A new fully unsupervised method, VMTAD, uses transformer architecture and a memory module to detect obstacles in dynamic agricultural scenes in real-time. It achieves state-of-the-art performance on a rapeseed dataset with 0.973 detection and 0.997 segmentation AUC, and a lightweight variant runs in 14 ms.

A study audits video, image, and audio deepfake benchmarks using linear probes on frozen self-supervised representations, finding that general-purpose representations can approach bespoke detector performance, suggesting benchmarks may reward general modality understanding rather than forensic skills.

This paper proposes a differentiable architecture search method to automatically discover the optimal fusion scheme for combining image and prompt tokens in visual prompt tuning. The approach jointly optimizes learnable prompts and their fusion mechanisms, introducing affine transformation and cross-attention as new fusion schemes. Experiments on 34 datasets demonstrate consistent improvements over baselines, revealing that hybrid fusion better leverages layer semantics in Vision Transformers.

Researchers introduce the Turbid Underwater Baseline (TUB) dataset and a new metric, PCD, to quantify information loss in underwater scenes with extreme turbidity. PCD correlates strongly with instance segmentation performance, outperforming common metrics.

GeMoE models token routing as a Minimum Description Length problem, using gating entropy to adaptively select experts, achieving 99.5% performance retention while increasing expert activation sparsity by 36.5%.

This study extends fMRI cognitive taskonomy from single-source to multi-source transfer across 23 Human Connectome Project task states, using Boolean Integer Programming (BIP) for budget-constrained task allocation. Training 1,127 models reveals directional, paradigm-structured single-source transfer and composition-dependent multi-source transfer. BIP prioritizes working-memory states (0-back and 2-back) under budget constraints, reflecting integrated perceptual, attentional, and executive processes. Findings highlight a cross-paradigm-limited motor cluster and high-priority working-memory states.

This paper introduces an innovative multi-task deep learning model that accurately predicts penetration state, depth, and weld seam morphology in laser penetration welding. The model uses weld pool images from a CMOS camera and welding parameters, integrating spatiotemporal features via CNNs and state space models. Test results show 99.35% accuracy for penetration state, 1.79 mm error for depth, and 95.65% accuracy for weld cross-section reconstruction.

Researchers developed a self-supervised framework using airborne LiDAR and optical imagery to estimate tree-level above-ground biomass in urban areas. The method achieved high accuracy in crown delineation and biomass estimation, revealing urban carbon stocks and changes over time without manual annotation.

This paper proposes LCG, a framework for long-context multi-image generation using Sparse Relational Attention (SRA) and Routing Consistency Constraint (RCC), along with a large synthetic dataset LCCD. Experiments show LCG outperforms baselines in prompt alignment and character consistency.

A hybrid approach combining image processing and deep learning assesses fruit freshness. An algorithm quantifies spoilage (0-100), and a CNN performs binary classification. Logistic regression fuses both outputs, later enabling the image processing algorithm to classify without CNN. Achieved >90% accuracy on apples and oranges with real-time performance and low computational requirements. Limitation requires isolated fruit on white/transparent background.

DocArena is a fully automated data curation pipeline that uses multimodal large language models (MLLMs) to transform raw documents into controllable, scalable training environments for document search agents. It requires no human annotation, generates reasoning-intensive QA pairs, and produces the DocArena-79K dataset spanning 8,336 documents across 16 domains and 49 languages. Experiments show that agents trained on DocArena achieve state-of-the-art performance on both retrieval accuracy and QA quality.

A systematic methodology transforms structured linguistic resources like Hindi WordNet into 1.25 million instruction-response pairs to fine-tune a 12B-parameter language model using resource-efficient LoRA and 4-bit quantization. Evaluation via a Hindi language learning chatbot shows superior pedagogical effectiveness (91.0) compared to general-purpose models (79.4-83.6) while maintaining competitive semantic performance. This work offers a practical alternative for low-resource languages, enabling specialized AI development for hundreds of languages with existing WordNet resources.

A new study reveals that larger language models outperform smaller ones in reasoning tasks due to constraint-guided reasoning. Using the AdvCluster framework, researchers found stable performance gaps of 6.43% and 7.38% across model pairs. The analysis identifies constraint identification and structured reasoning as key advantages.

This pilot study presents NEST-V1, a lightweight Transformer-based multimodal framework that generates emotion-conditioned Nepali Sign Language avatars from spoken input. On a dataset of 600 audio samples covering 4 common words and 3 emotional states, the system achieves 81.1% ASR accuracy and 79.21% emotion recognition accuracy with only 22.1M parameters, suitable for edge deployment. The work establishes a technical foundation for emotion-aware sign language translation in low-resource settings.

This study investigates using Nonviolent Communication (NVC) principles as lightweight prompt-level constraints to guide large language models (LLMs) toward more de-escalating dialogue behavior in emotionally charged situations. Through a dual-agent simulation framework across multiple models and user resistance levels, NVC-constrained prompting consistently reduces conversational escalation and stabilizes interactions with highly resistant users.

