VLX-Seek improves fine-grained VLM perception by turning fragile coordinate generation into region reference. It introduces addressable region tokens, a hybrid fine-grained region encoder, and compact object-centric reasoning for detection, counting, and open-vocabulary localization.
Read more →
VLX-Seek 面向端侧具身视觉，将细粒度感知任务从坐标生成改写为区域指代。通过区域 token、混合细粒度区域编码器和更紧凑的对象级推理，它让 VLM 在检测、计数和开放词汇定位中看得更准。
阅读全文 →
VLX-Flow turns live video streams into reusable model memory. It processes continuous chunks, maintains visual cache and semantic memory, and supports low-latency interaction without reprocessing the full video history for every query.
Read more →
VLX-Flow 面向一个越来越关键的问题：当视频从离线文件变成实时输入，多模态模型应该如何持续观察、持续记忆，并在任意时刻完成交互？VLX-Flow 将视频拆成连续片段，以增量方式更新视觉上下文和语义记忆，让模型不必在每次提问时重新处理完整历史。
阅读全文 →
Existing Vision-Language Models (VLMs) excel at holistic scene understanding but fail at precise, object-centric tasks like detection, primarily due to their inability to generate accurate coordinates. We propose VLM-FO1, an approach that solves this by transforming object detection from a generation to a retrieval problem. We treat bounding boxes as visual prompts, extract their features into unique "object tokens", and feed them directly to the model. This method dramatically improves performance, with VLM-FO1-3B reaching 44.4 mAP on COCO, rivaling specialized detectors and demonstrating strong capabilities on other region-based perception tasks.
Read more →
With the rapid advancement of Large Language Models (LLMs) and Vision Language Models (VLMs), AI technology is shifting from exam-oriented task completion to practical scenario-based complex problem-solving. Using LLMs and VLMs to tackle more realistic and intricate problems—rather than simply passing exams—is not only an inevitable direction of technological evolution but also a key requirement for industrial applications. We launched the first embodied AI agent—OmAgent, a reinforcement learning-based multimodal agent framework. Its feasibility has been verified in practical applications.
Read more →
随着大型语言模型（LLMs）与视觉语言模型（VLMs）的能力飞速发展，AI 技术正从「应试式」的任务达标转向「实战化」的复杂问题解决。用LLMs和VLMs去解决更实际更复杂的问题，而不是简单地通过“考试”， 这既是技术演进的必然方向，也是产业落地的核心诉求。我们推出全球首个具身智能 AI Agent——OmAgent，一个基于强化学习的多模态智能体框架，并在实际应用中验证了该路径的可行性。
Read more →
Key insights from our extensive experimentation with Reinforcement Learning for object detection in Vision Language Models, focusing on training methodologies, reward functions, and prompt engineering.
Read more →
A detailed exploration of how reinforcement learning enhances object detection performance compared to supervised fine-tuning in vision-language models.
Read more →