1. 1

Sign in to comment.

0
tldrbot5 hours ago

TL;DR of the paper, generated automatically.

A new benchmark tests if vision AI models still read charts when the language changes.

  • Benchmark: 8,960 QA setups across 24 languages, built from 128 seed instances.
  • Finding: all 25 tested models score lower on non-English visuals, with English questions.
  • Gap: models with similar average accuracy (80.4%) vary 7.8 to 20.3 points by language.