AB
AiBoss
project

Insight-V - A multimodal model that enhances long-chain visual reasoning capabilities

Insight-V is a multimodal model jointly developed by researchers from Nanyang Technological University, Tencent, and Tsinghua University. It enhances the capabilities of large-scale multimodal language models in long-chain visual reasoning. It is based on scalable data generation streams...

What is Insight-V?

Insight-V is a multimodal model jointly developed by researchers from Nanyang Technological University, Tencent, and Tsinghua University. It enhances the capabilities of large-scale multimodal language models in long-chain visual reasoning. Based on a scalable data generation pipeline, it produces high-quality reasoning data and employs a multi-agent system to decompose the visual reasoning task into two steps: reasoning and summarizing. Combined with a two-stage training process, this significantly improves the model's performance on visual reasoning benchmarks. Insight-V's design includes progressive data generation, multi-granularity evaluation, and an iterative DPO algorithm, enabling the system to exhibit outstanding performance in complex visual reasoning tasks.

Insight-V's main functions

  • Long-chain visual reasoningInsight-V can handle complex visual reasoning tasks, solving problems by generating detailed, progressive reasoning processes.
  • Data generation processThe system has a scalable data generation process, which can be used to produce long-chain, high-quality inference data and support complex multimodal tasks.
  • Multi-agent systemInsight-V employs a multi-agent architecture, breaking down the visual reasoning task into two independent steps: reasoning and summarizing, which are handled by dedicated agents.
  • Two-stage training processThe system implements a two-stage training strategy, including supervised fine-tuning and direct preference optimization (DPO), to enhance the model's inference capabilities.
  • Performance improvementIn multiple visual reasoning benchmark tests, Insight-V demonstrated significant performance improvements, outperforming other advanced models.

The technical principles of Insight-V

  • Progressive long-chain inference data generationIt collects single-step reasoning results based on a powerful multimodal model and generates the next reasoning action based on historical reasoning results.
  • Multi-granularity evaluationThe process involves filtering the inference data using real answers to exclude data with incorrect final answers. A reasoning step scoring model is then used to score the quality of the inference data, dividing it into subsets of different quality.
  • Model DesignIt is specifically designed to generate detailed, step-by-step reasoning processes. It adapts to inaccuracies in the reasoning path, selectively incorporating or ignoring certain elements.
  • Training strategySupervised fine-tuning of two agents cultivates their ability to reason and summarize progressively. Multi-round DPO training and sampling simulate an online environment to improve model performance.

Insight-V project address

Application scenarios of Insight-V

  • autonomous drivingIn autonomous vehicles, I helps the vehicle understand road signs, traffic signals, and the surrounding environment to make accurate driving decisions.
  • Robot VisionIn robotics, improving a robot's understanding of its environment enables it to navigate and recognize objects more effectively.
  • Intelligent monitoringIn the field of security monitoring, video streams are analyzed to identify abnormal behaviors or events and issue timely alerts.
  • Medical image analysisIt helps doctors analyze medical images, such as X-rays, CT scans, and MRIs, to identify diseases and lesions.
  • Education and trainingIn the field of education, it serves as a supplementary tool, providing demonstrations and explanations of visual problem-solving to enhance the learning experience.