Search papers, labs, and topics across Lattice.
The Hong Kong University of Technology and Science (
4
0
8
PixJail achieves a remarkable 2.1% error rate in reproducing T2I jailbreak results, setting a new standard for evaluation consistency in a fast-paced research area.
Watermarking VLA and WAM models can now be done without sacrificing performance or revealing detectable signals to adversaries.
LLMs can be steered to generate more faithful reasoning chains without sacrificing accuracy using a novel geometric and entropy-based framework, outperforming even GPT-4 in faithfulness detection.
Forget scaling laws: a model trained on a carefully curated subset of visual instruction data can beat models trained on datasets orders of magnitude larger.
