PolyBridgeBench (Loose Budget): leaderboard

PolyBridgeBench (Zhao et al., 2026) asks multimodal models to design complete load-bearing bridges for 189 Poly Bridge levels from a scene image and engineering constraints; designs pass deterministic legality checks, run in the native physics simulator, and a failed rollout is shown back to the model for repair. Success@3 is the percentage of levels with a working bridge within three model calls and three physics attempts; higher is better. Loose setting: the per-level material-cost limit is relaxed, which separates topology and load-bearing failures from budget failures.

Metric: Success@3 (%; share of all 189 executable Poly Bridge levels (149 core structural and 40 compositional); the model sees the scene image and structured constraints and returns a complete node, member and material topology; deterministic legality checks gate native dynamic physics execution, failed rollouts return visual evidence for repair; success when the vehicles cross within at most three model calls and three physics attempts; temperature 0.2 where exposed; loose budget: the per-level material-cost limit is relaxed, a diagnostic setting). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.

Top models

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

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