XModBench: A Tri-Modal Multiple-Choice QA Benchmark for Cross-Modal Consistency

XModBench is the first tri-modal benchmark designed to measure cross-modal consistency in omni-language models. It presents the same semantic content in audio, vision, and text modalities through 4-choice questions. The benchmark was created by RyanWW and accepted for presentation at ICLR 2026.

Use Cases

• Benchmarking model performance on tri-modal (audio/vision/text) understanding based on the described multiple-choice QA format.
• Evaluating cross-modal consistency in AI systems based on the benchmark's core design goal.
• Training or fine-tuning multimodal models on a structured QA task with context and choices.
• Analyzing the relative difficulty of questions presented in different sensory modalities.

Strengths

• Explicitly designed as the first benchmark for measuring cross-modal consistency in omni-language models.
• Features a tri-modal (audio, vision, text) multiple-choice QA structure.
• Accepted for presentation at the ICLR 2026 conference, suggesting peer review.

Limitations

• Description metadata is limited; actual data quality, size, and content require manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Freshness

Last updated 2026-05-18 17:40:09; freshness should be verified.