arXiv:2605.28111v1 Announce Type: new 
Abstract: Predicting how a cell will change its transcriptional state under a developmental signal or a genetic perturbation is the computational core of in-silico biology and the AI Virtual Cell program. Existing approaches either fit static control-to-treated maps that discard time, or solve multi-step ODE / Schr\"odinger-bridge problems on each dataset independently. We introduce Chreode, a one-step cell world model that predicts action-conditioned cell-state transitions through a structured residual transition operator. It shifts distributional evolution from inference time to training time, enabling single-pass generation while preserving a Waddington-inspired decomposition into downhill landscape flow, rotational in-tangent dynamics, and stochastic spread. The model is pretrained with a shared scVI encoder and a DiT-based dynamics backbone on a 2.4M-cell mouse embryonic atlas spanning 7 datasets. As a fine-tuning initialization, Chreode improves per-target Sinkhorn distance on Weinreb hematopoiesis and Veres islet differentiation over matched scratch models, PI-SDE, and PRESCIENT. As a transferable gene-state embedding for GEARS, the pretrained dynamics representation reduces shared-vocabulary DE20 mean squared error on Norman Perturb-seq from 0.2121 to 0.1858, a 12.4% relative improvement, without changing the GEARS training procedure. We interpret this transfer to perturbation prediction as evidence that pretrained developmental-trajectory dynamics encode differentiation primitives transferable to CRISPR-induced state shifts, since both involve cell-state transitions in a shared latent geometry. The pretrained backbone additionally produces zero-shot clonal fate scores on Weinreb that are competitive with strong dynamic-OT baselines.

تم تقديم Chreode كنموذج جديد لعالم الخلايا من خطوة واحدة مصمم للتنبؤ بكيفية تغيير الخلايا لحالاتها النسخية استجابةً للإشارات التنموية أو الاضطرابات الجينية. يستخدم هذا النموذج مشغل انتقال متبقي منظم، مما يسمح بتوليد فعال في تمريرة واحدة مع الحفاظ على تدفق المناظر الطبيعية المستوحاة من وادينغتون.

Chreode se ha presentado como un nuevo modelo celular de un solo paso diseñado para predecir cómo las células cambian sus estados transcripcionales en respuesta a señales de desarrollo o perturbaciones genéticas. Este modelo utiliza un operador de transición residual estructurado, lo que permite una generación eficiente en una sola pasada mientras mantiene un flujo de paisaje inspirado en Waddington.

Chreode a été introduit comme un nouveau modèle cellulaire à une étape, conçu pour prédire comment les cellules modifient leurs états transcriptionnels en réponse à des signaux de développement ou à des perturbations génétiques. Ce modèle utilise un opérateur de transition résiduel structuré, permettant une génération efficace en un seul passage tout en maintenant un flux de paysage inspiré de Waddington.

Chreode has been introduced as a novel one-step cell world model designed to predict how cells transition their transcriptional states in response to developmental signals or genetic perturbations. This model utilizes a structured residual transition operator, allowing for efficient single-pass generation while maintaining a Waddington-inspired landscape flow.

Chreode: A Cell World Model for One-Step Temporal Dynamics and Perturbation Prediction

arXiv:2606.13578v1 Announce Type: new 
Abstract: Scientific laboratories increasingly rely on AI systems to reason about experiments, but the physical act of doing science remains largely outside their reach. AI can help read literature, generate hypotheses, and plan protocols, yet the execution of those protocols at the bench still requires a human operator. Vision-Language-Action (VLA) models provide one possible interface between written protocols and robot execution, but existing policies are trained mostly on household and tabletop demonstrations and rarely encounter the instruments, transparent liquids, or fixed protocol workflows found in scientific laboratories. Closing this gap requires both laboratory-specific supervision and a unified learning framework that can accommodate the diverse robot embodiments used to execute experimental protocols. We therefore identify data and embodiment as central bottlenecks alongside model design. To address the data side, we build RoboGenesis, a simulation-based workflow and data engine that composes configured laboratory workflows from atomic skills, validates and filters rollouts, and exports structured demonstrations across supported robot profiles. On the policy side, we present LabVLA, trained with a two-stage recipe: FAST action token pretraining first makes the Qwen3-VL-4B-Instruct backbone action aware before any continuous control is learned, and flow matching posttraining then attaches a DiT action expert under knowledge insulation. On the LabUtopia benchmark, LabVLA achieves the highest average success rate among all evaluated baselines under both in-distribution and out-of-distribution settings.

