TL;DR: We retrieve manipulation-relevant human demonstrations from 100,000 web video clips using 3D hand-motion trajectories, improving dexterous-hand RL policies in Isaac Gym.

TL;DR: We compare model representations with recorded motion trajectories to assess robot-policy and backbone capability using a training-free offline score.

TL;DR: We show that RL post-training on simple arcade games improves multimodal LLMs’ out-of-domain math and spatial reasoning without worked solutions or diagrams during that post-training stage.

TL;DR: We reuse stale rollouts in a few large generation and optimization stages, matching standard GRPO’s reasoning performance at lower training cost.

TL;DR: We build an automated grounding, annotation, and filtering pipeline for multimodal medical pretraining, covering 25 million images across 10 modalities.

TL;DR: We use local and global aggregation in a hierarchical vision transformer to improve part and object segmentation.

TL;DR: We improve consistency in long-story visualization with a training-free method that refines story images using global reference context from the previous iteration.

TL;DR: We improve multi-agent LLM game performance and stability through self-play, persistent memory, and tournament-style prompt evolution.