Benchmark
CLEVR
A synthetic visual question answering dataset. Images are rendered from simple geometric shapes under controlled lighting and positions. Questions primarily test counting, spatial relationships, and attribute comparison, and are highly sensitive to fine-grained spatial information.
Timeline (last 90 days)
Nov 2
On the CLEVR dataset of the VQA task, Latent-Lens achieves an accuracy of 81.0±7.9%, versus 58.3±5.7% for the text communication baseline.
Unverified50%