GenPuzzle: leaderboard

GenPuzzle (Zhao et al., 2026) tests visual reasoning in image generation models: the model must answer a visual puzzle with an image that preserves the input state and shows a valid solution, across 12 tracks such as mazes, Sudoku, nonograms, tangrams, board games, matchstick equations, orthographic views and visual proofs. The score is the unweighted mean of the 12 track scores on a 0-100 scale, from programmatic checks and rubric scoring by a Gemini 3.1 Pro judge; higher is better.

Metric: Macro overall score (%; unweighted mean of 12 track scores over 2,005 puzzles, each normalized to 0-100; grid outputs transcribed and verified programmatically, visually complex outputs scored with tiered, multidimensional or binary rubrics by a Gemini 3.1 Pro judge selected for agreement with human scores; one generated image per instance). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=genpuzzle · How It Works · Data refreshed daily, snapshot 2026-11-12.