Benchmarking the Generality of Vision-Language-Action Models

TL;DR

This paper introduces MultiNet v1.0, a comprehensive benchmark to evaluate the cross-domain generality of vision-language models and action models, revealing current models' limitations in generalizing beyond training distributions.

Contribution

The paper presents MultiNet v1.0, a unified benchmark for assessing the generality of vision-language models across multiple capability regimes, addressing fragmented evaluation practices.

Findings

Models show significant performance drops on unseen domains and modalities.

Current models suffer from modality misalignment and output instability.

There is a notable gap between the goal of generalist AI and current model capabilities.

Abstract

Generalist multimodal agents are expected to unify perception, language, and control - operating robustly across diverse real world domains. However, current evaluation practices remain fragmented across isolated benchmarks, making it difficult to assess whether today's foundation models truly generalize beyond their training distributions. We introduce MultiNet v1.0, a unified benchmark for measuring the cross domain generality of vision language models (VLMs) and vision language action models (VLAs) across six foundational capability regimes. Visual grounding, spatial reasoning, tool use, physical commonsense, multi agent coordination, and continuous robot control. Evaluating GPT 5, Pi0, and Magma, we find that no model demonstrates consistent generality. All exhibit substantial degradation on unseen domains, unfamiliar modalities, or cross domain task shifts despite strong performance within their training distributions.These failures manifest as modality misalignment, output format instability, and catastrophic knowledge degradation under domain transfer.Our findings reveal a persistent gap between the aspiration of generalist intelligence and the actual capabilities of current foundation models.MultiNet v1.0 provides a standardized evaluation substrate for diagnosing these gaps and guiding the development of future generalist agents.Code, data, and leaderboards are publicly available.