Blueprint-Bench 2, created by Andon Labs, tests whether AI agents can build a coherent spatial model of an apartment from its interior photographs. Each agent processes 50 apartments, examining around 20 photos per apartment, and produces a 2D floor plan describing the rooms, their connections, and their relative sizes, with a persistent notepad carried across apartments. It is the second version of Andon Labs’ Blueprint-Bench, run in an agent-only setting.

We source results from the public Blueprint-Bench 2 leaderboard. Our chart reports the leaderboard’s normalized score, where 0 corresponds to the random baseline and 1 to a perfect floor plan; scores at or below the random baseline are clamped to 0. The underlying raw composite score and its standard error are kept in the data export.

Generated floor plans are compared to ground truth as room-connectivity graphs, combining the similarity of room-to-room connections with degree, density, room-count, door-count, and orientation terms, invariant to rotation and reflection. Scoring is deterministic, with no image-overlap metric or LLM judge. For reference, Andon Labs reports a human baseline normalized score of about 0.59 on a subset of apartments.