IKEA-Bench - Step Recognition: leaderboard

Metric: Accuracy (%) of four-way multiple-choice step recognition (match the video frames of one assembly step to one of four wordless manual diagrams, distractors from adjacent steps; 320 questions; chance 25) on IKEA-Bench, wordless IKEA manual diagrams against real assembly video from 29 products, Visual setting (diagram images only), zero-shot greedy decoding, unparsed answers counted wrong; higher is better. Source: arxiv.org. Saturation forecast: Around October 2027. 19 models tracked.

Top models

#ModelScore
1Gemini 3 Flash65.3
2Gemini 3.1 Pro (Preview)62.8
3Qwen 3.5 27B59.4
4Qwen 3.5 9B57.8
5Qwen 2.5 VL 7B49.1
6GLM-4.1V-9B (Thinking)48.4
7Qwen 3.5 2B44.4
8Gemma 3 27B43.1
9Gemma 3 4B39.4
10Gemma 3 12B35.3

Interactive version: theaggregate.ai/benchmark?slug=ikea-bench-step-recognition · How It Works · Data refreshed daily, snapshot 2026-10-07.