), which augments DPO with fine-grained token-level control while remaining fully offline and reference-based. TLRS-DPO introduces a token-level robust compensation term that converts the policy-reference log-ratio on rejected responses into localized risk scores, a curriculum-style schedule that gradually increases the strength of this compensation during training, and a risk-adaptive screening mechanism that selectively penalizes only high-risk tokens instead of entire sequences. This design bridges sequence-level preference fitting and token-level behavior control, and can be implemented as a drop-in modification to standard DPO trainers. Empirically, TLRS-DPO improves alignment quality and safety-relevant behavior with minimal impact on helpfulness compared to strong DPO baselines.