Code search, framed as information retrieval (IR), underpins modern software engineering and increasingly powers retrieval-augmented generation (RAG), improving code discovery, reuse, and the reliability of LLM-based coding.Yet existing code IR models remain largely text-centric and often overlook the visual and structural aspects inherent in programming artifacts such as web interfaces, data visualizations, SVGs, schematic diagrams, and UML.To bridge this gap, we introduce MMCoIR, the first comprehensive benchmark for evaluating multimodal code IR across five visual domains, and show through extensive evaluation the task is challenging.Therefore, we then propose CodeMMR, a unified retrieval model that jointly embeds natural language, code, and images into a shared semantic space through instruction-based multimodal alignment.CodeMMR achieves strong generalization across modalities and languages, outperforming competitive baselines (e.g., UniIR, GME, VLM2Vec) by an average of 10 points on nDCG@10.Moreover, integrating CodeMMR into RAG enhances code generation fidelity and visual grounding on unseen code generation tasks, underscoring the potential of multimodal retrieval as a core enabler for next-generation intelligent programming systems.

We introduce PAST-TIDE, our stance detection system addressing both subtasks of the StanceNakba Shared Task at NakbaNLP@LREC-COLING 2026. The main idea is statement tuning. We redefine stance as cloze-style masked language modeling (MLM), letting a verbalizer map label words to stance categories through the pre-trained MLM head rather than appending a randomly initialized classification head. We complement this with prototypical contrastive learning, which uses learnable class prototypes for batch-size independent contrastive training, and topic-conditional layer normalization for cross-topic Arabic stance detection. PAST-TIDE achieves macro-F1 scores of 0.75 for Subtask A and 0.74 for Subtask B on the official leaderboard, indicating that minimal architectural additions to a pre-trained model can remain competitive in low-resource settings.

Apricots (Prunus Armeniaca) are valuable stone fruits cultivated worldwide in temperate regions, generating $500 million in annual exports. However, disease and pests significantly threaten apricot production, impacting quality and yield. Brown rot and shot hole are the two major diseases affecting apricot yield worldwide. Early detection and targeted management strategies are critical to prevent their spread. Unfortunately, the lack of diverse labeled datasets hinders the performance of deep learning models for in-field disease detection. In this regard, we propose an innovative approach that leverages deep convolutional generative adversarial networks (DCGAN) and model-guided Cutmix (MGC) data augmentation to synthesize images of diseased apricots. The proposed synthesis process is driven by counterexamples, which are samples that the model fails to detect correctly. The use of DCGAN for background and conditional DCGAN for foreground generation allows fine-grained control over the synthesis of new samples. Further, the MGC uses SHapley Additive exPlanations of the object detection model to analyze its weaknesses and guide on-demand sample generation by replicating the counterexample's label distribution, aspect ratio, scale, and brightness/contrast. This expands and diversifies our custom apricot disease dataset, initially containing 1500 images. This addresses dataset imbalances and exposes the model to a dynamically augmented dataset, iteratively improving its performance. The proposed method was extensively evaluated on images of healthy and diseased apricots captured by unmanned aerial vehicles in different environmental conditions. MGC outperformed traditional augmentation methods with even smaller dataset, achieving 4% better mean average precision (mAP) score with Apricot-3K dataset. Additionally, the proposed method achieved 1-4 % improvements in mAP with many state-of-the-art object detection models when trained using the proposed framework on the Apricot-10K augmented dataset.

Background: Depression affects over 280 million people worldwide, yet diagnosis still depends on subjective clinical interviews and self-report questionnaires that are vulnerable to recall bias, social desirability effects, and limited access to trained professionals. Multimodal deep learning may enable more objective depression detection by combining behavioral, verbal, and physiological signals from multiple data sources. Methods: This systematic review followed the PRISMA 2020 guidelines. We searched IEEE Xplore, ACM Digital Library, Web of Science, SpringerLink, and ScienceDirect for peer-reviewed studies published between January 2018 and August 2025. Eligible studies used at least two measured input modalities for depression classification or severity assessment with quantitative evaluation. From an initial pool of 120 records, screening and eligibility assessment yielded 50 included studies. Results: Across clinical, social media, and laboratory settings, multimodal systems outperformed unimodal baselines. Audio-video fusion and electroencephalography combined with facial video or eye tracking produced the largest gains. Alignmentaware audio-text models improved accuracy when reliable timestamps were available. Audio-video pipelines on standard corpora reached precision values near 95%, with sensitivity and specificity frequently above 80%, though metric reporting varied across studies. Accuracy generally plateaued beyond two or three modalities; well-optimized bimodal systems often matched or outperformed larger combinations while remaining easier to deploy. Among fusion strategies, hybrid approaches that combine feature-level and decision-level information most often achieved the highest accuracy and stability, and attention-based or transformer-based fusion added gains in both performance and interpretability. Conclusions: Multimodal deep learning can improve the objectivity of depression assessment, but clinical adoption requires progress on several fronts: standardized evaluation protocols and metrics, transparent handling of temporal alignment and missing modalities, external validation in diverse populations, fairness and uncertainty evaluation, and deployment-oriented research on privacy-preserving and on-device inference. This systematic review was not prospectively registered.

