DRIFT: A Residual Flow Adapter for Decoding Continuous Outputs in Vision-Language Models

Open paper
Benchmark Model Rank Results
robot-manipulation-on-libero-spatialDRIFT (Qwen3-VL-2B)#31Success Rate: 98.8
robot-manipulation-on-simplerenv-widow-xDRIFT (Qwen3-VL-2B)#32Average: 0.615
robot-manipulation-on-libero-longDRIFT (Qwen3-VL-2B)#33Success Rate: 96.4
robot-manipulation-on-libero-objectDRIFT (Qwen3-VL-2B)#48Success Rate: 99.2
robot-manipulation-on-liberoDRIFT (Qwen3-VL-2B)#54Success Rate: 97.9
robot-manipulation-on-libero-goalDRIFT (Qwen3-VL-2B)#58Success Rate: 97.0
moment-retrieval-on-charades-staDRIFT (ET-Chat)#78mIoU: 43.8

Rank counts only results with a code link.