تهدف مبادرة LabVLA إلى سد الفجوة بين نماذج اللغة-الرؤية-العمل (VLA) وممارسات المختبرات العلمية، من خلال معالجة القيود التي تواجه أنظمة الذكاء الاصطناعي الحالية التي تكافح مع التنفيذ الفعلي للبروتوكولات العلمية. من خلال تطوير RoboGenesis، وهو إطار مصمم خصيصًا لبيئات المختبر، تسعى LabVLA إلى تحسين التفاعل بين البروتوكولات المكتوبة والتنفيذ الروبوتي.

La iniciativa LabVLA tiene como objetivo cerrar la brecha entre los modelos de Acción-Lenguaje-Visión (VLA) y las prácticas de laboratorio científico, abordando las limitaciones de los sistemas de IA actuales que luchan con la ejecución física de protocolos científicos. Al desarrollar RoboGenesis, un marco adaptado a entornos de laboratorio, LabVLA busca mejorar la interacción entre los protocolos escritos y la ejecución robótica.

L'initiative LabVLA vise à combler le fossé entre les modèles Vision-Language-Action (VLA) et les pratiques des laboratoires scientifiques, en abordant les limitations des systèmes d'IA actuels qui peinent à exécuter physiquement les protocoles scientifiques. En développant RoboGenesis, un cadre adapté aux environnements de laboratoire, LabVLA cherche à améliorer l'interaction entre les protocoles écrits et l'exécution robotique.

The LabVLA initiative aims to bridge the gap between Vision-Language-Action (VLA) models and scientific laboratory practices, addressing the limitations of current AI systems that struggle with the physical execution of scientific protocols. By developing RoboGenesis, a framework tailored for laboratory environments, LabVLA seeks to enhance the interaction between written protocols and robotic execution.

LabVLA: Grounding Vision-Language-Action Models in Scientific Laboratories

arXiv:2606.13676v1 Announce Type: new 
Abstract: Text-to-image (T2I) models contain rich spatial priors. Synthesizing photorealistic, cluttered scenes requires an understanding of geometry, including perspective and relative scale. Prior works adapt T2I models to leverage this prior for depth prediction, but they require dense depth data and involve complex recipes. We propose Modality Forcing, a simple, scalable post-training recipe for joint image-depth generation using a single DiT trained on sparse depth data. Modality Forcing enables conditional and joint generation of image and depth in any permutation by assigning separate noise levels per modality. Per-modality decoders let us train on sparse, real-world depth and achieve strong, generalizable depth prediction. We further show that Modality Forcing inherits the scalability of T2I pre-training: by training a set of T2I models from scratch (370M to 3.3B parameters), we find that larger models trained on more image data produce more accurate depth. Our strongest model is competitive with state-of-the-art monocular depth estimators and reduces AbsRel by 57% relative to existing joint image-depth generative models. These results provide strong evidence that image generation is a scalable pre-training objective for spatial perception. https://modality-forcing.github.io/

تم اقتراح نهج جديد يسمى Modality Forcing من أجل توليد الفضاء القابل للتوسع في نماذج النص إلى صورة (T2I)، مما يسمح بالتوليد المشترك للصورة والعمق باستخدام DiT واحد تم تدريبه على بيانات عمق نادرة. تبسط هذه الطريقة عملية توليد المشاهد الفوتوغرافية الواقعية من خلال تمكين التوليد الشرطي مع مستويات ضوضاء منفصلة لكل نمط.