Accurate detection of citrus fruit maturity is critical for optimizing harvest schedules and maximizing yield. Consumer-grade uncrewed aerial vehicles (UAVs) have emerged as cost-effective alternatives to traditional methods for detecting maturity, which rely on labor-intensive manual inspections. This paper presents a two-step, semi-supervised approach leveraging knowledge distillation (KD) and transfer learning for citrus maturity detection in UAV images. Specifically, we combine teacher-filtered pseudo-labels with a consistency-guided feature distillation signal to exploit abundant unlabeled UAV frames while using only a small labeled seed set. Firstly, a consistency-guided KD transfers knowledge from a pretrained detection transformer with collaborative hybrid assignment training (Co-DETR) to a lightweight student network by exploiting a small labeled and a large unlabeled dataset. The student network (Cit-DETR) is based on the highly efficient detection transformer (RT-DETR) having a ResNet18 backbone with selective kernel blocks and the hybrid encoder module. Step 2 uses a small labeled augmented dataset with maturity labels to fine-tune the Cit-DETR model for maturity detection. Experimental results on a custom UAV-captured citrus dataset demonstrate the effectiveness of our method, achieving 86.2% average precision in citrus detection and 91.0% mean average precision in ripeness detection. The model has been further optimized for real-time inference on edge devices or UAVs, enabling precision agriculture applications.

Off-policy learning methods seek to derive an optimal policy directly from a fixed dataset of prior interactions. This objective presents significant challenges, primarily due to the inherent distributional shift and value function overestimation bias. These issues become even more noticeable in reinforcement learning, where an agent trained on reward-free data must adapt to new tasks at test time without additional training. In this work, we address the off-policy problem in a zero-shot setting by discovering a theoretical connection of successor measures to stationary density ratios. Using this insight, our algorithm can infer optimal importance sampling ratios, effectively performing a stationary distribution correction with an optimal policy . We benchmark our method in motion tracking tasks on SMPL Humanoid, continuous control on ExoRL, and for the long-horizon OGBench tasks. Our technique seamlessly integrates into forward-backward representation frameworks and enables to new tasks in a regime. More broadly, this work bridges off-policy learning and zero-shot adaptation, offering benefits to both research areas.

Foundation models (FMs), powered by self-supervised learning (SSL), have redefined the capabilities of artificial intelligence, demonstrating exceptional performance in domains like natural language processing and computer vision. These advances present a transformative opportunity for brain signal analysis. Unlike traditional supervised learning, which is limited by the scarcity of labeled neural data, SSL offers a promising solution by enabling models to learn meaningful representations from unlabeled data. This is particularly valuable in addressing the unique challenges of brain signals, including high noise levels, inter-subject variability, and low signal-to-noise ratios. This survey systematically reviews the emerging field of bridging brain signals with foundation models through the innovative application of SSL. It explores key SSL techniques, the development of brain-specific foundation models, their adaptation to downstream tasks, and the integration of brain signals with other modalities in multimodal SSL frameworks. The review also covers commonly used evaluation metrics and benchmark datasets that support comparative analysis. Finally, it highlights key challenges and outlines future research directions. This work aims to provide researchers with a structured understanding of this rapidly evolving field and a roadmap for developing generalizable brain foundation models powered by self-supervision.

Automatic emotion recognition has attracted significant interest in healthcare, thanks to remarkable developments made recently in smart and innovative technologies. A real-time emotion recognition system allows for continuous monitoring, comprehension, and enhancement of the physical entity's capacities, along with continuing advice for enhancing quality of life and well-being in the context of personalized healthcare. Multimodal emotion recognition presents a significant challenge in terms of efficiently using the diverse modalities present in the data. In this article, we introduce a Multi-Stage Fusion Network (MSF-Net) for emotion recognition capable of extracting multimodal information and achieving significant performances. We propose utilizing the transformer-based structure to extract deep features from facial expressions. We exploited two visual descriptors, local binary pattern and Oriented FAST and Rotated BRIEF, to retrieve the computer vision- based features from the facial videos. A feature-level fusion network integrates the extraction of features from these modules, directing the output into the triplet attention technique. This module employs a three-branch architecture to compute attention weights to capture cross-dimensional interactions efficiently. The temporal dependencies in physiological signals are modeled by a Bi-directional Gated Recurrent Unit (Bi-GRU) in forward and backward directions at each time step. Lastly, the output feature representations from the triplet attention module and the extracted high-level patterns from Bi-GRU are fused and fed into the classification module to recognize emotion. The extensive experimental evaluations revealed that the proposed MSF-Net outperformed the state-of-the-art approaches on two popular datasets, BioVid Emo DB and MGEED. Finally, we tested the proposed MSF-Net in the Internet of Things environment to facilitate real-world scalable smart healthcare application.

