arXiv:2511.16147v1 Announce Type: new 
Abstract: In the field of large models (LMs) for natural language processing (NLP) and computer vision (CV), Parameter-Efficient Fine-Tuning (PEFT) has emerged as a resource-efficient method that modifies a limited number of parameters while keeping the pretrained weights fixed. This paper investigates the traditional PEFT approach, which applies modifications to all position indices, and questions its necessity. We introduce a new paradigm called Token-Selective PEFT (TS-PEFT), in which a function S selectively applies PEFT modifications to a subset of position indices, potentially enhancing performance on downstream tasks. Our experimental results reveal that the indiscriminate application of PEFT to all indices is not only superfluous, but may also be counterproductive. This study offers a fresh perspective on PEFT, advocating for a more targeted approach to modifications and providing a framework for future research to optimize the fine-tuning process for large models.

تقدم الورقة مفهوم الضبط الدقيق الفعال من حيث المعلمات الانتقائية للرموز (TS-PEFT)، وهو نهج جديد في معالجة اللغة الطبيعية ورؤية الكمبيوتر يقوم بتطبيق التعديلات بشكل انتقائي على مجموعة فرعية من مؤشرات المواقع. يتحدى هذا الأسلوب النهج التقليدي لضبط المعلمات الفعال (PEFT) الذي يعدل جميع المؤشرات بشكل عشوائي. تشير النتائج التجريبية إلى أن التطبيق المستهدف لـ TS-PEFT يمكن أن يعزز الأداء في المهام اللاحقة، مما يقترح تحولًا نحو استراتيجيات ضبط أكثر كفاءة.

El artículo presenta el Token-Selective Parameter-Efficient Fine-Tuning (TS-PEFT), un nuevo enfoque en el procesamiento del lenguaje natural y la visión por computadora que aplica modificaciones de manera selectiva a un subconjunto de índices de posición. Este método cuestiona el enfoque tradicional de Parameter-Efficient Fine-Tuning (PEFT), que modifica indiscriminadamente todos los índices. Los resultados experimentales indican que la aplicación dirigida de TS-PEFT puede mejorar el rendimiento en tareas posteriores, sugiriendo un cambio hacia estrategias de ajuste más eficientes.

Cet article présente le Token-Selective Parameter-Efficient Fine-Tuning (TS-PEFT), une nouvelle approche en traitement du langage naturel et vision par ordinateur qui applique sélectivement des modifications à un sous-ensemble d'indices de position. Cette méthode remet en question l'approche traditionnelle de Parameter-Efficient Fine-Tuning (PEFT), qui modifie indiscriminément tous les indices. Les résultats expérimentaux indiquent que l'application ciblée de TS-PEFT peut améliorer les performances sur les tâches en aval, suggérant un changement vers des stratégies de fine-tuning plus efficace…

The paper introduces Token-Selective Parameter-Efficient Fine-Tuning (TS-PEFT), a novel approach in natural language processing and computer vision that selectively applies modifications to a subset of position indices. This method challenges the traditional Parameter-Efficient Fine-Tuning (PEFT) approach, which indiscriminately modifies all indices. Experimental results indicate that the targeted application of TS-PEFT can enhance performance on downstream tasks, suggesting a shift towards more efficient fine-tuning strategies.

TS-PEFT: Token-Selective Parameter-Efficient Fine-Tuning with Learnable Threshold Gating

arXiv:2511.15875v1 Announce Type: new 
Abstract: The automated analysis of historical documents, particularly maps, has drastically benefited from advances in deep learning and its success across various computer vision applications. However, most deep learning-based methods heavily rely on large amounts of annotated training data, which are typically unavailable for historical maps, especially for those belonging to specific, homogeneous cartographic domains, also known as corpora. Creating high-quality training data suitable for machine learning often takes a significant amount of time and involves extensive manual effort. While synthetic training data can alleviate the scarcity of real-world samples, it often lacks the affinity (realism) and diversity (variation) necessary for effective learning. By transferring the cartographic style of an original historical map corpus onto vector data, we bootstrap an effectively unlimited number of synthetic historical maps suitable for tasks such as land-cover interpretation of a homogeneous historical map corpus. We propose an automatic deep generative approach and a alternative manual stochastic degradation technique to emulate the visual uncertainty and noise, also known as data-dependent uncertainty, commonly observed in historical map scans. To quantitatively evaluate the effectiveness and applicability of our approach, the generated training datasets were employed for domain-adaptive semantic segmentation on a homogeneous map corpus using a Self-Constructing Graph Convolutional Network, enabling a comprehensive assessment of the impact of our data bootstrapping methods.

