HVSBench: Human Visual System Benchmark for Multimodal Models

HVSBench is a benchmark for evaluating how well multimodal large language models align with human perceptual behavior. It contains 8,657 question-answer examples and references 7,507 unique raw images. The dataset was released by author garrying on Hugging Face and was last updated on May 14, 2026.

Use Cases

• Benchmarking model performance on human visual system tasks based on the described categories of prominence, subitizing, prioritizing, free-viewing, and searching.
• Training models to better align with human perceptual behavior using the provided question-answer examples.
• Analyzing the alignment between AI and human vision across different cognitive tasks mentioned in the description.

Strengths

• Contains 8,657 question-answer examples for evaluation.
• References 7,507 unique raw images, providing a substantial visual component.
• Covers multiple human visual system tasks as described: prominence, subitizing, prioritizing, free-viewing, and searching.

Limitations

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

Provenance

Source

garrying on Hugging Face

Collection Method

Packaged as parquet shards with embedded image bytes, derived from a 10% test subset described in an associated paper.

Freshness

Last updated 2026-05-14 09:27:12; freshness should be verified.