Automatically identifying violence in videos is critical, and combining visual and audio cues is often the most effective approach that provides complementary information for violence detection. However, existing research on fusing these cues is computationally demanding and limited. To address this issue, we propose a novel fused vision-based graph neural network (FV-GNN) for violence detection using audiovisual information. This approach combines local and global features from both audio and video, leveraging a residual learning strategy to extract the most informative cues. Furthermore, FV-GNN utilizes dynamic graph filtering to analyze the inherent relationships between audio and video samples, enhancing violence recognition. The network consists of three branches: integrated, specialized, and scoring. The integrated branch captures long-range dependencies based on similarity, while the specialized branch focuses on local positional relationships. Finally, the scoring branch assesses the predicted violence likelihood against reality. We extensively explored the use of graphs for modeling temporal context in videos and found FV-GNN to be particularly well-suited for real-time violence detection. Our experiments demonstrate that FV-GNN outperforms current state-of-the-art methods on the XD-Violence datasets.

As generative AI continues to evolve, Vision Language Models (VLMs) have emerged as promising tools in various healthcare applications. One area that remains relatively underexplored is their use in human activity recognition (HAR) for remote health monitoring. VLMs offer notable strengths, including greater flexibility and the ability to overcome some of the constraints of traditional deep learning models. However, a key challenge in applying VLMs to HAR lies in the difficulty of evaluating their dynamic and often non-deterministic outputs. To address this gap, we introduce a descriptive caption data set and propose comprehensive evaluation methods to evaluate VLMs in HAR. Through comparative experiments with state-of-the-art deep learning models, our findings demonstrate that VLMs achieve comparable performance and, in some cases, even surpass conventional approaches in terms of accuracy. This work contributes a strong benchmark and opens new possibilities for the integration of VLMs into intelligent healthcare systems. Code and dataset are available at: https://github.com/gouga10/VLMs-HAR-RHMS.git .

Modern continuous-time generative models typically induce V-shaped flows: each sample travels independently along a nearly straight trajectory from the prior to the data. Although effective, this independent movement overlooks the hierarchical structures that exist in real-world data. To address this, we introduce Y-shaped generative flows, a framework in which samples travel together along shared pathways before branching off to target-specific endpoints. Our formulation is theoretically justified, yet remains practical, requiring only minimal modifications to standard velocity-driven models. We implement this through a scalable, neural network-based training objective. Experiments on synthetic, image, and biological datasets demonstrate that our method recovers hierarchy-aware structures, improves distributional metrics over strong flow-based baselines, and reaches targets in fewer steps.

Offline reinforcement learning (RL) learns exclusively from static datasets, without further interaction with the environment. In practice, such datasets vary widely in quality, often mixing expert, suboptimal, and even random trajectories. The choice of algorithm therefore depends on dataset fidelity. Behavior cloning can suffice on high-quality data, whereas mixed- or low-quality data typically benefits from offline RL methods that stitch useful behavior across trajectories. Yet in the wild it is difficult to assess dataset quality a priori because the data's provenance and skill composition are unknown. We address the problem of estimating offline dataset quality without training an agent. We study a spectrum of proxies from simple cumulative rewards to learned value based estimators, and introduce the Bellman Wasserstein distance (BWD), a value aware optimal transport score that measures how dissimilar a dataset's behavioral policy is from a random reference policy. BWD is computed from a behavioral critic and a state conditional OT formulation, requiring no environment interaction or full policy optimization. Across D4RL MuJoCo tasks, BWD strongly correlates with an oracle performance score that aggregates multiple offline RL algorithms, enabling efficient prediction of how well standard agents will perform on a given dataset. Beyond prediction, integrating BWD as a regularizer during policy optimization explicitly pushes the learned policy away from random behavior and improves returns. These results indicate that value aware, distributional signals such as BWD are practical tools for triaging offline RL datasets and policy optimization.

