Search papers, labs, and topics across Lattice.
School of Cyber Science and Engineering, Qufu Normal University, Qufu, Shandong, China
8
2
15
1
Z-Reward achieves nearly 90% human preference accuracy by transforming subjective visual preferences into nuanced score distributions, outperforming traditional reward models.
MiniMax-M2 proves that massive parameter counts don't always translate to better agentic performance; strategic activation of a smaller subset can unlock frontier-level intelligence.
LLMs can finally extract cyberattack patterns from threat reports with high precision, thanks to a novel "diverge-then-converge" approach that mimics how human analysts validate findings.
Fine-tuning efficient few-step diffusion models no longer requires sacrificing their speed, thanks to a self-distillation approach that preserves inference capabilities.
LLMs can't even reproduce published physics papers end-to-end, with the best model scoring only 34% on a new benchmark designed for this purpose.
Image protection schemes are surprisingly brittle: off-the-shelf image-to-image models can bypass them with a simple "denoise" prompt, often outperforming specialized attacks.
Dataset distillation can be sped up by 18x on ImageNet-1K without sacrificing accuracy by focusing optimization on high-loss regions.
PANC achieves a staggering 162% increase in attack effectiveness against Transformer-based visual trackers while compressing adversarial noise to just 10%.
