Seeing Isn't Believing: Uncovering Blind Spots in Evaluator Vision-Language Models

Large Vision-Language Models (VLMs) are increasingly used to evaluate outputs of other models, for image-to-text (I2T) tasks such as visual question answering, and text-to-image (T2I) generation tasks. Despite this growing reliance, the reliability of these Evaluator VLMs remains under explored. In this work, we systematically evaluate the reliability of Evaluator VLMs across both I2T and T2I tasks. We introduce targeted perturbations that degrade output quality along key error dimensions, including object hallucinations, spatial reasoning, factual grounding, and visual fidelity. These perturbations test whether Evaluator VLMs can reliably account for these quality degrading errors in their evaluations. Using a comprehensive benchmark of over 4000 perturbed instances spanning 40 perturbation dimensions, we evaluate 4 prominent VLMs using single-answer scoring, pairwise comparison, and reference-guided paradigms. Our findings reveal that current VLM evaluators exhibit substantial blind spots: they often fail to detect perturbed outputs - in some cases exceeding 50%, struggle particularly with fine-grained compositional and spatial errors, and are often insensitive to hallucinated content that contradicts the input image. Pairwise comparison proves more reliable, though failure rates persist. These results highlight the unreliable nature of current Evaluator VLMs and urge caution in their deployment for benchmarking and development decisions. Code and data have been made publicly available.

Transformer-Based Rhythm Quantization of Performance MIDI Using Beat Annotations

Rhythm transcription is a key subtask of notation-level Automatic Music Transcription (AMT). While deep learning models have been extensively used for detecting the metrical grid in audio and MIDI performances, beat-based rhythm quantization remains largely unexplored. In this work, we introduce a novel deep learning approach for quantizing MIDI performances using a priori beat information. Our method leverages the transformer architecture to effectively process synchronized score and performance data for training a quantization model. Key components of our approach include dataset preparation, a beat-based pre-quantization method to align performance and score times within a unified framework, and a MIDI tokenizer tailored for this task. We adapt a transformer model based on the T5 architecture to meet the specific requirements of rhythm quantization. The model is evaluated using a set of score-level metrics designed for objective assessment of quantization performance. Through systematic evaluation, we optimize both data representation and model architecture. Additionally, we apply performance and score augmentations, such as transposition, note deletion, and performance-side time jitter, to enhance the model's robustness. Finally, a qualitative analysis compares our model's quantization performance against state-of-the-art probabilistic and deep-learning models on various example pieces. Our model achieves an onset F1-score of 97.3% and a note value accuracy of 83.3% on the ASAP dataset. It generalizes well across time signatures, including those not seen during training, and produces readable score output. Fine-tuning on instrument-specific datasets further improves performance by capturing characteristic rhythmic and melodic patterns. This work contributes a robust and flexible framework for beat-based MIDI quantization using transformer models.

Towards Adaptive Open-Set Object Detection via Category-Level Collaboration Knowledge Mining

Existing object detectors often struggle to generalize across domains while adapting to emerging novel categories. Adaptive open-set object detection (AOOD) addresses this challenge by training on base categories in the source domain and adapting to both base and novel categories in the target domain without target annotations. However, current AOOD methods remain limited by weak cross-domain representations, ambiguity among novel categories, and source-domain feature bias. To address these issues, we propose a category-level collaboration knowledge mining strategy that exploits both inter-class and intra-class relationships across domains. Specifically, we construct a clustering-based memory bank to encode class prototypes, auxiliary features, and intra-class disparity information, and iteratively update it via unsupervised clustering to enhance category-level knowledge representation. We further design a base-to-novel selection metric to discover source-domain features related to novel categories and use them to initialize novel-category classifiers. In addition, an adaptive feature assignment strategy transfers the learned category-level knowledge to the target domain and asynchronously updates the memory bank to alleviate source-domain bias. Extensive experiments on multiple benchmarks show that our method consistently surpasses state-of-the-art AOOD methods by 1.1-5.5 mAP.

Zero-Shot Retail Theft Detection via Orchestrated Vision Models: A Model-Agnostic, Cost-Effective Alternative to Trained Single-Model Systems

