Surgical Scene Understanding Benchmark for Multimodal LLMs

SurgMLLMBench is a multimodal benchmark dataset for training and evaluating interactive multimodal large language models in surgical scene understanding. It integrates diverse surgical video content. The dataset was created by KIST-HARILAB and was last updated in February 2026.

Use Cases

• Benchmarking MLLM performance on surgical scene question-answering tasks using video frames and text prompts.
• Training models for surgical instrument recognition from integrated video data.
• Evaluating MLLM reasoning on surgical workflow steps described in the benchmark's textual annotations.

Strengths

• Designed specifically as a benchmark for the complex task of surgical scene understanding.
• Integrates multimodal data (video and text) for interactive model evaluation.

Limitations

• Unknown dataset size, row count, and specific video duration limits evaluation scope.
• Lack of accessible column and sample data details hinders preliminary analysis.

Provenance

Source

KIST-HARILAB, presented in the associated academic paper.

Collection Method

null

Time Range

null

Freshness

Last updated February 2026.

Geography

null