Search papers, labs, and topics across Lattice.
3
0
7
9
Achieving a 71.81% zero-shot Pass@1 score, the DeepSeek-V4-Flash model outperforms leading competitors by leveraging a novel optimization framework on Ascend SuperPOD.
Shared descriptive patterns in skill libraries can obscure task-relevant signals, but SkillSight reveals and calibrates these biases to boost retrieval accuracy by over 20%.
MLLMs don't fuse vision and language uniformly: targeted interventions guided by layer-wise attention analysis can significantly boost multimodal reasoning without retraining.
