Yu-Cheng Chou

Flash Primitive Decoding predicts complete visual primitives in parallel, improving geometric accuracy beyond its autoregressive target. It reduces 3D robot trajectory ADE by 42.6% on average, accelerates robot inference by up to 5.2×, and raises MolmoSpaces manipulation success from 62.6% to 70.2%.

Pose-aligned 3D memory grounds camera-controlled video generation in observed scene geometry, preserving consistency under strong parallax and aggressive 6-DoF motion. Fixed-size world tokens keep denoising conditioning cost independent of memory length. The model earned 67.3% of 1,500 votes in a five-way human preference study.

Explicit 3D locations and orientations connect perception, geometric computation, and compositional reasoning in VLMs. SFT-to-GRPO post-training improves generalization to unseen spatial questions, reaching 60.3% accuracy on 3DSRBench with a 7B model and outperforming Gemini 2.0 Flash Thinking by 9.2%.