Search papers, labs, and topics across Lattice.
1
0
3
1
Counterintuitively, giving web-navigating LLMs *more* HTML to read—not less—unlocks better action grounding and overall performance, but only if the model is capable enough and given sufficient "thinking" tokens.