This study investigates the machine unlearning techniques within the context of large language models (LLMs), referred to as LLM unlearning. LLM unlearning offers a principled approach to removing the influence of undesirable data (e.g., sensitive or illegal information) from LLMs, while preserving their overall utility without requiring full retraining. Despite growing research interest, there is no comprehensive survey that systematically organizes existing work and distills key insights; here, we aim to bridge this gap. We begin by introducing the definition and the paradigms of LLM unlearning, followed by a comprehensive taxonomy of existing unlearning studies. Next, we categorize current unlearning approaches, summarizing their strengths and limitations. Additionally, we review evaluation metrics and benchmarks, providing a structured overview of current assessment methodologies. Finally, we outline promising directions for future research, highlighting key challenges and opportunities in the field.

Recently, small models with latent recursion have obtained promising results on complex reasoning tasks. These results are typically explained by the theory that such recursion increases a networks depth, allowing it to compactly emulate the capacity of larger models. However, the performance of recursively added layers remains behind the capabilities of one pass models with the same feed-forward depth. This means that in the looped version, not every recursive step effectively contributes to depth. This raises the question: when and why does latent reasoning improve performance, and when does it result in dead compute? In our work, we demonstrate that latent recursive reasoning provides answer to this question. We show that latent recursive reasoning can be formalized as a policy improvement algorithm. Building on these insights, we propose to use a training schemes from reinforcement learning and diffusion methods for latent reasoning models. Using the Tiny Recursive Model as our testbed, we show that with our modifications we can avoid dead compute steps and reduce the total number of forward passes by 18x while maintaining performance. Broadly speaking, we show how a policy improvement perspective on recursive steps can explain model behavior and provide insights for further improvements.

Advancements in face recognition (FR) technologies have amplified privacy concerns, necessitating methods that protect identity while maintaining recognition utility. Existing face anonymization methods typically focus on obscuring identity but fail to meet the requirements of biometric template protection, including revocability, unlinkability, and irreversibility. We propose FaceAnonyMixer, a cancelable face generation framework that leverages the latent space of a pre-trained generative model to synthesize privacy-preserving face images. The core idea of FaceAnonyMixer is to irreversibly mix the latent code of a real face image with a synthetic code derived from a revocable key. The mixed latent code is further refined through a carefully designed multi-objective loss to satisfy all cancelable biometric requirements. FaceAnonyMixer is capable of generating high-quality cancelable faces that can be directly matched using existing FR systems without requiring any modifications. Extensive experiments on benchmark datasets demonstrate that FaceAnonyMixer delivers superior recognition accuracy while providing significantly stronger privacy protection, achieving over an 11

The field of Multimodal Emotion Recognition (MER) has made considerable advancements in recent years; however, the opportunity to leverage the synergistic relationships between different modalities remains largely untapped. This paper introduces an MER approach employing a Joint Multi-Scale Multimodal Transformer (JMMT) with recursive cross-attention for naturalistic recognition of emotions by enhancing and capturing inter- and intra-modal relationships across both (visual and audio) modalities. We compute multi-scale attention weights based on cross-correlations between multi-scale joint representations of combined and individual cues to capture inter and intra-modal dynamics. As a result of individual modalities, recursive inputs are fed back during the fusion for further refinement of features. Our JMMT model presents a cost-effective solution for consumer devices by capturing synergistic characteristics across visual and audio inputs. The JMMT model outperforms the state-of-the-art (SOTA) methods in MER systems, which were evaluated by IEMOCAP and MELD datasets.

Unlearning methods for vision-language models (VLMs) have primarily adapted techniques from large language models (LLMs), relying on weight updates that demand extensive annotated forget sets. Moreover, these methods perform unlearning at a coarse granularity, often leading to excessive forgetting and reduced model utility. To address this issue, we introduce SAUCE, a novel method that leverages sparse autoencoders (SAEs) for fine-grained and selective concept unlearning in VLMs. Briefly, SAUCE first trains SAEs to capture high-dimensional, semantically rich sparse features. It then identifies the features most relevant to the target concept for unlearning. During inference, it selectively modifies these features to suppress specific concepts while preserving unrelated information. We evaluate SAUCE on two distinct VLMs, LLaVA-v1.5-7B and LLaMA-3.2-11B-Vision-Instruct, across two types of tasks: concrete concept unlearning (objects and sports scenes) and abstract concept unlearning (emotions, colors, and materials), encompassing a total of 60 concepts. Extensive experiments demonstrate that SAUCE outperforms state-of-the-art methods by 18.04 comparable model utility. Furthermore, we investigate SAUCE's robustness against widely used adversarial attacks, its transferability across models, and its scalability in handling multiple simultaneous unlearning requests. Our findings establish SAUCE as an effective and scalable solution for selective concept unlearning in VLMs.