Search papers, labs, and topics across Lattice.
Affiliation:
3
0
8
0
Localization-sensitive questions can see up to 40% performance improvement by dynamically selecting reasoning strategies based on question type.
Loop Memory Attention enables models to revisit earlier computations, leading to a 2.2% accuracy boost in MathQA tasks compared to fixed loop depths.
A single agent configuration shows a stark 68% performance drop when deprived of critical private conventions, underscoring the hidden knowledge gaps in LLM evaluations.
