Yaowu Fan, Tao Han, Dazhao Du, Andy J. Ma, Jia Wan
Recent advances in computer vision have produced powerful specialized models for detection, segmentation, counting, and other visual tasks. However, these models are typically optimized for isolated tasks, limiting their ability to support general-purpose visual intelligence, particularly when complex language understanding and dense small-object perception are required. We propose VisHarness, a trainable visual agent that decouples high-level perception, reasoning, and decision-making from low-level task execution. Rather than training a task-specific model, we train VisHarness to harness a set of carefully designed heterogeneous visual experts. This paradigm preserves the agent's general intelligence while leveraging the precision of specialized visual experts for specific tasks. With only lightweight training, VisHarness learns a generalizable policy for harnessing visual experts, enabling it to solve fundamental vision tasks in diverse complex scenarios through multi-turn interactions with these experts. To enable efficient on-policy reinforcement learning in a live environment, we introduce dynamic visual memory archiving to reduce the visual-token overhead that rapidly accumulates during multi-turn interactions with visual experts. Experiments on four representative benchmarks spanning reasoning segmentation, generalized referring expression segmentation, dense small-object detection, and referring expression counting show that VisHarness substantially outperforms existing general-purpose models and is competitive with or superior to task-specific models. Our code, model weights, and training data are publicly available at https://github.com/fyw1999/VisHarness.
Haoxi Li, Sikai Bai, Jianfei Ma, Qinglin Hou, Xiaocheng Lu, Tao Han, Jie Zhang, Song Guo
Learning from experience is essential for building the next generation of agents that continually expand their capabilities. Pretrained vision-language models (VLMs) already internalize rich world knowledge, yet this knowledge is rarely leveraged to drive policy improvement through online interaction. Realizing this potential efficiently remains challenging, as policy updates shift the agent's interaction distribution, requiring the world model to adapt accordingly, while continually updating large pretrained models is computationally expensive. To address this challenge, we introduce Dwell, a model-based reinforcement learning (RL) framework that reframes a VLM as a recurrent latent transition operator, endowing it with the ability to predict, adjust, and act in latent space. Leveraging the VLM's own representations and its dual role as both world model and policy, the agent continually learns from real interactions to anticipate and evaluate action outcomes, and from latent imagination to improve its behavior. Across five interactive visual environments, Dwell achieves the strongest overall performance among trainable methods while substantially reducing the compute required for imagination, showing that effective world models for VLM agents need not reconstruct the future; they need to predict futures that are useful for action. Our code is publicly available at https://github.com/Tom111-lhx/Dwell