Large language models excel at short-context reasoning but degrade over long conversations due to context window limits. ContextForge recycles context via structured query generation, external memory retrieval, and controlled synthesis, reducing token overhead while maintaining answer quality. On a 15-turn healthcare benchmark, ContextForge improved consistency and reduced token consumption.

A new study finds that using linguistic features such as assertive certainty, explicit moral vocabulary, and emotion words in fine-tuning data significantly shifts LLM reasoning toward stronger pro-animal-welfare stances, while hedged language and concrete sensory description dilute that stance. The research offers practical guidance for animal-welfare advocates.

A new study evaluates LLM performance on statics problems using a model distillation approach. LLMs perform well on text-only problems but accuracy drops when diagrams and multi-step reasoning are introduced. The decline is primarily due to difficulties in multi-step reasoning, not image recognition limits.

A study finds that post-training for helpfulness (SFT and RL) significantly degrades animal compassion values instilled during mid-training, while coding post-training better preserves them. Helpfulness training also causes a large drop in English general moral reasoning but not cross-lingually, whereas the compassion degradation transfers consistently across languages. This suggests mid-trained values are encoded more deeply and cross-lingually than reasoning improvements from domain-specific post-training. The paper recommends coding post-training for value-preserving model development.

A new benchmark, Know2Guess, aims to evaluate LLMs' ability to distinguish between knowledge-based answering and guessing, considering data contamination. It includes 1,200 items across five domains and tests models like FLAN-T5, Qwen2.5-Instruct, and Llama-3-Instruct. Qwen2.5-3B-Instruct shows best reliability but still has calibration issues.

HierBias is a novel hierarchical context-conditioned media bias detector that formally models document context for bias prediction, theoretically proving reduced Bayes error and improving sample efficiency via multi-task learning. It achieves 0.853 F1 and 0.723 MCC on BABE and BASIL, surpassing state-of-the-art.

This research proposes a neural network method incorporating topological data analysis to improve flood detection accuracy and interpretability. Using the SEN12-FLOOD dataset, topological features extracted from images are integrated into neural networks, demonstrating that topological descriptors carry meaningful flood signals independently and enhance the robustness and interpretability of existing networks.

Modern machine learning systems have evolved into complex socio-technical architectures that actively mediate human opportunity. The field of algorithmic fairness addresses how models optimized for predictive accuracy can systematically disadvantage marginalized groups. This thesis (arXiv:2606.26200) identifies two fundamental limitations: reliance on deterministic point estimates for auditing and treatment of individuals as isolated entities devoid of structural context.

A new FHPLF model integrates hash learning with federated learning, using binary gradient matrices, projected Hamming distance, and a privacy-enhanced upload strategy to improve accuracy, efficiency, and privacy.

This paper introduces Clue-Guided Group Discovery (CGGD), a paradigm that progressively recovers money laundering groups from initial clues through analyst interaction. The proposed Clue2Group framework constructs a compact local investigation context, uses a multi-semantic local-temporal GNN to estimate risk fields, and integrates evidence to recover criminal groups. Experiments on large-scale AML benchmarks demonstrate its practical potential for real-world investigations.

This paper challenges the assumption that setting LLM-as-judge temperature to 0 ensures deterministic grading, showing flips occur due to default temperature and residual non-reproducibility even under greedy decoding.

KG-TRACE is a novel neuro-symbolic framework that integrates the WHO mutation knowledge graph as a structured biological constraint into a neural genomic model, dynamically weighting neural evidence against symbolic biological knowledge via a learned epistemic trust gate. Evaluated on the CRyPTIC M. tuberculosis cohort, it achieves an AUROC of 0.9760 for isoniazid and introduces the Biological Grounding Ratio (BGR) metric to quantify alignment between neural attributions and established biology.

This paper provides a comprehensive review of Neural Architecture Search (NAS) methods applied to Generative Adversarial Networks (GANs), comparing search strategies, evaluation metrics, and performance outcomes. It highlights the superiority of evolutionary and gradient-based methods in certain contexts, the need for robust evaluation metrics beyond IS and FID, and the importance of diverse datasets.

This study reframes phototaxis in unicellular algae as an information-driven sensorimotor process, linking a POMDP with biochemical reaction dynamics via CRN-ODEs. Using inverse reinforcement learning on 30 Chlamydomonas trajectories, it infers behavioral objectives and shows that run–tumble alternation emerges as an information-acquisition strategy, demonstrating how intracellular biochemical networks support adaptive information-seeking behavior.

A GPU-native population optimizer, χ-sao (Convergence-Halt-Invert-Stick-And-Oscillate), exploits a deliberate convergence-anticonvergence oscillation cycle to escape local traps while freezing confirmed modes. On all 42 functions of the Simon Fraser University optimization benchmark suite across dimensions d ∈ {2,4,8,16,32,64}, χ-sao achieves 100% mode recovery where all CPU baselines collapse at d ≥ 8 on the hardest multimodal functions, with speedups up to 34× over basin-hopping on Michalewicz d=64 and up to 39× on Rotated Hyper-Ellipsoid d=64. Under substantial likelihood noise (σ_noise up to 1.0), mode detection remains 100% reliable. The algorithm is available as an open-source Python package on PyPI.