The Furniture Assembly Benchmark is Epoch AI’s benchmark for spatial reasoning about physical construction. Given a photo of a partially or fully assembled piece of IKEA furniture and the official assembly manual, a model must decide whether the build is correct so far and, if not, identify the step in which the mistake was made and describe it. Spotting assembly errors requires connecting a technical diagram to a real object, a capability relevant to tasks like repairing appliances or fixing a car that is largely untested by existing visual reasoning benchmarks.

The benchmark consists of photos taken during the assembly of three IKEA products of increasing complexity: the Ställ shoe cabinet, the Tonstad bed frame and the Gullaberg dresser. Some photos show correct builds; others contain a deliberate, clearly visible mistake, and in many cases assembly continued past the mistake to make it harder to spot.

For the full launch analysis, see our publication, AI has improved significantly at reasoning about IKEA furniture assembly.

Epoch has used photos of IKEA products and IKEA assembly manuals solely for the purposes of research and commentary. This work is conducted independently of IKEA. Epoch does not distribute or sell access to the underlying copyrighted material.

Photos. We purchased three pieces of IKEA furniture, chosen to span a range of difficulty according to the IKEA Complexity Index: the Ställ shoe cabinet (14 photos), the Tonstad bed frame (25 photos) and the Gullaberg dresser (21 photos), 60 photos in total. We photographed the builds at various points during assembly. Some photos show a correct build; for the others we intentionally made realistic assembly mistakes, and often continued building past the mistake so that the error is visible but not the most recent step. For every photo containing a mistake, the issue is clearly visible to someone who knows what to look for.

Task. For each photo, the model is told which product is being built and which step was just completed, and is given the photo, the official assembly manual as a PDF, a zoom tool to view parts of the image at higher resolution, and a Python interpreter. It must decide whether the build is correct so far. If it finds a mistake, it must report the step in which the mistake was made and describe the issue; if it finds several, it must list all of them. If the build is correct, it must direct the user to the next step. Each sample runs in an agentic sandbox with a budget of 80 steps; the model is warned when it is close to the limit and asked for a final answer when it reaches it. In practice, fewer than 5% of samples hit the limit.

Grading. Grading is multistep. If the build has no mistake and the model says so, the answer is marked correct. If the build has a mistake, the model’s answer is first checked against the answer key for the step(s) in which the mistake(s) were made; if it identifies the correct step(s), its description of the mistake is then checked by an LLM grader (GPT-5.6 Sol), which is prompted to be lenient: as long as the model identifies the mistaken step and roughly describes the issue, it receives credit. The reported score is accuracy across the 60 photos; error bars show ± standard error.

We evaluate each model at the highest reasoning effort its API supports. For open-weights models we used the official developer API endpoints (Moonshot and Alibaba), although speed and quality for open models can vary significantly by provider. Some models lack image support entirely and cannot be evaluated.

Human baseline. In September 2026, we measured human performance on the benchmark. Fifteen volunteers with no previous experience assembling these IKEA products each answered every photo from one of the three products, five participants per product. Participants saw the same photos and assembly manuals as the models, and had to decide whether each build was correct so far and, if not, name the step(s) in which mistakes were made. Their task was slightly easier than the models’: they did not have to describe the mistakes, so they were graded on the steps alone, with no LLM grader. Participants spent a median of 77 seconds per photo.

The average participant scored 63% across the 60 photos. No participant answered all 60 photos, so the best human score of 84% is calculated by combining the top-scoring participant on each product.