PyVision: Agentic Vision with Dynamic Tooling

TL;DR

PyVision introduces a flexible framework enabling multimodal large language models to autonomously generate, refine, and utilize Python-based tools for visual reasoning, leading to significant performance improvements across benchmarks.

Contribution

It presents a novel interactive framework that allows models to dynamically create and adapt tools for visual reasoning, surpassing static workflow limitations.

Findings

Achieves +7.8% performance on V* with GPT-4.1

Boosts performance by +31.1% on VLMsAreBlind-mini with Claude-4.0-Sonnet

Develops a taxonomy of dynamically created tools

Abstract

LLMs are increasingly deployed as agents, systems capable of planning, reasoning, and dynamically calling external tools. However, in visual reasoning, prior approaches largely remain limited by predefined workflows and static toolsets. In this report, we present PyVision, an interactive, multi-turn framework that enables MLLMs to autonomously generate, execute, and refine Python-based tools tailored to the task at hand, unlocking flexible and interpretable problem-solving. We develop a taxonomy of the tools created by PyVision and analyze their usage across a diverse set of benchmarks. Quantitatively, PyVision achieves consistent performance gains, boosting GPT-4.1 by +7.8% on V* and Claude-4.0-Sonnet by +31.1% on VLMsAreBlind-mini. These results point to a broader shift: dynamic tooling allows models not just to use tools, but to invent them, advancing toward more agentic visual reasoning.