MMLongCite: A Benchmark for Evaluating Long-Context Vision-Language Models

MMLongCite is a benchmark dataset created by Jonaszky123 to evaluate the faithfulness of long-context vision-language models. It covers 4 task categories, including Single-Source Visual Reasoning, Multi-Source Visual Reasoning, Vision Grounding, and Video Understanding, encompassing 8 distinct long-context tasks. The dataset was last updated on May 7, 2026.

Use Cases

• Benchmarking model faithfulness based on citation tasks
• Evaluating visual reasoning capabilities across single and multiple sources
• Testing vision grounding performance in long-context scenarios
• Assessing video understanding abilities in extended contexts

Strengths

• Covers 4 distinct task categories for evaluation
• Encompasses 8 specific long-context tasks
• Designed specifically for evaluating long-context vision-language models

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

huggingface

Freshness

Last updated 2026-05-07 04:05:08