GPT-5.5, Claude Fable 5 Score Below 11% on ActiveVision Benchmark Testing Multi-Step Visual Inspection
A new evaluation framework called ActiveVision tests whether multimodal language models can perform iterative visual inspection, requiring AI systems to scan, trace, and compare details across multiple steps instead of answering from a single static image. In testing, GPT-5.5 scored 10.6% accuracy and Claude Fable 5 scored 3.5%.
The benchmark results place models well behind human performance, which averaged 96.1% on the same tasks. Coding agents utilizing vision tools reached between 24.7% and 50.6%, primarily because they failed to detect errors in their visual inputs. Researchers indicate the findings suggest multimodal systems need a functional perception and reasoning loop rather than simply processing additional tokens.
From the sources (2 posts)
@askalphaxiv"An Exam for Active Observers" Even though MLLMs can describe images, they still struggle to actively look at them. This paper introduces ActiveVision, a benchmark where models must scan, trace, and compare visual details over multiple st
@ofirpressRT @JiaruiZ58876329: Vision is a loop, not a glance. We introduce ActiveVision, a benchmark testing whether models can repeatedly observe,…