PuzzleMate - Piece-to-Location Matching with Natural Language Reference (Aria): leaderboard
A task of PuzzleMate (Dasgupta et al., 2026), an egocentric jigsaw-puzzle benchmark for multimodal assistants built on three 48-piece puzzles: the model describes in a short natural-language answer which of several unlabeled missing slots a given piece belongs to, on photos from a low-resolution Project Aria glasses camera. Scored as accuracy in percent as judged by human annotators; higher is better.
Metric: Accuracy (%; correct as judged by human annotators for clarity and spatial accuracy; task: describing in natural language, within 35 tokens, which of several unlabeled missing slots a given piece belongs to; three 48-piece jigsaw puzzles photographed from an egocentric viewpoint, Project Aria glasses camera (low-resolution, noisy egocentric images)). Source: arxiv.org. Saturation forecast: Around November 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 52 | #172 |
| 2 | Gemini 2.5 Pro | 48 | #262 |
| 3 | Gemma 3 27B | 41 | #1638 |
| 4 | Gemma 3 12B | 37 | #2013 |
| 5 | Gemma 3 4B | 28 | #5104 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=puzzlemate-piece-to-location-matching-with-natural-language-reference-aria · How It Works · Data refreshed daily, snapshot 2026-11-12.