Publications

We introduce Vero, a family of fully open vision-language models (VLMs) designed for general visual reasoning across diverse domains such as charts, science, and spatial understanding. By scaling reinforcement learning (RL) data with the 600K-sample Vero-600K dataset and task-routed rewards, Vero achieves state-of-the-art performance on 30 challenging benchmarks, demonstrating that broad data coverage is the key driver of strong RL scaling. All data, code, and models are publicly released.

We propose a hierarchical input-dependent state space model that leverages the linear scalability of SSMs for decision-making over full-length videos. Our framework couples a temporally consistent visual encoder with an SSM head to propagate temporal information. The temporal module consists of two components: a local aggregation block for fine-grained dynamics and a global relation block for long-range dependencies.