Search papers, labs, and topics across Lattice.
3
0
4
State-of-the-art VLMs struggle to execute scientific research workflows, revealing a critical gap in their ability to interact with specialized software and produce verifiable results.
Placing reasoning traces before context in transformers can boost long-context reasoning accuracy by up to 100% while using exponentially less memory.
A staggering 42.5% of correct answers in long-document VQA can't be derived from terminal working memory alone, revealing a crucial oversight in current evaluation methods.