Retail theft costs the global economy over \$100 billion annually, yet existing AI-based detection systems require expensive custom model training on proprietary datasets and charge \$200-500/month per store. We present Paza, a zero-shot retail theft detection framework that achieves practical concealment detection without training any model. Our approach orchestrates multiple existing models in a layered pipeline - cheap object detection and pose estimation running continuously, with an expensive vision-language model (VLM) invoked only when behavioral pre-filters trigger. A multi-signal suspicion pre-filter (requiring dwell time plus at least one behavioral signal) reduces VLM invocations by 240x compared to per-frame analysis, bounding calls to <=10/minute and enabling a single GPU to serve 10-20 stores. The architecture is model-agnostic: the VLM component accepts any OpenAI-compatible endpoint, enabling operators to swap between models such as Gemma 4, Qwen3.5-Omni, GPT-4o, or future releases without code changes - ensuring the system improves as the VLM landscape evolves. We evaluate the VLM component on the DCSASS synthesized shoplifting dataset (169 clips, controlled environment), achieving 89.5% precision and 92.8% specificity at 59.3% recall zero-shot - where the recall gap is attributable to sparse frame sampling in offline evaluation rather than VLM reasoning failures, as precision and specificity are the operationally critical metrics determining false alarm rates. We present a detailed cost model showing viability at \$50-100/month per store (3-10x cheaper than commercial alternatives), and introduce a privacy-preserving design that obfuscates faces in the detection pipeline. The source code is available at https://github.com/xHaileab/Paza-AI.

Language Prompt vs. Image Enhancement: Boosting Object Detection With CLIP in Hazy Environments

Object detection in hazy environments is challenging because degraded objects are nearly invisible and their semantics are weakened by environmental noise, making it difficult for detectors to identify. Common approaches involve image enhancement to boost weakened semantics, but these methods are limited by the instability of enhanced modules. This paper proposes a novel solution by employing language prompts to enhance weakened semantics without image enhancement. Specifically, we design Approximation of Mutual Exclusion (AME) to provide credible weights for Cross-Entropy Loss, resulting in CLIP-guided Cross-Entropy Loss (CLIP-CE). The provided weights assess the semantic weakening of objects. Through the backpropagation of CLIP-CE, weakened semantics are enhanced, making degraded objects easier to detect. In addition, we present Fine-tuned AME (FAME) which adaptively fine-tunes the weight of AME based on the predicted confidence. The proposed FAME compensates for the imbalanced optimization in AME. Furthermore, we present HazyCOCO, a large-scale synthetic hazy dataset comprising 61258 images. Experimental results demonstrate that our method achieves state-of-the-art performance. The code and dataset will be released.

Explicit Dropout: Deterministic Regularization for Transformer Architectures

Dropout is a widely used regularization technique in deep learning, but its effects are typically realized through stochastic masking rather than explicit optimization objectives. We propose a deterministic formulation that expresses dropout as an additive regularizer directly incorporated into the training loss. The framework derives explicit regularization terms for Transformer architectures, covering attention query, key, value, and feed-forward components with independently controllable strengths. This formulation removes reliance on stochastic perturbations while providing clearer and fine-grained control over regularization strength. Experiments across image classification, temporal action detection, and audio classification show that explicit dropout matches or outperforms conventional implicit methods, with consistent gains when applied to attention and feed-forward network layers. Ablation studies demonstrate stable performance and controllable regularization through regularization coefficients and dropout rates. Overall, explicit dropout offers a practical and interpretable alternative to stochastic regularization while maintaining architectural flexibility across diverse tasks.

DroneScan-YOLO: Redundancy-Aware Lightweight Detection for Tiny Objects in UAV Imagery

Aerial object detection in UAV imagery presents unique challenges due to the high prevalence of tiny objects, adverse environmental conditions, and strict computational constraints. Standard YOLO-based detectors fail to address these jointly: their minimum detection stride of 8 pixels renders sub-32px objects nearly undetectable, their CIoU loss produces zero gradients for non-overlapping tiny boxes, and their architectures contain significant filter redundancy. We propose DroneScan-YOLO, a holistic system contribution that addresses these limitations through four coordinated design choices: (1) increased input resolution of 1280x1280 to maximize spatial detail for tiny objects, (2) RPA-Block, a dynamic filter pruning mechanism based on lazy cosine-similarity updates with a 10-epoch warm-up period, (3) MSFD, a lightweight P2 detection branch at stride 4 adding only 114,592 parameters (+1.1%), and (4) SAL-NWD, a hybrid loss combining Normalized Wasserstein Distance with size-adaptive CIoU weighting, integrated into YOLOv8's TaskAligned assignment pipeline. Evaluated on VisDrone2019-DET, DroneScan-YOLO achieves 55.3% mAP@50 and 35.6% mAP@50-95, outperforming the YOLOv8s baseline by +16.6 and +12.3 points respectively, improving recall from 0.374 to 0.518, and maintaining 96.7 FPS inference speed with only +4.1% parameters. Gains are most pronounced on tiny object classes: bicycle AP@50 improves from 0.114 to 0.328 (+187%), and awning-tricycle from 0.156 to 0.237 (+52%).

