List your company

Layout Generation on MesaTask-CTRC benchmark 1.0

46.5Quality Pass Rate (τ=7)PhyScene3D
0.94812.77424.636.426Jun 1, 2026

Evaluation results

MethodLinks
PhyScene3DBackbone=Qwen-3 VL 8B, Stages=SFT + Physically-Aware RL (GRPO)2026.0646.539.127.37.859.459.918.299.158.9341.63.86
GPT-4oParadigm=Zero-shot Generation, Source=Closed-source LLM2026.0627.622.112.76.958.799.547.488.188.1968.97.87
MesaTaskParadigm=End-to-End Model, Type=Supervised Baseline2026.0621.118.614.87.789.239.788.159.058.878.38.19
I-Design-tableParadigm=Agent-Based Solver2026.06199.74.54.398.29.25.345.556.5339.15.94
GPT-4o-miniParadigm=Zero-shot Generation, Source=Closed-source LLM2026.0617.913.48.86.188.459.156.827.347.598013.61
ReferenceDescription=Samples drawn from MesaTask-10k2026.0617.115.111.57.839.259.528.259.188.8781.58.19
Holodeck-tableParadigm=Agent-Based Solver2026.062.70.602.356.217.383.613.434.62.70.47