AI Agent Papers2026
Turns agent decisions into conditional predictions from historical evidence, using one shared pretrained predictor across diverse tasks. With Qwen3.5-27B, it reaches 84.92% mean accuracy across 50 benchmarks.
Fuhao Li, Wenxuan Song +8The Hong Kong University of Science and Technology (Guangzhou), OLA Dimensions +2
Integrates 3D spatial supervision into continual pre-training for robot vision-language-action policies, helping spatial understanding transfer to downstream tasks. Across eight real-world deployment conditions, it improves average success over π0.5 by 24 percentage points.