Plan-Bench

Plan-Bench: Can Existing VLMs Go Well Beyond Perception to Reason and Plan in Embodied Worlds?

Plan-Bench is a benchmark designed to evaluate the high-level reasoning and multi-step planning capabilities of Vision-Language Models (VLMs) in embodied environments, independently of low-level action execution.

Dataset Overview

Plan-Bench contains 699 curated samples across three core dimensions and 16 subtasks.

Category	Samples	Subtasks
Spatial Cognition	259	6
Temporal Reasoning	210	5
Multi-Step Planning	230	5
Total	699	16

The benchmark includes:

• 183 single-choice questions
• 516 open-ended questions
• 3,007 images
• Both single-image and multi-image inputs

Evaluation

Plan-Bench supports execution-decoupled evaluation of VLMs using an LLM-as-judge protocol with answer hints for semantically equivalent responses.