Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models

Insight-V is a research framework that explores long-chain visual reasoning in multimodal large language models (MLLMs). It introduces a scalable pipeline for generating long and robust reasoning data for complex multimodal tasks without human labor, using a progressive strategy and multi-granularity assessment to ensure quality. To address the challenge of directly supervising MLLMs with long reasoning data, Insight-V employs a multi-agent system: a reasoning agent dedicated to long-chain reasoning and a summary agent that judges and summarizes results. An iterative DPO algorithm further enhances the reasoning agent's stability and quality. Built on LLaVA-NeXT, Insight-V achieves significant performance gains on challenging multimodal benchmarks requiring visual reasoning while maintaining or improving performance on perception-focused tasks.