LetTokenBurn
Sign in
Auto86%Harbor: pendingMultimodalKnowledge & Reasoning

HAPACT

HAPACT: A Benchmark For Human-Centric Physical Impact Localization in Movies

A video benchmark for localizing physical impacts temporally and spatially on the human body, containing 2.8K video-query pairs from movies with SMPL-X annotations.

PaperRepo

How to run

ltb · hapact
$ npm i -g lettokenburn-cli
$ ltb bench info hapact

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

11 sourced entries
HarnessSourceConfidenceLink
GPT-5.442.6N/AIndependentmediumOct 16, 2026arxiv.org
GPT-4o28.9N/AIndependentmediumOct 16, 2026arxiv.org
Gemini 3.1 Pro27.2N/AIndependentmediumOct 16, 2026arxiv.org
Qwen3-VL-8B-Instruct22.9N/AIndependentmediumOct 16, 2026arxiv.org
Qwen3-VL-32B21.4N/AIndependentmediumOct 16, 2026arxiv.org
InternVL3.5-8B-Instruct14.0N/AIndependentmediumOct 16, 2026arxiv.org
Grounded-Video-LLM8.0N/AIndependentmediumOct 16, 2026arxiv.org
UniTime7.9N/AIndependentmediumOct 16, 2026arxiv.org
TimeLens5.1N/AIndependentmediumOct 16, 2026arxiv.org
TRACE4.2N/AIndependentmediumOct 16, 2026arxiv.org
VTimeLLM1.0N/AIndependentmediumOct 16, 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

HAPACT comprises over 2.8K video-query pairs from 212 movies, featuring frame-level temporal and SMPL-X spatial annotations for human-centric impact events. It evaluates temporal retrieval and spatial localization, showing that existing methods struggle with fine-grained impact reasoning.