A new study from Carnegie Mellon University finds that AI vision systems like CLIP often generate image captions based on stereotypes associated with the image's label rather than by actually analyzin…
Researchers propose Group-Contrastive Forward-Forward (GCFF), a biologically plausible learning algorithm that yields monosemantic neurons organized into hierarchies of increasing abstraction without …
Researchers show that the frozen predictor from Joint-Embedding Predictive Architectures (JEPAs) can be transferred across encoder families to improve occluded feature completion. In experiments, pair…
A new open-source project demonstrates how multimodal vector search works under the hood, using OpenAI CLIP (ViT-B/32) to embed text and images into a 512-dimensional vector space and Qdrant (HNSW ind…
A new study reveals that intermediate representations from models like CLIP and DINOv2, combined with semantically aligned references, significantly improve synthetic image attribution accuracy. The r…
A new study from arXiv (2607.12052v1) finds that training-free synthetic image attribution accuracy peaks at intermediate representation levels of CLIP and DINOv2, with reference selection methods cri…
Researchers have developed a parameter-efficient AI framework based on CLIP models that enhances animal re-identification by adapting to visual shifts over time, using continuous metadata conditioning…
Researchers have identified three core principles—divisibility, transferability, and stability—that govern compositional generalization in AI models, revealing that linear structures in neural represe…
Researchers introduced SFT+RL, a two-stage framework combining supervised fine-tuning and reinforcement learning on CLIP's encoder, achieving a 10.2% improvement in clean accuracy and 15.8% in adversa…
Researchers developed a multilingual, multimodal NLP framework for early detection of fake news and violence-driven mob activity, achieving 98% accuracy on a fused dataset of 138,256 Bangla and Englis…
A new decision framework for cross-habitat marine species recognition shows that pairing a frozen DINOv2 foundation model with a linear classifier and just 10-20 labeled images per species achieves re…
Researchers propose RSGPNet, a training-free geometric prompting framework for open-vocabulary semantic segmentation in remote sensing, which improves segmentation accuracy by leveraging object geomet…
Researchers introduced HANCLIP, a family of vision-language models that restructure the embedding space to handle negation by combining hyperbolic and angular objectives. Trained on 20,000 image-text …
Researchers propose TheProfessor, a multi-teacher prompt distillation method for vision-language models that ensembles a domain-finetuned teacher and a zero-shot teacher. On four datasets, confidence-…
Researchers at ACL 2026 revealed PriceBlind, a visual adversarial attack that exploits a vulnerability called Visual Dominance Hallucination in multimodal large language models used as financial agent…
FocusOnFun team's Vortex system achieved 90.5% in the Preliminary Round and 'Excellent' overall performance with 'Outstanding' QA results at the Ho Chi Minh City AI Challenge 2025. The multimodal vide…
Noise Contrastive Estimation (NCE) and InfoNCE are machine learning methods for estimating statistical distributions by distinguishing real data from noise, with applications in language modeling, spe…
Researchers have developed L-VARC, a novel framework that enhances visual reasoning on the Abstraction and Reasoning Corpus (ARC) by integrating a language-guided Learning Using Privileged Information…
Researchers have developed Bayesian Conditional Priors (BCP) Estimation, a gradient-free test-time adaptation method that improves multi-label recognition with frozen Vision-Language Models under dist…
NVIDIA released a workflow for converting FP8-quantized CLIP model checkpoints into TensorRT engines, enabling faster inference and higher GPU throughput for production deployment. The process involve…