自进化代码图像推理：通过可执行反思提升视觉计算能力

AI 摘要

该研究提出Code-with-Image范式，将视觉推理从语言链转移到Python程序执行，并构建CwI-Bench基准测试。作者提出Self-Reflection over Executable Reasoning，一种无需训练的自我进化循环，通过观察和可执行反射改进技能库，显著提升GPT-5.6-luna和Qwen3.5-27B在基准上的准确率，且技能可跨模型迁移。

正文节选

Self-Evolving Code-with-Image Reasoning Tianze Yang1,2, Liang Wu1, Ruitong Sun2, Yucheng Shi2, Yanqiao Wang2, Mayank Darbari1, Ninghao Liu3, Jin Sun2, Liangjie Hong1 1Nokia 2University of Georgia 3The Hong Kong Polytechnic University 1. Introduction Multimodal models are increasingly asked visual questions that no single glance can settle. The thinking-with-images paradigm [28, 42, 12, 26, 41, 38, 13, 24] responds by letting the model act: instead of answering in one pass, it iteratively manipul

发布时间：2026-08-13 12:00

抓取时间：2026-08-13 12:57

来源机构：arXiv