VLDBench: Large-Scale Multimodal Disinformation Detection

VectorInstitute released VLDBench in January 2026 as a large-scale benchmark for evaluating Vision-Language Models (VLMs) and Large Language Models (LLMs) on multimodal disinformation detection. The framework provides a testing ground for AI safety by presenting models with deceptive content that integrates both visual and textual modalities.

Use Cases

• Evaluating VLM accuracy in identifying inconsistencies within multimodal disinformation labels.
• Benchmarking LLM reasoning when processing deceptive image-text pairs.
• Testing AI safety guardrails against synthetic media features within the benchmark framework.

Strengths

• Multimodal focus targeting both VLMs and LLMs
• Updated as of January 2026
• Developed by VectorInstitute for AI safety research

Limitations

• Unknown record count and column schema in provided metadata
• Disinformation patterns are subject to rapid temporal staleness

Provenance

Source

VectorInstitute

Freshness

Updated January 2026.