Se ha propuesto un nuevo enfoque llamado Modality Forcing para la generación espacial escalable en modelos de texto a imagen (T2I), permitiendo la generación conjunta de imágenes y profundidad utilizando un solo DiT entrenado con datos de profundidad escasos. Este método simplifica el proceso de síntesis de escenas fotorealistas al habilitar la generación condicional con niveles de ruido separados para cada modalidad.

Une nouvelle approche appelée Modality Forcing a été proposée pour la génération spatiale évolutive dans les modèles de texte à image (T2I), permettant la génération conjointe d'images et de profondeur à l'aide d'un seul DiT formé sur des données de profondeur rares. Cette méthode simplifie le processus de synthèse de scènes photoréalistes en permettant une génération conditionnelle avec des niveaux de bruit séparés pour chaque modalité.

A new approach called Modality Forcing has been proposed for scalable spatial generation in text-to-image (T2I) models, allowing for joint image and depth generation using a single DiT trained on sparse depth data. This method simplifies the process of synthesizing photorealistic scenes by enabling conditional generation with separate noise levels for each modality.

Modality Forcing for Scalable Spatial Generation

arXiv:2605.20795v2 Announce Type: replace 
Abstract: Flow matching based video generative models have been increasingly relying on prepended Vision-Language Models (VLMs) to handle complex, instruction-based video editing. The prevailing assumption underlying this paradigm is that a connector module can seamlessly align the VLM's rich multi-modal reasoning with the original text embedding space of DiTs. However, we hypothesize that this alignment acts as a severe semantic bottleneck, degrading fine-grained structural variables. Verifying this is challenging, as end-to-end evaluations conflate alignment failures with generation errors, and natural datasets lack disentangled annotations. To rigorously investigate this, we propose a controlled data processing pipeline based on video composition that results in TRACE-Edit, a diagnostic dataset focusing on relation-based editing. Leveraging this dataset, we propose a comprehensive diagnostic protocol to analyze two important designs of meta-query and connector in the existing video editing models. Systematic evaluation of four representative model cases reveals that fine-grained structural semantics can be severely degraded during alignment. Our findings overturn the assumption of lossless semantic transfer, identifying the VLM-to-DiT alignment as a major bottleneck and providing a new diagnostic foundation for future multi-modal alignment architectures.

لقد سلطت التطورات الأخيرة في تكنولوجيا تحرير الفيديو الضوء على قيود نماذج اللغة والرؤية (VLM) في محاذاتها مع المحولات الديناميكية للصور (DiT). تقترح دراسة جديدة مجموعة بيانات تشخيصية تُعرف باسم TRACE-Edit، تهدف إلى فهم الاختناقات الدلالية التي تحدث أثناء عملية المحاذاة هذه، والتي قد تعيق فعالية تحرير الفيديو القائم على التعليمات.

Los avances recientes en la tecnología de edición de video han resaltado las limitaciones de los Modelos de Visión-Lenguaje (VLM) en su alineación con los Transformadores de Imágenes Dinámicas (DiT). Un nuevo estudio propone TRACE-Edit, un conjunto de datos diagnóstico que busca comprender los cuellos de botella semánticos que ocurren durante este proceso de alineación, lo que podría obstaculizar la efectividad de la edición de video basada en instrucciones.

Des avancées récentes dans la technologie de montage vidéo ont mis en évidence les limites des modèles de vision-langage (VLM) en matière d'alignement avec les transformateurs d'images dynamiques (DiT). Une nouvelle étude propose TRACE-Edit, un ensemble de données diagnostiques visant à comprendre les goulets d'étranglement sémantiques qui se produisent lors de ce processus d'alignement, ce qui pourrait entraver l'efficacité des montages vidéo basés sur des instructions.

Recent advancements in video editing technology have highlighted the limitations of Vision-Language Models (VLMs) in aligning with Dynamic Image Transformers (DiTs). A new study proposes TRACE-Edit, a diagnostic dataset aimed at understanding the semantic bottlenecks that occur during this alignment process, which may hinder the effectiveness of instruction-based video editing.

Chreode: A Cell World Model for One-Step Temporal Dynamics and Perturbation Prediction

Was this article worth reading? Share it

Ready to build your own newsroom?