A Variational Message Passing Framework for Multi-Sensor Multi-Object Tracking using Raw Radar Signals

The growing proliferation of unmanned aerial vehicles (UAVs) poses major challenges for reliable airspace surveillance, as drones are typically small, have low radar cross-sections, and often move slowly in cluttered environments. These characteristics make the joint tasks of detecting, localizing, and tracking multiple objects difficult for conventional detect-then-track (DTT) approaches, which rely on pre-processed measurements and may discard informative low-signal-to-noise ratio (SNR) signal components. To overcome these limitations, we propose a variational message passing (VMP)-based direct multiobject tracking (MOT) method that operates directly on raw radar signals and explicitly accounts for an unknown and time-varying number of objects. The proposed method is formulated for MIMO multi-radar systems and performs data fusion by jointly processing the signals of all radar sensors using a probabilistic model. A superimposed signal model is employed to capture correlations in the raw sensor data caused by closely spaced objects, and a hierarchical Bernoulli-Gamma model is introduced to jointly model object existence, reflectivities, and the reliability of individual radar-object links. Using a mean-field approximation, we derive message updates, yielding a computationally efficient VMP algorithm that simultaneously performs object detection, track formation, state estimation, and nuisance parameter learning directly from the radar signal. Simulation results in synthetic scenarios with weak and closely-spaced objects show that the proposed direct-MOT method outperforms a conventional pipeline based on super-resolution estimation followed by belief propagation (BP)-based tracking, particularly in low-SNR and clutter-rich conditions, demonstrating the advantages of direct signal-level inference and coherent multi-radar fusion.

Sparse Hypergraph-Enhanced Frame-Event Object Detection with Fine-Grained MoE

Integrating frame-based RGB cameras with event streams offers a promising solution for robust object detection under challenging dynamic conditions. However, the inherent heterogeneity and data redundancy of these modalities often lead to prohibitive computational overhead or suboptimal feature fusion. In this paper, we propose Hyper-FEOD, a high-performance and efficient detection framework, which synergistically optimizes multi-modal interaction through two core components. First, we introduce Sparse Hypergraph-enhanced Cross-Modal Fusion (S-HCF), which leverages the inherent sparsity of event streams to construct an event-guided activity map. By performing high-order hypergraph modeling exclusively on selected motion-critical sparse tokens, S-HCF captures complex non-local dependencies between RGB and event data while overcoming the traditional complexity bottlenecks of hypergraph computation. Second, we design a Fine-Grained Mixture of Experts (FG-MoE) Enhancement module to address the diverse semantic requirements of different image regions. This module employs specialized hypergraph experts tailored for object boundaries, internal textures, and backgrounds, utilizing a pixel-level spatial gating mechanism to adaptively route and enhance features. Combined with a load-balancing loss and zero-initialization strategy, FG-MoE ensures stable training and precise feature refinement without disrupting the pre-trained backbone's distribution. Experimental results on mainstream RGB-Event benchmarks demonstrate that Hyper-FEOD achieves a superior accuracy-efficiency trade-off, outperforming state-of-the-art methods while maintaining a lightweight footprint suitable for real-time edge deployment.

LLM-as-Judge Framework for Evaluating Tone-Induced Hallucination in Vision-Language Models

Vision-Language Models (VLMs) are increasingly deployed in settings where reliable visual grounding carries operational consequences, yet their behavior under progressively coercive prompt phrasing remains undercharacterized. Existing hallucination benchmarks predominantly rely on neutral prompts and binary detection, leaving open how both the incidence and the intensity of fabrication respond to graded linguistic pressure across structurally distinct task types. We present Ghost-100, a procedurally constructed benchmark of 800 synthetically generated images spanning eight categories across three task families: text-illegibility, time-reading, and object-absence, each designed under a negative-ground-truth principle that guarantees the queried target is absent, illegible, or indeterminate by construction. Every image is paired with five prompts drawn from a structured 5-Level Prompt Intensity Framework, holding the image and task identity fixed while varying only directive force, so that tone is isolated as the sole independent variable. We adopt a dual-track evaluation protocol: a rule-based H-Rate measuring the proportion of responses in which a model crosses from grounded refusal into unsupported positive commitment, and a GPT-4o-mini-judged H-Score on a 1-5 scale characterizing the confidence and specificity of fabrication once it occurs. We additionally release a three-stage automated validation workflow, which retrospectively confirms 717 of 800 images as strictly compliant. Evaluating nine open-weight VLMs, we find that H-Rate and H-Score dissociate substantially across model families, reading-style and presence-detection subsets respond to prompt pressure in qualitatively different ways, and several models exhibit non-monotonic sensitivity peaking at intermediate tone levels: patterns that aggregate metrics obscure.