AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

LEGO-Puzzles: How Good Are MLLMs at Multi-Step Spatial Reasoning?

Kexian Tang, Junyao Gao, Yanhong Zeng, Haodong Duan, Sun Yanan, Zhening Xing, Wenran Liu, Kai Chen, Kaifeng Lyu

ECCV 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 53 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
raise required reasoning steps k from 1 to 8 in PLAN-k-Step-VQAGPT-5 · PLAN-k-Step-VQAexact match accuracy: 90% -> 0%Hurt
move from Next-Step to Ordering taskGPT-5 and most MLLMs · LEGO-Puzzles Elementaryaccuracy: e.g. Qwen3-VL-32B 62.0 -> 34.0Hurt
proprietary vs open-source MLLMsMLLMs · LEGO-Puzzles Elementaryaccuracy: 72.0 (GPT-5) vs 57.0 (best open-source)Improved
MLLMs vs human expertsbest MLLMs · LEGO-Puzzles-Liteaccuracy: 71.8 vs 93.6 humanHurt
change output from multiple choice to image generationGPT-5 · LEGO-Puzzles planningaccuracy: 0 correct generative trajectoriesHurt
use LEGO-Puzzles Height/Adjacency vs 3DSRBenchproprietary MLLMs · 3DSRBench correlationPearson correlation: PCC 0.93 (Height), 0.98 (Adjacency)Improved
Height task on MLLMsMLLMs · LEGO-Puzzles Heightaccuracy: 21/29 models below random thresholdHurt
scale model size within familiesQwen2.5-VL · LEGO-Puzzles Elementaryaccuracy: 31.5 -> 46.6 (7B -> 72B)Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy