Search papers, labs, and topics across Lattice.
3
0
5
0
SkillFuzz uncovers over 1,000 implicit intents in skill compositions, revealing hidden risks that traditional auditing methods miss.
Fine-tuning on the new ProCUA-SFT dataset boosts UI-TARS 7B's performance from a dismal 8-10% to an impressive 45.0% on OSWorld tasks, highlighting the critical role of high-quality training data.
SSR not only resolves destructive parameter collisions in LoRA merging but also guarantees mathematical optimality, setting a new standard for efficiency in diffusion model training.
