We introduce polylithic systems, where independent producers contribute specialized capabilities to a shared agentic system. Our experiments show that coordinated specialists from multiple producers outperform monolithic generalists, and even small producers that underperform in isolation can improve the overall system by 5–13%.

We introduce VisGym, a gymnasium of 17 environments for evaluating and training VLMs. The suite spans symbolic puzzles, real-image understanding, navigation, and manipulation, and provides flexible controls over difficulty, input representation, planning horizon, and feedback.

FrontierCS is a benchmark of unsolved, open-ended, verifiable, and diverse computer science challenges designed to evolve alongside AI capabilities, featuring problems like polyomino packing that remain difficult even for advanced models.

We propose YOLO-Count, a differentiable open-vocabulary object counting model that tackles both general counting challenges and enables precise quantity control for text-to-image (T2I) generation.

CharXiv reveals significant shortcomings in MLLMs’ chart understanding, showing a large performance gap between models and humans.

We close the performance gap between screenshot Language Models and text-only Language Models on language understanding tasks with our PTP objective.

We propose TutorChat and TutorEval, a dataset of long synthetic dialogues about textbooks and a question-ansering benchmark consisting questions about long chapters from STEM textbooks written by human experts.

We introduce token-wise consistency terms between the image content and object segmentation maps in training text-to-image models for enhanced multi-category instance composition and photorealism.

We provide a two-way integration for the widely-adopted ControlNet method by integrating four external condition generation algorithms into a single dense image labeling method, and by integrating its individually trained image generation processes into a single model.

We establish a framework that effectively and efficiently finetunes a language model with embodied experience while retaining its language modeling abilities.

We investigate whether internal models learned by modern model-based RL algorithms can be leveraged to solve new, distinctly different tasks faster.