MInfoVisQA: 24-language benchmark exposes cross-language robustness gaps in MLLMs
MBZUAI · hf · 2026-11-19
MBZUAI, Georgia Tech, UTokyo and UIUC release MInfoVisQA, a parallel benchmark for information-rich visual QA. It builds 8,960 QA configurations from 128 seed instances across 24 languages by disentangling numerical-structural data from text and programmatically syncing translated labels. Evaluating 25 MLLMs under Language-Aligned QA and Cross-Language QA settings, it reveals substantial robustness gaps when surface languages change.
More from Research
- Tsinghua-led ALIVE makes streaming video models interactive with proactive monitoring and feedback-seeking — Tsinghua · 2026-11-19
- Stanford removes the decoder: direct pixel prediction from diffusion transformers cuts cost, lifts quality — StanfordUniversity · 2026-11-19
- AnisoFlow learns anisotropic prediction targets for flow models, boosting FID without touching architecture — StanfordUniversity · 2026-11-19
- Debiasing Without Annotations: Cross-View Consensus Boosts Worst-Group Accuracy — Weiwei Li · 2026-11-19
- KAIST Proposes PROLA: Low-Rank Adaptation for Spatiotemporal Weather Downscaling — MLVLab · 2026-11-19
- NTU's PACT Adapts Time-Series Foundation Models Without Forgetting Old Conditions — NanyangTechnologicalUniversity · 2026-11-19