Search papers, labs, and topics across Lattice.
Peking University
2
0
4
VLMs can achieve better multimodal reasoning simply by dynamically rescaling positional indices based on information density, without any training or architectural changes.
ID-LoRA slashes trainable parameters by up to 46% compared to standard LoRA while boosting performance across diverse benchmarks, offering a sweet spot between efficiency and effectiveness for fine-tuning LLMs.