لقد تحسنت التحليلات الآلية للخرائط التاريخية بشكل كبير بفضل التقدم في التعلم العميق، وخاصة في رؤية الكمبيوتر. ومع ذلك، فإن ندرة بيانات التدريب المعلّمة للخرائط التاريخية المحددة تمثل تحديًا. لمعالجة ذلك، تم اقتراح طريقة لتوليد خرائط تاريخية اصطناعية من خلال نقل الأسلوب الخرائطي لخرائط أصلية إلى بيانات متجهة، مما يتيح إنشاء عدد غير محدود من عينات التدريب لمهام التعلم الآلي.

El análisis automatizado de mapas históricos ha mejorado significativamente gracias a los avances en el aprendizaje profundo, especialmente en visión por computadora. Sin embargo, la escasez de datos de entrenamiento anotados para corpus de mapas históricos específicos representa un desafío. Para abordar esto, se ha propuesto un método para generar mapas históricos sintéticos transfiriendo el estilo cartográfico de mapas originales a datos vectoriales, lo que permite la creación de un número ilimitado de muestras de entrenamiento para tareas de aprendizaje automático.

L'analyse automatisée des cartes historiques a considérablement progressé grâce aux avancées de l'apprentissage profond, en particulier dans la vision par ordinateur. Cependant, la rareté des données d'entraînement annotées pour des corpus de cartes historiques spécifiques pose un défi. Pour y remédier, une méthode de génération de cartes historiques synthétiques en transférant le style cartographique de cartes originales sur des données vectorielles a été proposée, permettant la création d'un nombre illimité d'échantillons d'entraînement pour les tâches d'apprentissage automatique.

The automated analysis of historical maps has significantly improved due to advancements in deep learning, particularly in computer vision. However, the scarcity of annotated training data for specific historical map corpora poses a challenge. To address this, a method for generating synthetic historical maps by transferring the cartographic style of original maps onto vector data has been proposed, enabling the creation of an unlimited number of training samples for machine learning tasks.

Automatic Uncertainty-Aware Synthetic Data Bootstrapping for Historical Map Segmentation

arXiv:2511.16080v1 Announce Type: cross 
Abstract: Modern data processing workflows frequently encounter ragged data: collections with variable-length elements that arise naturally in domains like natural language processing, scientific measurements, and autonomous AI agents. Existing workflow engines lack native support for tracking the shapes and dependencies inherent to ragged data, forcing users to manage complex indexing and dependency bookkeeping manually. We present Operon, a Rust-based workflow engine that addresses these challenges through a novel formalism of named dimensions with explicit dependency relations. Operon provides a domain-specific language where users declare pipelines with dimension annotations that are statically verified for correctness, while the runtime system dynamically schedules tasks as data shapes are incrementally discovered during execution. We formalize the mathematical foundation for reasoning about partial shapes and prove that Operon's incremental construction algorithm guarantees deterministic and confluent execution in parallel settings. The system's explicit modeling of partially-known states enables robust persistence and recovery mechanisms, while its per-task multi-queue architecture achieves efficient parallelism across heterogeneous task types. Empirical evaluation demonstrates that Operon outperforms an existing workflow engine with 14.94x baseline overhead reduction while maintaining near-linear end-to-end output rates as workloads scale, making it particularly suitable for large-scale data generation pipelines in machine learning applications.

أوبيرون هو محرك سير عمل قائم على لغة راست مصمم للتعامل مع البيانات غير المنتظمة، والتي تتكون من مجموعات تحتوي على عناصر ذات أطوال متغيرة، وغالبًا ما توجد في مجالات مثل معالجة اللغة الطبيعية والقياسات العلمية. يقدم المحرك صيغة جديدة للأبعاد المسماة مع علاقات اعتماد واضحة، مما يسمح للمستخدمين بالإعلان عن خطوط أنابيب مع تعليقات توضيحية للأبعاد يتم التحقق من صحتها. يقوم أوبيرون بجدولة المهام ديناميكيًا مع اكتشاف أشكال البيانات أثناء التنفيذ.

