Search papers, labs, and topics across Lattice.
4
0
6
Summary tokens are not just compressed context, but also dynamic KV cache indices, unlocking flexible memory compression and reduced compute in hybrid attention.
Giving agent harness optimizers unconstrained edit scope reliably degrades performance; enforcing graduated, periodic permission scheduling ($1\to2\to3$) instead yields an immediate 20% jump in held-out task completion.
Non-monotonic performance in experimental trajectories reveals that later rounds can sometimes undermine earlier successes, challenging conventional assumptions in industrial ML research.
