arXiv:2605.27696v2 Announce Type: replace-cross 
Abstract: Discrete visual tokenizers translate images into ordered sequences of codes, providing a natural representation for structural description of scenes. Yet existing adaptive tokenizers either require post-hoc search or select among a discrete set of pre-trained rates, rather than learning a continuous per-image sequence length coupled to the model and scene, and they typically train against pixel reconstruction, emphasizing texture rather than structure. We propose STROP, a discrete visual tokenizer architecture that forms structural scene representations and simultaneously learns how long an image's visual program should be. Using a four-phase curriculum supervised by local rate--distortion probes against frozen DINOv3 features, STROP optimizes a dedicated length head that estimates the active prefix length in a single forward pass. By bypassing pixel-level reconstruction gradients, the codebook is shaped entirely by the quality of higher-level latent representations. Program length grows with scene complexity, and signs of compositional structure emerge both in downstream dense-prediction transfer and in direct inspection of the learned code vocabulary.

قدم الباحثون STROP، وهي بنية جديدة لمحول بصري منفصل تقوم بترجمة الصور إلى تسلسلات مرتبة من الرموز، مما يحسن طول البرامج البصرية للصور مع التركيز على التمثيل الهيكلي بدلاً من إعادة بناء البكسل. تستخدم هذه الطريقة منهجية من أربع مراحل لتحسين كفاءة النموذج في تقدير أطوال البادئات النشطة في عملية واحدة.

Los investigadores han presentado STROP, una nueva arquitectura de tokenizador visual discreto que traduce imágenes en secuencias ordenadas de códigos, optimizando la longitud de los programas visuales para imágenes mientras se centra en la representación estructural en lugar de la reconstrucción de píxeles. Este enfoque utiliza un currículo de cuatro fases para mejorar la eficiencia del modelo en la estimación de longitudes de prefijo activas en una sola pasada hacia adelante.

Des chercheurs ont introduit STROP, une nouvelle architecture de tokenizer visuel discret qui traduit les images en séquences ordonnées de codes, optimisant la longueur des programmes visuels pour les images tout en se concentrant sur la représentation structurelle plutôt que sur la reconstruction pixel par pixel. Cette approche utilise un curriculum en quatre phases pour améliorer l'efficacité du modèle dans l'estimation des longueurs de préfixes actifs en un seul passage avant.

Researchers have introduced STROP, a novel discrete visual tokenizer architecture that translates images into ordered sequences of codes, optimizing the length of visual programs for images while focusing on structural representation rather than pixel reconstruction. This approach utilizes a four-phase curriculum to enhance the model's efficiency in estimating active prefix lengths in a single forward pass.

Structure over Pixels: Learning Variable-Length Visual Programs

arXiv:2601.22108v2 Announce Type: replace 
Abstract: Continued pretraining is optimized with fixed self-supervised tasks but selected by downstream performance, creating a coarse feedback loop in which practitioners evaluate checkpoints, change data mixtures or objectives, and restart runs, while individual updates remain blind to target capabilities. We ask whether a small set of verifiable downstream examples can provide step-level feedback without directly supervising the learner. We introduce V-pretraining, which decouples a learner trained only with a self-supervised loss from a lightweight task designer that constructs targets or views for unlabeled batches. Given the current learner and batch, V-pretraining scores a candidate construction by predicting the first-order reduction in downstream loss after the induced self-supervised update. The designer maximizes this value; the learner then applies the update with targets or views detached, so downstream labels never update learner parameters. We instantiate V-pretraining as adaptive top-K soft targets for language modeling and learned views or masks for self-supervised vision. Across both modalities, V-pretraining improves target capabilities without degrading generalization. Under wall-clock-matched continued pretraining, it improves GSM8K Pass@1 for Qwen models using 1,024 GSM8K examples only as feedback, including a +7.4 point single-run gain for Qwen2.5-0.5B. In vision, it improves DINOv3 transfer to ADE20K semantic segmentation and NYUv2 depth estimation while preserving ImageNet linear accuracy, suggesting that feedback-guided task construction can improve target capabilities without collapsing general-purpose representations.

تم تقديم نهج جديد للتدريب المستمر، يسمى V-pretraining، والذي يفصل المتعلم عن مصمم المهام، مما يسمح بتغذية راجعة أكثر فعالية تعتمد على الأداء في المهام اللاحقة دون إشراف مباشر. تهدف هذه الطريقة إلى تحسين التعلم الذاتي الإشراف من خلال التنبؤ بانخفاض الخسارة في المهام اللاحقة بعد التحديثات.

Se ha introducido un nuevo enfoque para el preentrenamiento continuo, denominado V-preentrenamiento, que separa al aprendiz del diseñador de tareas, permitiendo una retroalimentación más efectiva basada en el rendimiento en tareas posteriores sin supervisión directa. Este método busca optimizar el aprendizaje auto-supervisado al predecir la reducción de la pérdida en tareas posteriores tras las actualizaciones.

Une nouvelle approche de préentraînement continu, appelée V-préentraînement, a été introduite, séparant l'apprenant du concepteur de tâches, permettant un retour d'information plus efficace basé sur la performance en aval sans supervision directe. Cette méthode vise à optimiser l'apprentissage auto-supervisé en prédisant la réduction de la perte en aval après les mises à jour.

A new approach to continued pretraining, termed V-pretraining, has been introduced, which separates the learner from the task designer, allowing for more effective feedback based on downstream performance without direct supervision. This method aims to optimize self-supervised learning by predicting the reduction in downstream loss following updates.

Structure over Pixels: Learning Variable-Length Visual Programs

Was this article worth reading? Share it

Ready to build your own newsroom?