Operon es un motor de flujo de trabajo basado en Rust diseñado para manejar datos irregulares, que consisten en colecciones con elementos de longitud variable comúnmente encontrados en campos como el procesamiento del lenguaje natural y las mediciones científicas. El motor introduce un formalismo de dimensiones nombradas con relaciones de dependencia explícitas, permitiendo a los usuarios declarar tuberías con anotaciones de dimensiones que se verifican por su corrección. Operon programa dinámicamente tareas a medida que se descubren formas de datos durante la ejecución.

Operon est un moteur de flux de travail basé sur Rust, conçu pour gérer les données irrégulières, qui se composent de collections avec des éléments de longueur variable, couramment rencontrées dans des domaines tels que le traitement du langage naturel et les mesures scientifiques. Le moteur introduit un formalisme de dimensions nommées avec des relations de dépendance explicites, permettant aux utilisateurs de déclarer des pipelines avec des annotations de dimensions vérifiées pour leur exactitude. Operon planifie dynamiquement les tâches à mesure que les formes de données sont découvertes pe…

Operon is a Rust-based workflow engine designed to handle ragged data, which consists of collections with variable-length elements commonly found in fields like natural language processing and scientific measurements. The engine introduces a formalism of named dimensions with explicit dependency relations, allowing users to declare pipelines with dimension annotations that are verified for correctness. Operon dynamically schedules tasks as data shapes are discovered during execution, addressing the limitations of existing workflow engines.

Operon: Incremental Construction of Ragged Data via Named Dimensions

arXiv:2511.15906v1 Announce Type: new 
Abstract: Generative models for structure-based drug design are often limited to a specific modality, restricting their broader applicability. To address this challenge, we introduce FuncBind, a framework based on computer vision to generate target-conditioned, all-atom molecules across atomic systems. FuncBind uses neural fields to represent molecules as continuous atomic densities and employs score-based generative models with modern architectures adapted from the computer vision literature. This modality-agnostic representation allows a single unified model to be trained on diverse atomic systems, from small to large molecules, and handle variable atom/residue counts, including non-canonical amino acids. FuncBind achieves competitive in silico performance in generating small molecules, macrocyclic peptides, and antibody complementarity-determining region loops, conditioned on target structures. FuncBind also generated in vitro novel antibody binders via de novo redesign of the complementarity-determining region H3 loop of two chosen co-crystal structures. As a final contribution, we introduce a new dataset and benchmark for structure-conditioned macrocyclic peptide generation. The code is available at https://github.com/prescient-design/funcbind.

FuncBind هو إطار جديد مصمم لتصميم الأدوية القائم على الهيكل، يستخدم الحقول العصبية لتوليد جزيئات تحتوي على جميع الذرات مشروطة بالأهداف. تتيح هذه الطريقة نموذجًا موحدًا قادرًا على التعامل مع أنظمة ذرية متنوعة، بما في ذلك الجزيئات الصغيرة والكبيرة، والأحماض الأمينية غير التقليدية. يحقق FuncBind أداءً تنافسيًا في توليد هياكل جزيئية متنوعة، بما في ذلك الجزيئات الصغيرة والببتيدات الدائرية، مشروطة بالهياكل المستهدفة.

FuncBind es un nuevo marco diseñado para el diseño de fármacos basado en estructuras que utiliza campos neuronales para generar moléculas de todos los átomos condicionadas por objetivos. Este enfoque permite un modelo unificado capaz de manejar sistemas atómicos diversos, incluidas moléculas pequeñas y grandes, así como aminoácidos no canónicos. FuncBind demuestra un rendimiento competitivo en la generación de diversas estructuras moleculares, incluidas pequeñas moléculas y péptidos macrocíclicos, condicionadas por estructuras objetivo.

