Temporal Localization
Identify which connection happened first, last, next, or most recently relative to an assembly state.
Overview
Assembly videos expose gaps in visual temporal reasoning.
Flat-Pack Bench is a multiple-choice benchmark for fine-grained spatio-temporal understanding. Each question asks a model to reason about parts, contact events, assembly order, final connectivity, or object identity across time in real furniture assembly videos.
Benchmark
Four question families target complementary assembly skills.
Temporal Ordering
Recover the exact order in which parts or part-pairs become physically connected.
Mating
Determine whether two labeled parts are connected in the fully assembled furniture.
Tracking
Match labeled parts across two states, requiring object identity through occlusion and motion.
Prompt construction
Questions combine a video with visual prompts that label the relevant parts. Non-tracking questions use one annotated image, while tracking questions use an Image A to Image B correspondence prompt.
Dataset Viewer
Switch between curated examples and the full benchmark browser.
0 curated samples
Use arrow keys
0 questions
Use J/K or arrow keys
Evaluation
Models are scored as multiple-choice video question answerers.
Inputs
The evaluation compares keyframe videos, trimmed videos, and visual prompt construction variants such as separate images, collages, and concatenated prompt media.
Outputs
Models answer with an option label. We report micro-average accuracy and per-category accuracy for TORD, TLOC, TRACK, and MATE.
Results
Current LVLMs remain far below human performance.
micro-average accuracy
InternVL3-78B in the paper's main benchmark setting
OpenAI GPT-5 in the paper's main benchmark setting
| Model | Prompt | Video | Micro | 95% CI | TORD | TLOC | TRACK | MATE |
|---|
Citation
BibTeX
TBA