Search papers, labs, and topics across Lattice.
4
0
6
Visual understanding, not knowledge, is the critical barrier in Document Visual Question Answering, with smaller models showing surprising adaptability through targeted finetuning.
Vision transformers and frequency-domain representations unlock surprisingly strong performance in automated analysis of ultra-widefield retinal images for diabetic retinopathy.
CrimeNERdb offers a crucial resource for improving crime-related information extraction, filling a significant gap in NER datasets.
Fine-tuning a small Llama 3.1 model on the new EduEVAL-DB dataset allows it to rival the pedagogical risk detection capabilities of a much larger model (Gemini 2.5 Pro), suggesting a path to effective and efficient AI tutoring tools.