FuncBind est un nouveau cadre conçu pour la conception de médicaments basée sur la structure, qui utilise des champs neuronaux pour générer des molécules toutes atomiques conditionnées par des cibles. Cette approche permet un modèle unifié capable de traiter des systèmes atomiques divers, y compris des molécules petites et grandes, ainsi que des acides aminés non canoniques. FuncBind démontre des performances compétitives dans la génération de diverses structures moléculaires, y compris des petites molécules et des peptides macrocycliques, conditionnées par des structures cibles.

FuncBind is a new framework designed for structure-based drug design that utilizes neural fields to generate target-conditioned, all-atom molecules. This approach allows for a unified model capable of handling diverse atomic systems, including small and large molecules, and non-canonical amino acids. FuncBind demonstrates competitive performance in generating various molecular structures, including small molecules and macrocyclic peptides, conditioned on target structures.

Unified all-atom molecule generation with neural fields

arXiv:2505.03201v3 Announce Type: replace 
Abstract: Integrated Gradients (IG) is a widely used attribution method in explainable AI, particularly in computer vision applications where reliable feature attribution is essential. A key limitation of IG is its sensitivity to the choice of baseline (reference) images. Multi-baseline extensions such as Expected Gradients (EG) assume uniform weighting over baselines, implicitly treating baseline images as equally informative. In high-dimensional vision models, this assumption often leads to noisy or unstable explanations. This paper proposes Weighted Integrated Gradients (WG), a principled approach that evaluates and weights baselines to enhance attribution reliability. WG introduces an unsupervised criterion for baseline suitability, enabling adaptive selection and weighting of baselines on a per-input basis. The method not only preserves core axiomatic properties of IG but also provides improved theoretical guarantees on the quality of explanation over EG. Experiments on commonly used image datasets and models show that WG consistently outperforms EG, yielding 10 to 35 percent improvements in attribution fidelity. WG further identifies informative baseline subsets, reducing unnecessary variability while maintaining high attribution accuracy. By moving beyond the idea that all baselines matter equally, Weighted Integrated Gradients offers a clearer and more reliable way to explain computer-vision models, improving both understanding and practical usability in explainable AI.

تقدم الورقة Weighted Integrated Gradients (WG)، وهي طريقة متقدمة لتخصيص الميزات في الذكاء الاصطناعي القابل للتفسير، خاصة في تطبيقات رؤية الكمبيوتر. تعالج WG القيود المفروضة على Integrated Gradients (IG) من خلال اختيار وتقييم الصور الأساسية بشكل تكيفي، مما يعزز موثوقية التخصيص. تحافظ هذه الطريقة على الخصائص الأساسية لـ IG بينما تحسن جودة الشروحات، مما يجعلها مساهمة مهمة في هذا المجال.

El artículo presenta los Weighted Integrated Gradients (WG), un método avanzado para la atribución de características en IA explicable, especialmente en aplicaciones de visión por computadora. Los WG abordan las limitaciones de los Integrated Gradients (IG) al seleccionar y ponderar de manera adaptativa las imágenes de referencia, mejorando así la fiabilidad de la atribución. Este método preserva las propiedades fundamentales de los IG mientras mejora la calidad de las explicaciones, lo que representa una contribución significativa al campo.

Cet article présente les Weighted Integrated Gradients (WG), une méthode avancée d'attribution des caractéristiques dans l'IA explicable, en particulier dans la vision par ordinateur. Les WG répondent aux limites des Integrated Gradients (IG) en sélectionnant et en pondérant de manière adaptative les images de référence, améliorant ainsi la fiabilité de l'attribution. Cette méthode préserve les propriétés fondamentales des IG tout en améliorant la qualité des explications, ce qui en fait une contribution significative au domaine.

The paper introduces Weighted Integrated Gradients (WG), an advanced method for feature attribution in explainable AI, particularly in computer vision. WG addresses the limitations of Integrated Gradients (IG) by adaptively selecting and weighting baseline images, improving attribution reliability. This method preserves the core properties of IG while enhancing the quality of explanations, making it a significant contribution to the field.

TS-PEFT: Token-Selective Parameter-Efficient Fine-Tuning with Learnable Threshold Gating

Was this article worth reading? Share it