Linhao Zhang

Turning messy execution traces into structured episodic memory stops agents from repeating failed terminal commands, boosting autonomous environment setup to a 75% Pass@1 rate across nine programming languages.

Pretraining loss actively misleads video data curation: broad data coverage systematically beats aggressive quality filtering across downstream benchmarks in both Wan 2.1 and V-JEPA 2.1.

Depth information can drastically improve object counting accuracy in crowded scenes, achieving over 60% reduction in counting errors.