arXiv:2510.21412v2 Announce Type: replace 
Abstract: Human intelligence effortlessly interprets visual scenes along a rich spectrum of semantic dimensions. However, existing approaches to language-grounded visual concept learning are limited to a few predefined primitive axes, such as color and shape, and are typically explored in synthetic datasets. In this work, we propose a scalable framework that adaptively identifies image-related concept axes and grounds visual concepts along these axes in real-world scenes. Leveraging a pretrained vision-language model and our universal prompting strategy, our framework identifies a diverse image-related axes without any prior knowledge. Our universal concept encoder adaptively binds visual features to the discovered axes without introducing additional model parameters for each concept. To ground visual concepts along the discovered axes, we optimize a compositional anchoring objective, which ensures that each axis can be independently manipulated without affecting others. We demonstrate the effectiveness of our framework on subsets of ImageNet, CelebA-HQ, and AFHQ, showcasing superior editing capabilities across diverse real-world concepts that are too varied to be manually predefined. Our method also exhibits strong compositional generalization, outperforming existing visual concept learning and text-based editing methods. The code is available at https://github.com/whieya/Language-grounded-VCL.

تم اقتراح إطار جديد لتعلم المفاهيم البصرية المرتبطة باللغة، والذي يهدف إلى تحسين كيفية تفسير الآلات للمشاهد البصرية. على عكس الطرق التقليدية التي تعتمد على محاور محدودة مثل اللون والشكل، تتكيف هذه الطريقة المبتكرة لتحديد محاور المفاهيم ذات الصلة في الصور. هذه الخطوة مهمة لأنها قد تؤدي إلى أنظمة ذكاء اصطناعي أكثر تطوراً قادرة على فهم والتفاعل مع العالم بطريقة تشبه البشر.

Se ha propuesto un nuevo marco para el aprendizaje de conceptos visuales anclados en el lenguaje, que busca mejorar la forma en que las máquinas interpretan escenas visuales. A diferencia de los métodos tradicionales que dependen de ejes limitados como el color y la forma, este enfoque innovador se adapta para identificar ejes de conceptos relevantes en las imágenes. Este avance es significativo, ya que podría llevar a sistemas de IA más sofisticados capaces de comprender e interactuar con el mundo de una manera más similar a la humana.

Un nouveau cadre pour l'apprentissage des concepts visuels ancrés dans le langage a été proposé, visant à améliorer la façon dont les machines interprètent les scènes visuelles. Contrairement aux méthodes traditionnelles qui s'appuient sur des axes limités comme la couleur et la forme, cette approche innovante s'adapte pour identifier les axes de concepts pertinents dans les images. Cette avancée est significative car elle pourrait conduire à des systèmes d'IA plus sophistiqués capables de comprendre et d'interagir avec le monde de manière plus humaine.

A new framework for language-grounded visual concept learning has been proposed, which aims to enhance how machines interpret visual scenes. Unlike traditional methods that rely on limited axes like color and shape, this innovative approach adapts to identify relevant concept axes in images. This advancement is significant as it could lead to more sophisticated AI systems capable of understanding and interacting with the world in a more human-like manner.

Bridging the gap to real-world language-grounded visual concept learning

arXiv:2512.06154v1 Announce Type: new 
Abstract: Learning invariant graph representations for out-of-distribution (OOD) generalization remains challenging because the learned representations often retain spurious components. To address this challenge, this work introduces a new tool from information theory called Partial Information Decomposition (PID) that goes beyond classical information-theoretic measures. We identify limitations in existing approaches for invariant representation learning that solely rely on classical information-theoretic measures, motivating the need to precisely focus on redundant information about the target $Y$ shared between spurious subgraphs $G_s$ and invariant subgraphs $G_c$ obtained via PID. Next, we propose a new multi-level optimization framework that we call -- Redundancy-guided Invariant Graph learning (RIG) -- that maximizes redundant information while isolating spurious and causal subgraphs, enabling OOD generalization under diverse distribution shifts. Our approach relies on alternating between estimating a lower bound of redundant information (which itself requires an optimization) and maximizing it along with additional objectives. Experiments on both synthetic and real-world graph datasets demonstrate the generalization capabilities of our proposed RIG framework.

تقدم دراسة جديدة إطارًا يسمى التعلم البياني invariant-guided بواسطة المعلومات الزائدة (RIG)، والذي يستخدم تحليل المعلومات الجزئية (PID) لتحسين التعميم خارج التوزيع (OOD) في تعلم تمثيلات الرسوم البيانية. تهدف هذه الطريقة إلى التخفيف من احتفاظ التمثيلات المتعلمة بالمكونات الزائفة من خلال تعظيم المعلومات الزائدة مع عزل المجموعات السببية.

Un nuevo estudio presenta un marco denominado aprendizaje gráfico invariante guiado por la redundancia (RIG), que utiliza la descomposición de información parcial (PID) para mejorar la generalización fuera de distribución (OOD) en el aprendizaje de representaciones gráficas. Este enfoque busca mitigar la retención de componentes espurios en las representaciones aprendidas al maximizar la información redundante mientras aísla los subgrupos causales.

Une nouvelle étude présente un cadre appelé apprentissage graphique invariant guidé par la redondance (RIG), qui utilise la décomposition d'information partielle (PID) pour améliorer la généralisation hors distribution (OOD) dans l'apprentissage des représentations graphiques. Cette approche vise à atténuer la rétention de composants fallacieux dans les représentations apprises en maximisant l'information redondante tout en isolant les sous-graphes causaux.

A new study introduces a framework called Redundancy-guided Invariant Graph learning (RIG), which utilizes Partial Information Decomposition (PID) to enhance out-of-distribution (OOD) generalization in graph representation learning. This approach aims to mitigate the retention of spurious components in learned representations by maximizing redundant information while isolating causal subgraphs.

Bridging the gap to real-world language-grounded visual concept learning

Was this article worth reading? Share it

LucidQuery AI

The Visualizer

Attentive AI