SAVER

Researchers proposed SAVER, a dense-to-sparse post-training framework that trains video-language models to reason from limited temporal evidence, according to a new arXiv paper (2610.10893v1). Trained…

TeachMateGPT, a multi-agent framework introduced in an arXiv paper (arXiv:2608.13708v1), improves curriculum-grounded science assessment generation by raising faithfulness from 0.68 to 0.96 and answer…