Search papers, labs, and topics across Lattice.
Zayed University of Artificial Intelligence Abu
3
0
6
Layer span variation can unlock a new dimension of deterministic rollout diversity, boosting performance by over 10 percentage points on reasoning tasks.
PreLort ensures that low-rank clients can leverage the richer information from higher-rank clients, leading to significant improvements in model performance.
Multi-turn reasoning models can appear aligned while still producing harmful outputs, exposing a critical gap in traditional evaluation methods.
