Auto71%Harbor: pendingMultimodal

MLLMs-ZS-HOI

Benchmarking Multimodal Large Language Models for Zero-Shot Human-Object Interaction

A benchmark framework evaluating MLLMs on zero-shot Human-Object Interaction recognition and detection using adapted HICO, HICO-DET, and V-COCO datasets with novel QA annotations.

How to run

ltb · mllms-zs-hoi
$ npm i -g lettokenburn-cli
$ ltb bench info mllms-zs-hoi

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

8 sourced entries
HarnessSourceConfidenceLink
Qwen2-VL-7B52.9N/AIndependentmediumOct 13, 2026arxiv.org
Qwen2-VL-2B51.4N/AIndependentmediumOct 13, 2026arxiv.org
MiniCPM-V 148.2N/AIndependentmediumOct 13, 2026arxiv.org
MiniCPM-V 247.5N/AIndependentmediumOct 13, 2026arxiv.org
LLaVA-OV-0.5B46.1N/AIndependentmediumOct 13, 2026arxiv.org
InternVL2.5-2B43.5N/AIndependentmediumOct 13, 2026arxiv.org
InternVL2.5-1B41.8N/AIndependentmediumOct 13, 2026arxiv.org
Phi-3.5-Vision40.4N/AIndependentmediumOct 13, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

The benchmark transforms standard HOI datasets (HICO, HICO-DET, V-COCO) into question-answer formats with new annotations. It introduces a taxonomy of zero-shot prompting strategies and visual prompting methods like Objects as Context.

Scoring converts token-level logits into interaction scores to enable mAP calculation. Experiments show state-of-the-art zero-shot performance, achieving 53.50 mAP on HICO and 23.69 mAP on HICO-DET.